doi
stringlengths 28
28
| title
stringlengths 19
311
| abstract
stringlengths 217
5.08k
| plain language summary
stringlengths 115
4.83k
| article
stringlengths 3.87k
161k
|
|---|---|---|---|---|
10.1371/journal.ppat.1007719 | T-cell derived acetylcholine aids host defenses during enteric bacterial infection with Citrobacter rodentium | The regulation of mucosal immune function is critical to host protection from enteric pathogens but is incompletely understood. The nervous system and the neurotransmitter acetylcholine play an integral part in host defense against enteric bacterial pathogens. Here we report that acetylcholine producing-T-cells, as a non-neuronal source of ACh, were recruited to the colon during infection with the mouse pathogen Citrobacter rodentium. These ChAT+ T-cells did not exclusively belong to one Th subset and were able to produce IFNγ, IL-17A and IL-22. To interrogate the possible protective effect of acetylcholine released from these cells during enteric infection, T-cells were rendered deficient in their ability to produce acetylcholine through a conditional gene knockout approach. Significantly increased C. rodentium burden was observed in the colon from conditional KO (cKO) compared to WT mice at 10 days post-infection. This increased bacterial burden in cKO mice was associated with increased expression of the cytokines IL-1β, IL-6, and TNFα, but without significant changes in T-cell and ILC associated IL-17A, IL-22, and IFNγ, or epithelial expression of antimicrobial peptides, compared to WT mice. Despite the increased expression of pro-inflammatory cytokines during C. rodentium infection, inducible nitric oxide synthase (Nos2) expression was significantly reduced in intestinal epithelial cells of ChAT T-cell cKO mice 10 days post-infection. Additionally, a cholinergic agonist enhanced IFNγ-induced Nos2 expression in intestinal epithelial cell in vitro. These findings demonstrated that acetylcholine, produced by specialized T-cells that are recruited during C. rodentium infection, are a key mediator in host-microbe interactions and mucosal defenses.
| The nervous system is an active contributor to the regulation of immune responses. Prior studies have identified a unique CD4+ T-cell population that can relay signals from the sympathetic nervous system. These specialized T-cells express the enzyme choline acetyltransferase (ChAT) and produce acetylcholine (ACh). Release of ACh in response to neurotransmitters from the sympathetic innervation was previously shown to aberrant immune cell activation, reducing mortality during septic shock. Also, these CD4+ ChAT+ T-cells were previously found to control host-commensal interactions in naïve mice, but their role during enteric bacterial infection was unknown. Here we demonstrate that infection with C. rodentium induces ChAT+ T-cell recruitment and that expression of ChAT by this T-cell population significantly augments host defenses. These data support a diverse and expanding role of ACh in host immune responses.
| The recently revealed degree of integration between the nervous and immune systems are remarkable [1]. While it is well accepted that neurotransmitters can act on immune cells to alter cell activation and consequently host immune response, recent evidence demonstrates that select immune cell populations not only respond but can also produce neurotransmitters. Among these immune cells are the CD4+ T-cells that express choline acetyltransferase (ChAT), the enzyme required for acetylcholine (ACh) biosynthesis [2–4]. These T-cells are crucial intermediaries between the nervous and immune system, functioning to relay neuronal signals and prevent aberrant immune cell activation. Neural inhibition of inflammation can inhibit innate immune cell function in preclinical models of inflammatory bowel disease [5], rheumatoid arthritis [6], ischemia reperfusion injury [7, 8], and post-operative ileitis [9]. Immune regulation in this pathway requires norepinephrine (NE) released from neurons to activate β2 adrenergic receptors (β2AR) on ChAT+ T-cells causing the release of ACh [2].
Mucosal immunity is crucial to restricting access of commensal and pathogenic bacteria to the host. Host defenses are comprised of overlapping mechanisms that bind, flush away, exclude, or kill pathogenic enteric bacteria [10]. These roles are in part fulfilled by differentiated intestinal epithelial cells (IECs) that not only act as a physical barrier, but also produce and release mucus [11], bactericidal antimicrobial peptides [12, 13], and free radicals such as nitric oxide (NO) that are bactericidal or bacteriostatic [14, 15]. Loss of these protective mechanisms can result in aberrant immune responses to otherwise innocuous commensal bacteria, increased mucosal inflammation, or susceptibility to infection. In addition, mucosal homeostasis and host-resistance to pathogens is dependent on composition of the intestinal microbiota, with bacterial species that can reduce, or enhance susceptibility to pathogens including Citrobacter rodentium [16–18]. Physiological processes that govern these mechanisms of host defense and host-bacterial interactions are therefore paramount to the health of the host.
In the gastrointestinal tract, ACh enhances mucosal protection by controlling IEC functions ranging from release of mucus and antimicrobial peptides to increasing ion and fluid secretion [12, 19, 20]. Together, these mechanisms of mucosal defense maintain homeostatic interactions between the host and commensal microbiota, while limiting access of pathogens such as C. rodentium. Although the source of ACh regulating these functions of IEC has long been attributed to ChAT+ secretomotor neurons within the gastrointestinal tract, we and others have previously described ChAT+ T-cells that serve as essential non-neuronal sources of ACh [2–4]. This unique source of ACh appears to participate in mucosal immunity and host commensal interactions. As evidence of this, conditional ablation of ChAT in T-cells was found to reduce production of antimicrobial peptides in the small intestine of naïve mice, and induce changes in the jejunal but not colonic microbiota composition [13]. Despite these key observations, the role of ACh released from specialized T-cells during enteric infection is unknown. With these issues in mind, we have used ChAT-GFP reporter mice, and conditional ablation of ChAT in T-cells to assess the role of T-cell derived ACh in host mucosal immune function during C. rodentium infection.
Using this approach, we have identified that ChAT+ T-cells are recruited to the colon during C. rodentium infection, and that conditional ablation of ChAT in T-cells significantly increases C. rodentium burden in the colon. This increased susceptibility to infection is due to decreased expression nitric oxide synthase isoform 2 in IEC, with ACh acting to enhance IFNγ-induced gene transcription.
Mice used in this study are on a C57BL/6 background and were originally purchased from Jackson laboratories (Bar Harbor, ME), including CXCR5-/-, ChAT-GFP (B6.Cg-Tg(RP23-268L19-EGFP)2Mik/J)), ChATf/f and LCK.Cre to establish a breeding colony. ChAT T-cell conditional knockout (cKO) mice were produced by breeding ChATf/f and LCK.Cre mice to generate LCK.Cre- ChATf/f (WT) and LCK.Cre+ ChATf/f (cKO) mice. This breeding scheme permitted use of littermate cKO and WT controls. At 6–8 weeks of age, mice were gavaged with either LB, or Citrobacter rodentium (108 CFU (colony-forming unit), strain DBS100, generously provided by Dr. Andreas Baumler). In a subset of experiments, colitis was induced by administration of dextran sodium sulfate (DSS, 3% v/v) in the drinking water for 5 days followed by normal water for 3 days as previously published [21].
All procedures were approved by the Institutional Animal Care and Use Committee at UC Davis under protocol number 20170 in accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals. Mice were euthanized by CO2 asphyxiation followed by cervical dislocation according to American Veterinary Medical Association guidelines for collection of tissues.
IEC and lymphocytes were isolated from the colonic lamina propria according to a standard protocol [22, 23]. In brief the colons from mice were removed, cut open along the mesenteric boarder, washed in PBS, before being place in Hanks buffered salt solution (HBSS) with EDTA (5 mM) to remove IEC. For experiments to ascertain the amount of NOS2 expressed by IEC, dissociated cells suspensions were incubated for 20 min with Fc blocking (CD16/32) antibody (TONBO Biosciences, San Diego, CA) and subjected to staining using Live/Dead aqua viability stain (ThermoFisher Scientific, Waltham, MA), anti- CD45-FITC (ThermoFisher Scientific) and -EpCAM-BV421 (BD Biosciences, San Jose, CA). IEC were then subjected to fixation and permeabilization using Fix/Perm buffer, and washed with permeabilization according to manufacturer’s instructions (BD Biosciences). Intracellular staining was then conducted using anti-NOS2-APC (ThermoFisher Scientific). Lymphocytes were isolated from colonic tissues subjected to removal of IEC as above, followed by digestion of 1 cm cut fragments in Collagenase IV (40 mg/ml, Sigma Aldrich). Released lymphocytes were incubated in RPMI media (10% FBS Pen/strep L-glutamine (ThermoFisher Scientific)) supplemented with PMA/Ionomycin (Cell Stimulation Cocktail 1X, eBioscience), with Golgi-Stop (BD Biosciences) for 4 hours. Samples were subjected to staining with Live/Dead Aqua (ThermoFisher Scientific), anti -CD3-BV605 (BioLegend, San Diego, CA), anti-CD4-APC (TONBO Biosciences), followed by fixation and intracellular staining for with anti -IL-22-PerCPeFluo710 (ThermoFisher Scientific), -IL-17A-PE (BD Biosciences), -IFNγ-PECy7 (ThermoFisher Scientific) and -GFP-AF488 (BioLegend). All samples were then acquired on a BD LSRII, and assessed using FlowJo (Treestrar, Ashland OR).
A 1 cm segment of colon was cut and placed into a pre-weighed microcentrifuge tube to determine tissue weight. Samples were homogenized in sterile PBS using a 5 mm sterile stainless-steel bead (Qiagen, Germantown MD) in a bead beater (Qiagen). Sample homogenates were then diluted in sterile PBS, and (100 μL) plated onto MacConkey agar plates, and colonies counted after 16 h of incubation at 37°C.
Analysis of gene expression was performed by quantitative real-time PCR (qRT-PCR) as previously described [3]. Briefly, a 1 cm long segment of colon was homogenized in Trizol (Invitrogen, Carlsbad, CA) using a 5 mm stainless steel bead in a bead beater (Qiagen). RNA was extracted as directed by manufacturer’s instructions, with isolated RNA dissolved in ultrapure H2O (Invitrogen). Synthesis of cDNA was performed using an iSCRIPT reverse transcriptase kit (Bio-Rad, Hercules, CA), and Real time qPCR was performed for the following targets using the indicated primer pairs from Primerbank [24]: IL-1β forward 5’-CTGTGACTCATGGGATGATGATG-3’, reverse 5’-CGGAGCCTGTAGTGCAGTTG-3’, IL-6 forward 5’- TAGTCCTTCCTACCCCAATTTCC-3’, reverse 5’- TTGGTCCTTAGCCACTCCTTC-3’, IL-17A forward 5’- TTTAACTCCCTTGGCGCAAAA-3’, reverse 5’-CTTTCCCTCCGCATTGACAC-3’, IL-22 forward 3’-ATGAGTTTTTCCCTTATGGGGAC-5’, reverse 3’-CTGGAAGTTGGACACCTCAA-5’, IFNγ forward 5’- GCCACGGCACAGTCATTGA-3’, reverse 5’- TGCTGATGGCCTGATTGTCTT-3’, Tnfα forward 5’-CCCTCACACTCAGATCATCTTCT-3’, reverse 5’-GCTACGACGTGGGCTACAG-3’, NOS2 forward 5’-GTTCTCAGCCCAACAATACAAGA-3’, reverse 5’- GTGGACGGGTCGATGTCAC-3’, Chi3L3 forward, CTCTGTTCAGCTATTGGACGC, reverse 5’- CGGAATTTCTGGGATTCAGCTTC-3’, MRC1 forward 5’- CTCTGTTCAGCTATTGGACGC-3’, reverse 5’- CGGAATTTCTGGGATTCAGCTTC, Retnla forward 5’- CCAATCCAGCTAACTATCCCTCC- 3’, reverse 5’- ACCCAGTAGCAGTCATCCCA -3’, Arg1 forward 5’-CTCCAAGCCAAAGTCCTTAGAG-3’, reverse 5’-AGGAGCTGTCATTAGGGACATC-3’, Cxcl13 forward 5’- GGCCACGGTATTCTGGAAGC-3’, reverse 5’- GGGCGTAACTTGAATCCGATCTA-3’, Ccl1 forward 5’- GGCTGCCGTGTGGATACAG-3’, reverse 5’- AGGTGATTTTGAACCCACGTTT-3’, Ccl8 5’- TCTACGCAGTGCTTCTTTGCC-3’, reverse 5’- AAGGGGGATCTTCAGCTTTAGTA -3’, Ccl19 forward 5’- GGGGTGCTAATGATGCGGAA-3’, reverse 5’- CCTTAGTGTGGTGAACACAACA-3’, Ccl21 forward 5’- GTGATGGAGGGGGTCAGGA-3’, reverse 5’- GGGATGGGACAGCCTAAACT-3’, Defa3 forward 5'- GAGAGATCTGGTATGCTATTG-3', reverse 5'- AGCAGAGTGTGTACATTAAATG-3', Defa5 forward 5'- TCAAAAAAGCTGATATGCTATTG-3', reverse 5'- AGCTGCAGCAGAATACGAAAG-3', Defa20 forward 5'- GAGAGATCTGATATGCTATTG-3', reverse 5'- AGAACAAAAGTCGTCCTGAG-3', Defa21 forward 5'- GAGAGATCTGATCTGCCTTTG-3', reverse 5'- CAGCGCAAAAAAGGTCCTGC-3', Defa22 forward 5'- GAGAGATCTGATCTGCCTTTG-3', reverse 5'- CAGCGCAAAAAAGGTCCTGC-3', Defa23 forward 5'- GAGAGATCTGGTATGCTATTG-3', reverse 5'- AGCAGAGCGTGTATATTAAATG-3', Defa24 forward 5'- GAGAGATCTGGTATGCTATTG-3', reverse 5'- AGCAGAGCATGTACAATAAATG-3', Defa26 forward 5'- ATTGTAGAAAAAGAGGCTGTAC-3', reverse 5'- AGCAGAGTGTGTACATTAAATG-3', Itln1 forward 5'- ACCGCACCTTCACTGGCTTC-3', reverse 5'- CCAACACTTTCCTTCTCCGTATTTC-3', Reg3g forward 5'- CCTCAGGACATCTTGTGTC-3', reverse 5'- TCCACCTCTGTTGGGTTCA-3', Lyz1 forward 5'- GCCAAGGTCTACAATCGTTGTGAGTTG-3', reverse 5'- CAGTCAGCCAGCTTGACACCACG-3', Actb forward 5’-GGCTGTATTCCCCTCCATCG-3’, reverse 5’- CCAGTTGGTAACAATGCCATGT-3’. CIIta forward 5’-TGCGTGTGATGGATGTCCAG-3’, reverse 5’-CCAAAGGGGATAGTGGGTGTC-3’, Irf1 5’-ATGCCAATCACTCGAATGCG-3’, reverse 5’-TTGTATCGGCCTGTGTGAATG-3’. Amplification and data acquisition were conducted using a QuantStudio6 (ThermoFisher Scientific). Data were analyzed using the delta delta CT method normalizing gene expression to Actb in each sample followed by normalization to experimental control sample.
Fresh fecal samples from mice were collected and stored at -80°C until analysis. Pellets were extracted with nano-pure water (10 mg/mL) and gently agitated overnight at 4°C. The homogenized samples were centrifuged at 21,000 g for 5 min. Supernatants (100 μl) were transferred and centrifuged at 21,000 g again for 20 min. For each sample, 20 μl of the supernatant was mixed with 20 μl of 100 mM N-(3-Dimethylaminopropyl)-N0-ethylcarbodiimide hydrochloride (1-EDC HCl) (Sigma-Aldrich, cat. # E7750) in 5% pyridine (Sigma-Aldrich cat. # 270407) and 40 μL of 200 mM 2-Nitrophenylhydrazine (2-NPH) (Sigma-Aldrich, cat. # N21588) in 80% acetonitrile (ACN) (Sigma-Aldrich) with 50 mM HCl. The mixture was incubated at 40°C for 30 min. After reacting, 400 ml of 10% ACN was added to the solution. Then 1 μl of the solution was injected into an Agilent 6490 triple quadruple mass spectrometer for analysis. Chromatographic separations were carried out on an Agilent C18 stationary phase (2.1 x 50 mm, 1.8 um) column. Mobile phases were 100% ACN (B) and water with 10% ACN (A). The analytical gradient was as follows: time = 0 min, 10% B; time = 4.5 min, 10% B; time = 10 min, 35% B; time = 10.1 min, 85% B; time 11.6 min, 90% B; time 12 min, 90% B. Flow rate was 0.3 ml/min and injection volume was 1 μL. Samples were held at 4°C in the autosampler, and the column was operated at 40°C. The MS was operated in selected reaction monitoring (SRM) mode, where a parent ion is selected by the first quadrupole, fragmented in the collision cell, then a fragment ion selected for by the third quadrupole. Product ions, collision energies, and cone voltages were optimized for each analyte by direct injection of individual synthetic standards. The MS was operated in positive ionization modes with the capillary voltage set to 1.8 kV. Source temperature was 200°C and sheath gas temperature 200°C. Gas flow was 11 L/min, sheath gas flow was 7 L/min, and collision gas flow was 0.2 mL/min. Nebulizer pressure (nitrogen) was set to 25 psi. Argon was used as the collision gas. A calibration curve was generated using authentic standards for each compound.
Colonic tissue specimens were fixed in 10% normal buffered formalin for 24 h prior to gradual dehydration in ethanol, embedded in paraffin and 6 μm thick cross sections were cut onto glass slides. Slides with tissue sections were de-paraffinized and rehydrated according to standard protocols, stained with hematoxylin and eosin to allow for evaluation of histopathology. Crypt lengths were measured using bright field microscopy on these sections with FIJI (Fiji Is Just ImageJ) [25], measuring at least 20 crypts per animal.
Slides with colonic tissue section were also used for confocal analysis with antibodies raised against specific proteins of interest according to standard protocols. In brief, after slides were de-paraffinized and rehydrated, antigen retrieval was performed in citrate buffer (10 mM, pH 6.0, 30 min., 95°C). After blocking in 5% BSA and normal donkey serum, samples were incubated in primary antibody overnight (16 h 4°C). Primary and secondary antibodies used are detailed in Table 1. Slides were washed extensively (3 x 5 mins) in TBS-tween20 and incubated in appropriately labeled secondary antibodies (Invitrogen) for 1 h at room temperature, washed, counterstained with DAPI in TBS-tritonX100 0.1% v/v, washed and mounted in Prolong gold (Invitrogen). Staining using anti-mouse CDH1 (E-cadherin) was revealed using a mouse on mouse kit according to manufacturer’s instructions (Vector laboratories, Burlingame, CA). Slides were imaged on a Leica SP8 STED 3X confocal microscope with a 63X 1.4 NA objective. Areas larger than the field of view of the objective were acquired using a tiling approach, whereby adjacent images were acquired with a 10% overlap.
Analysis of standard confocal data sets was performed by opening Leica image format files in Imaris Stitcher (v9.0, Bitplane Scientific) to fuse overlapping fields of view together. These reconstructed areas were then analyzed using Imaris software. Expression of NOS2 in IEC was determined by creating a mask based on regions of CDH1 staining (i.e. IEC) that contained DAPI+ cells. This defined region was then interrogated for the number of IEC present, and the intensity of NOS2 staining. Counting of DAPI+ Ki67+ IEC, or T-cells (CD3+ DAPI+) cells were performed in a similar manner in 3–5 fused fields of view from each animal counted.
Following excision of the intestine, segments of colon were cut along the mesenteric border to allow for mounting in the Ussing chamber (Physiologic Instruments, San Diego, CA). Tissues were maintained in oxygenated Kreb’s buffer consisting (in mM) of: 115 NaCl, 1.25 CaCl2, 1.2 MgCl2, 2.0 KH2PO4 and 25 NaHCO3 at pH 7.35 ± 0.02 and maintained at 37°C. Additionally, glucose (10 mM) was added to the serosal buffer as a source of energy, which was balanced osmotically by mannitol (10 mM) in the mucosal buffer. Agar–salt bridges were used to monitor potential difference (PD) across the tissue, and to inject the required short‐circuit current (Isc) to maintain the PD at zero by an automated voltage clamp. Data from the voltage clamp (i.e. Isc, and PD) was continuously acquired using acquisition software (Physiologic Instruments). Baseline Isc values were obtained after equilibrium had been achieved approximately 15 min after the tissues were mounted. Isc, an indicator of active ion transport, was expressed in μA/cm2. After tissues reached stable short-circuit current for 15 minutes, stimulation of ion secretion was induced by addition of the muscarinic receptor agonist carbachol (20 μM). After returning to baseline the adenylate cyclase activating compound Forskolin ([FSK], 20 μM) was added and response recorded.
CMT-93 cell line (ATCC cat. CCL-223) from an induced carcinoma of mouse rectum was cultured in DMEM supplemented with 10% heat-inactivated fetal bovine serum (FBS supplemented with Pen Strep (ThermoFisher Scientific cat. 15070063), and maintained at 37°C in a humidified atmosphere with 5% CO2. Cells were seeded in a 6-well plate at a density of 1×106 cell/well and incubated for 48 h. The medium was replaced with serum-free DMEM for 2 h. Then, cells were treated with IFNγ 1 ng/ml (Peprotech cat. 315–05), Carbachol (100 μM, Sigma Aldrich) or both for 3h. Cells were collected and stored at -80°C on Trizol. RNA extraction, cDNA synthesis and qPCR were performed as described above.
Data were analyzed using one-way analysis of variance (ANOVA) in Prism (Graphpad, San Diego CA), with a P value of less than 0.05 denoted as significant.
Only sparse numbers of ChAT+ T-cells have been observed in the intestine of naïve mice [13], however the potential role of ChAT+ T-cells in the mucosal immune response during enteric bacterial infection has not been established. To assess if ChAT+ T-cells are recruited during infection, ChAT-GFP+ mice were infected with C. rodentium and the number of CD3+ ChAT-GFP+ T-cells determined by confocal microscopy on days 6, 10, 21, and 30 post-infection (p.i.). Mice infected with C. rodentium had a significant increase in the number of CD3+ChAT-GFP+ T-cells in the colon beginning 10 days p.i. which persisted until 30 days p.i. (Fig 1A & 1B). In order to characterize these recruited cells, flow cytometry was conducted on isolated lamina propria lymphocytes (LPL). These colonic lamina propria ChAT-GFP+ T-cells (Single, Live, CD3+, CD4+) 10 days post-C. rodentium infection produced IFNγ, IL-17A, and IL-22 (Fig 2A). Quantification revealed that ChAT-GFP+ T-cells express more IFNγ, IL-17A, and IL-22 by MFI (mean fluorescence intensity) compared to ChAT-GFP- T-cells (Fig 2B). Despite this, it is important to note that the frequency of ChAT-GFP+ T-cells actively producing IFNγ and IL-17A were significantly less compared to ChAT-GFP- T-cells. ChAT-GFP+ IL-22+ T-cell population appears to be persistent in the naïve colon and does not increase significantly during infection. These data demonstrate that the ChAT-GFP+ T-cells are not unique to Th1/Th17/Th22 T-cells subsets, and can be polarized to these three different phenotypes (Fig 2).
In contrast to the recruitment induced by C. rodentium infection, induction of colonic inflammation by the chemical irritant DSS failed to increase the number of ChAT-GFP+ T-cells compared to naïve control, despite evidence of overt inflammation (S1A Fig). Together, these results suggest that ChAT+ T-cells are a specific component of the host response to C. rodentium infection and their recruitment is driven by specific signals and rather than a simple consequence of intestinal inflammation.
The functional role of T-cell-derived ACh during C. rodentium infection was determined using a T-cell conditional knockout (cKO) approach. Accordingly, infected ChAT T-cell cKO mice had increased CFU/g of C. rodentium in colonic tissue at day 10 p.i. as compared to infected WT mice (Fig 3A). To determine if the increased bacterial burden of C. rodentium resulted in altered localization of the pathogen in the colon, confocal microscopy analysis using antibodies directed against C. rodentium was performed (Fig 3B). Compared to WT, we observed increased C. rodentium in the colonic lumen, adjacent to IEC (CDH1+ DAPI+), and the presence of microcolonies within the colonic crypts in ChAT T-cell cKO mice. Despite the increased bacterial burden in ChAT T-cell cKO mice, no significant increase in the number of proliferating (DAPI+ CDH1+ Ki67+) IEC cells (Fig 3D &3E), histopathological damage, or crypt hyperplasia was observed compared to infected WT mice (Fig 3C). Together these data indicate that T-cell derived ACh is a critical component of host defense during C. rodentium infection but does not influence epithelial barrier integrity or effect crypt hyperplasia. To assess what factors could contribute to recruitment of these CD3+ ChAT+ T-cells during C. rodentium infection, we performed qRT-PCR for chemokines that are cognate ligands for previously identified receptors expressed by this population of T-cells [4, 26]. The pattern of Cxcl13 expression closely mirrors the recruitment of ChAT+ T-cells (S2A Fig), with significantly increased expression beginning 10 days and lasting until day 30 p.i. while significantly increased expression of Ccl1, Ccl8, Ccl19 and Ccl21 occurred between 21 and 30 days p.i., well after recruitment of ChAT+ T-cells began. Critically, infection of mice deficient in CXCR5, the cognate receptor for CXCL13, did not experience increased C. rodentium bacterial burden or pathology (S2B Fig). In addition, assessment of intestinal physiology using Ussing chambers revealed no significant differences in conductance, baseline or evoked short-circuit current responses to carbachol or forskolin in naïve WT or ChAT T-cell cKO mice (S3 Fig).
To determine the immunological consequences of conditional ablation of ChAT in T-cells during C. rodentium infection, qRT-PCR was conducted on colon from LB control and infected WT and ChAT T-cell cKO mice for proinflammatory gene expression. At day 10 p.i., expression of Il-1β, Il-6, and Tnfα were significantly increased in C. rodentium infected mice compared to LB control mice (Fig 4). Expression of these cytokines was significantly enhanced 10 days p.i. in the ChAT T-cell cKO mice compared to WT infected animals. In contrast, expression of Ifnγ, Il-17a, Il-22were increased 10 days p.i. to a similar extent in WT and ChAT T-cell cKO mice (Fig 4). These data indicate that ablation of ChAT in T-cells can significantly alter the host immune response to C. rodentium, but in a manner that does not alter local Th1, Th17, or Th22 responses.
As ChAT T-cell conditional knockout mice were previously observed to have reduced expression of antimicrobial peptides [13], we questioned if this could result in an increased C. rodentium burden. Using qRT-PCR we observed no significant differences in antimicrobial peptide expression in the small intestine or colon (S4 Fig) in naïve WT and ChAT T-cell cKO mice. As expected, colonic expression of RegIIIγ was significantly increased after C. rodentium infection in both WT and ChAT T-cell cKO mice, however there was no difference between the two genotypes in the terminal ileum or colon (S4A & S4B Fig). As the commensal microbiota actively produces bioactive metabolites, we assessed if production of short-chain fatty acids (SCFA) was different in WT compared to ChAT T-cell cKO mice. Mass spectrometry revealed significant changes in specific SCFA during C. rodentium infection. Significantly reduced lactic acid was observed in infected WT but not in ChAT T-cell cKO mice. Butyric acid was significantly enhanced in both WT and ChAT T-cell cKO infected mice compared to uninfected WT or cKO control mice. While significantly increased production of pyruvic acid was detected in the feces from uninfected ChAT T-cell cKO mice, infection reduced the concentration of this metabolite to levels observed in uninfected or C. rodentium infected control mice. (S5 Fig). Together these findings indicate that the increased C. rodentium burden in ChAT T-cell cKO mice was not due to an inability to produce antimicrobial peptides or alterations in SCFA produced by the microbiota.
The increased expression of certain pro-inflammatory cytokines coupled with increased colonic C. rodentium burden in ChAT T-cell cKO mice lead us to ascertain if innate effector responses were intact. First, we considered if lack of T-cell derived ACh could increase differentiation of alternatively activated macrophages, disrupting the ability to mount and effect innate responses to C. rodentium. As indicated in Fig 5, no significant differences were noted in arginase1 (Arg1), mannose receptor C-type 1 (Mrc-1), chitinase-like 3 (Chi3l3), or resistin-like molecule α (Retnla) expression by qRT-PCR in colonic tissues between WT and ChAT T-cell cKO mice. Expression of Nos2 (“iNOS”) however was significantly abrogated 10 days p.i. in ChAT T-cell cKO mice compared to infected WT. These data indicate that lack of T-cell derived ACh does not increase alternatively activated/M2 macrophage polarization, but significantly impacts the expression of Nos2. As numerous cell types can express NOS2, we assessed the localization and quantity of NOS2 protein by confocal microscopy on colonic tissue from infected WT and ChAT T-cell cKO mice and uninfected controls. As indicated in Fig 6, IEC (CDH1+ DAPI+) were the predominant cell type that were immunoreactive of NOS2 during C. rodentium infection. In keeping with the qRT-PCR data, C. rodentium induced NOS2 expression was significantly reduced in ChAT T-cell cKO compared to WT mice (Fig 6A). Quantification of NOS2 expression in IEC further demonstrate reduced NOS2 expression in C. rodentium infected ChAT T-cell cKO mice (Fig 6B). This reduced ability to increase NOS2 expression in IEC during 10 days p.i. was further validated by flow cytometry conducted on IEC (Single, live, CD45-, EpCAM+) from naïve and infected WT and ChAT T-cell cKO mice (Fig 6C & 6D). These data demonstrate that T-cell deficiency in ChAT significantly impairs C. rodentium induced increases of NOS2 expression in IEC.
As ACh has been previously demonstrated to induce NOS2 expression in lung epithelial cells [27], we sought to determine if ACh could induce similar effects in IEC. The mouse colonic epithelial cell line CMT-93 was treated with IFNγ ± carbachol (ACh mimetic), with Nos2, irf1, and CIIta expression assessed by qRT-PCR. As expected, stimulation with IFNγ (1 ng/mL, 3 h, time and dose determined empirically) induced expression of Ciita, Irf1, and Nos2. Co-treatment with carbachol further significantly increased expression of Nos2 compared to IFNγ alone, but did not enhance Ciita or Irf1 expression (Fig 7). Treatment with carbachol alone failed to significantly increase expression of any of the target genes. These results suggest that cholinergic signaling in IEC can synergistically enhance select IFNγ induced genes including Nos2.
The finding that the nervous system is an active participant during inflammation has been an unexpected and intriguing finding. At the interface between these two systems is a unique type of T-cells that can produce and release ACh in response to sympathetic neurotransmitters [2, 28]. Although the predominant focus on these ChAT+ T-cells has been on their ability to reduce the severity of disease in a number of clinically relevant immunopathologies [5–9], ChAT+ T-cells can also help to establish host-commensal interactions. While a prior study demonstrated increased diversity of the small intestinal microbiota in ChAT T-cell cKO mice, due to reduced antimicrobial peptide expression [13], the role of these cells during enteric bacterial infection was not known.
Using a combination of ChAT-GFP reporter and ChAT T-cell cKO mice, our studies are the first to demonstrate recruitment of ChAT+ T-cells and a functional role for these cells during C. rodentium infection. These recruited ChAT-GFP+ T-cells do not appear to be restricted to a unique Th subset, with ChAT-GFP+ T-cells found to produce IFNγ, IL-17A, or IL-22 in agreement with prior studies [3].
As we and others have previously demonstrated that ChAT+ T-cells express the chemokine receptors CXCR5 [4] and CCR8 [29]; we sought to characterize the production of cognate ligands to these receptors during C. rodentium infection, Cxcl13 and Ccl8 respectively. Our analysis indicates that Cxcl13 but not Ccl8 expression is induced beginning 10 days p.i. until day 30 p.i., a period during infection that closely mirrors when the number of ChAT+ T-cells increased in the colon. This temporal pattern of chemokine expression is corroborated by other studies demonstrating increased Ccl8 during C. rodentium infection [30]. CXCL13 is well established as critical to organization of secondary lymphatic organs [31], tertiary lymphoid tissues and recruitment of IL-22 producing ILC3 [32] and can be induced by vagal nerve stimulation [33]. Despite this, our studies using CXCR5 KO mice indicate that this signaling axis is either not critical or functionally redundant with respect to the host response to C. rodentium 10 days post-infection.
The importance of ACh derived from T-cells to host mucosal immune response during enteric bacterial infection was determined using a ChAT T-cell conditional knockout. Highlighting the host protective role of ACh, we observed an increased bacterial burden following enteric C. rodentium infection in ChAT T-cell cKO compared to WT mice. This increased bacterial burden in ChAT T-cell cKO mice was associated with significantly increased expression of the pro-inflammatory cytokines Il-1β, Il-6, Tnfα, with equivalent expression of Ifnγ, Il-17a, or Il-22 compared to infected WT mice. Loss of T-cell derived Ach however did not impinge on IL-22 production, typically produced by ILC or T-cells in response to C. rodentium infection [34]. Together these findings, supported by the literature, suggest that ChAT+ T-cells are important in eliciting host-protective responses.
Mucosal immunity is comprised of a multitude of overlapping mechanisms that serve to protect the host from pathogens including the production and secretion of antimicrobial peptides. T-cell derived ACh has been implicated in regulation of host-microbial interactions at the mucosal surface by controlling antimicrobial peptide production. Conditional ablation of ChAT in CD4+ cells using CD4.Cre ChATf/f mouse line resulted in reduced lysozyme, defensin A, and ang4 expression in the small intestine, consequently increasing the diversity of commensal microbiota in the jejunum but not the cecum, or colon [13]. In contrast to these findings we noted no significant reductions in antimicrobial peptide expression in ChAT T-cell cKO compared to WT mice. As expected [35], expression of RegIIIy was significantly enhanced in WT and ChAT T-cell cKO mice during C. rodentium infection irrespective of genotype. These data suggest that the increased bacterial burden in ChAT T-cell cKO mice was not due to a deficit in antimicrobial peptide expression.
Host production of free radicals including NO are critical factors in protection against several bacterial pathogens [36–38]. NO also functions as a short-lived cell signaling molecule and is produced by three distinct isoforms of nitric oxide synthase that are each uniquely regulated in a tissue- or context-dependent manner [37]. In contrast to the constitutively expressed NOS found in endothelium or neurons, bacterial products or inflammation can induce NOS2 expression in a variety of cell types [37, 39]. Infection with C. rodentium increases NOS2 expression, functioning to limit bacterial burden and disease [14, 40]. In agreement with this literature, our data demonstrate that IEC in the colon are the predominant cell type expressing NOS2 during C. rodentium infection in WT mice. Conditional ablation of ChAT in T-cells, however, resulted in significantly reduced Nos2 expression compared to WT mice. Confocal microscopy on colonic tissue sections and flow cytometry experiments confirmed NOS2 expression was significantly increased in colonic IEC of WT mice, but not in ChAT T-cell cKO mice during infection. Together these data demonstrate that lack of T-cell derived ACh significantly reduced the induction of NOS2 in IEC during C. rodentium infection. Although NOS2 expression is characteristically elicited by IFNγ-induced activation of STAT1-dependent gene transcription [41], expression of this cytokine was not affected by ChAT T-cell deficiency. Additionally, we observed that acetylcholine mimetics significantly enhance IFNγ-induced Nos2 expression in IEC in vitro, in agreement with previously reported experiments in lung epithelial cells [27].
There are striking similarities in the aberrant host response to C. rodentium infection in ChAT T-cell cKO and the previously described Nos2-/- mice [14]. For example, both mouse lines exhibit increased bacterial burden at day 10 p.i. without resulting in increased mortality or enhanced colonic histopathology [14]. Although Nos2 deficiency in mice, or inhibition of NO production increases Th17 differentiation [42], no significant increase in Il-17a expression was observed in the colonic tissue from C. rodentium infected ChAT T-cell cKO mice compared to WT mice. This is likely due to the short half-life of NO in biological fluids [43], and the expression of NOS2 in colonic IEC far from differentiating T-cells in draining lymph nodes.
Our data further substantiate the unique role of ACh producing ChAT+ T-cells in modulating immune function. These unique T-cells appear to function as a critical component of the mucosal immune system, limiting the number and detrimental effects of enteric bacterial pathogens. How these specialized T-cells that are recruited to the colon, become activated, and release ACh during C. rodentium infection warrants future study. Given the requirement for NE signaling through the β2AR receptor on ChAT+ T-cells in septic shock [2], activation by the sympathetic innervation is a strong possibility. Supporting this contention, Salmonella typhimurium induces activation of the sympathetic innervation within the small intestine, and the release of NE adjacent to muscularis macrophages [44]. While it is tempting to speculate that infection induced activation of a neuronal circuit is host protective, it is important to note that host NE induces bacterial expression of virulence genes by enteric pathogens such as C. rodentium [45] and enterohemorrhagic Escherichia coli [46]. Future studies will only further illuminate the integrated nature of the nervous system and immune system with ChAT+ T-cells as a critical node mediated host protection during enteric bacterial infection.
|
10.1371/journal.pcbi.1003835 | Mesoscopic Model and Free Energy Landscape for Protein-DNA Binding Sites: Analysis of Cyanobacterial Promoters | The identification of protein binding sites in promoter sequences is a key problem to understand and control regulation in biochemistry and biotechnological processes. We use a computational method to analyze promoters from a given genome. Our approach is based on a physical model at the mesoscopic level of protein-DNA interaction based on the influence of DNA local conformation on the dynamics of a general particle along the chain. Following the proposed model, the joined dynamics of the protein particle and the DNA portion of interest, only characterized by its base pair sequence, is simulated. The simulation output is analyzed by generating and analyzing the Free Energy Landscape of the system. In order to prove the capacity of prediction of our computational method we have analyzed nine promoters of Anabaena PCC 7120. We are able to identify the transcription starting site of each of the promoters as the most populated macrostate in the dynamics. The developed procedure allows also to characterize promoter macrostates in terms of thermo-statistical magnitudes (free energy and entropy), with valuable biological implications. Our results agree with independent previous experimental results. Thus, our methods appear as a powerful complementary tool for identifying protein binding sites in promoter sequences.
| Binding of specific proteins to particular sites in the DNA sequence is a fundamental issue for gene regulation in molecular biology and genetic engineering. A deep understanding of cell physiology requires the analysis of a plethora of genes involving characterization of their promoter architectures that determine their regulation and gene transcription. In order to locate the promoter elements of a given gene, experimental determination of its transcription start site (TSS) is required. This is an expensive, time-consuming task that, depending on our requirements, could be simplified using computational analysis as a first approach. Nevertheless, most computational methods lack a physical basis on the protein-DNA interaction mechanism. We adopt here this strategy, by using a simple model for protein-DNA interaction to find TSS in a bunch of cyanobacteria promoters. We make use of physical tools to characterize these TSS and to relate them with biological properties as the relative strength of the promoter. Our study shows how a model based on a coarse-grained description of a biomolecule can give valuable insight on its biological function.
| Transcriptional regulation is the main mechanism for gene control in prokaryotes. In order to adapt optimal protein expression to nutritional and environmental conditions, a cascade of transcriptional regulators works as signal transducers determining the accessibility of RNA polymerase to bacterial promoters. In the last years, high throughput approaches have been confirmed as powerful tools for a better understanding of the regulatory networks that govern key aspects of cell physiology, such as the mechanisms leading to pathogenesis or the acclimation to xenobiotics and hostile environments, among others [1]–[4].
However, successful transcriptome sequencing requires the generation of comprehensive transcriptome profiles that rely on the isolation of a sufficiently large number of reads to detect those biologically relevant transcripts, that represent a relatively small proportion of the cDNA library [5]. Moreover, those procedures are time consuming and, in many cases, the budget for sequencing costs constrains the total number of reads that can be obtained [6], [7].
Therefore, computational methods emerge as valuable complementary approaches for prediction or further validation of high throughput results [8], [9]. Mostly, a statistical approach to the study of sequences is adopted, leading to a general lack of methods based on the physical mechanism of protein-DNA interactions. A possibility to tackle the problem is the microscopic study of protein-DNA interaction [10]–[12], but this approach demands huge computer facilities and it is restricted to few base pairs up to the date. In this sense, coarse-grained models arise as powerful tools to model biological systems, speeding up the computation and allowing to get a deeper insight in the physical interactions [13], [14]. Adopting this strategy, we develop a coarse-grained model that allows for the analysis of promoter sequences and the identification and characterization of protein binding sites, likely related to transcriptional activity in the genome of the nitrogen-fixing cyanobacterium Anabaena PCC 7120.
Cyanobacteria are the only prokaryotes able to perform oxygenic photosynthesis, being key contributors to fixation. The ability of some cyanobacterial strains to fix atmospheric nitrogen or the formation of harmful blooms by toxigenic species, among other properties, evidence their ecological relevance [15]. Besides, cyanobacteria are an excellent model for the study of multicellularity in prokaryotes [16] and potential sources for novel drugs derived from their secondary metabolites [17].
The genome of Anabaena PCC 7120 contains 7,211,789 base pairs (bp) and 6,223 genes organized in a 6,413,771 bp chromosome and 6 plasmids [18]. Anabaena PCC 7120 has been used for long time as a model for the study of prokaryotic cell differentiation and nitrogen fixation [19]. More recently, the experimental definition of a genome wide map of transcriptional start sites (TSSs) of Anabaena together with the analysis of transcriptome variations resulting from the adaptation to nitrogen stress have provided a holistic picture of this complex process [20].
The problem of protein-DNA recognition is a widely debated issue, yet far to be fully understood. In this sense, it has been widely reported how the physical properties of the DNA chain result in key functional consequences in this process. DNA local structure highly influences some transcription factors (TFs) binding [21]–[23]. Thermal stability and bubble formation (i.e. local long-lived transient openings in the DNA strands) has also been extensively reported to correlate with several DNA functions, such as the recombination rate, single nucleotide polymorphism, DNA replication or gene transcription [24]–[27]. In this regard, the relation between bubble formation and the location of protein binding sites, is a lengthly, controversial debate, greatly nourished by the study of Peyrard-Bishop-Daxouis (PBD) model [28], [29]. This mesoscopic model was initially intended to reproduce the DNA melting transition, though it has been widely used afterwards for studying bubble formation on DNA promoters, likely correlated with biological relevant sites in the sequence, such as the TSS or the TATA box [30]–[35].
Despite the lack of consensus on whether PBD model is suitable for predicting protein binding sites [36]–[39], strong evidence supports this idea, showing clear correlation between regions with high propensity to form bubbles, and the presence of binding sites of DNA-interacting proteins such as RNA polymerase, [30]–[32], [40] or some TFs [33], [34], [41], [42]. Even more, succeeding revisions of this model showed clear relation between flexibility profiles and location of TSSs [43]. Grounded on these evidences, we propose a physical model for protein-DNA interaction in promoters [44], based on the coupling of a generic particle with the sequence-dependent bubble formation. This simple model is combined with a suitable analysis method [45] allowing the detection of biologically relevant sites, namely TSSs, on promoters of a prokaryote genome.
In order to prove the capacity of prediction of the computational methods developed in [44] and [45] for identifying the TSSs of a promoter, we have analyzed the result of simulating the dynamics of nine promoters of Anabaena PCC 7120. We have analyzed the simulations outputs and built systematically the relevant macrostates of the system. In every case, our analysis algorithm finds the TSS as one of these states, yielding in addition thermodynamic parameters (e.g. free energy, entropy) that allow their physical characterization and thus further biological discussion. In this regard, our method arises as a complementary tool that, from physical principles, finds protein binding sites (we focus on TSSs) and characterizes them, allowing to discuss the strength -in terms of RNA production- of such sites, something not achievable by statistical methods. Remarkably, in this case the base pair sequence is the only previous information required. Thus, our numerical outcomes are independent numerical predictions to be confronted with previous or future experimental results.
We base our model on a modification of the PBD model [28]–[31], [35] to include the interaction with a generic particle as a sliding protein coupled with the sequence. PBD model reduces the complexity of DNA to a set of units that represent the base pairs of the chain (see Fig. 1). The only degrees of freedom are the coordinates which stand for the opening of each base pair. The total Hamiltonian of the model accounts for two phenomenological interactions, the intra-base and the inter-base potentials, , where is the linear momentum of the base pair and its reduced mass.
The potential describes the inter-base pair or stacking interactions. The election is the anharmonic potential [28] whose elastic constant is for small openings but drops to for large . The parameter sets the length scale for this behavior.
The original PBD model uses Morse potential for the intra-base pair interaction. Nevertheless, a successful modification includes an entropic barrier which accounts for solvent interactions with open base pairs [35], [46], [47]. This modification sharpens the thermal denaturation and stabilizes the bubbles, reproducing in a more realistic way the experiments [35], [46], [47]. We include this effect adding a gaussian barrier [35], thus . Sequence dependence is introduced only in this potential term as the interaction is stronger if the base pair is C-G than if it is A-T (see Text S1 for the complete set of parameters). Sequence-dependence can be also introduced in the stacking potential parameters, a modification that accounts for flexibility properties of the DNA chain [40], [43], [48].
Inspired on the one-dimensional diffusion stage of DNA-interacting proteins [49], we include a new degree of freedom to the traditional PBD model. This new degree of freedom consists on a brownian particle that moves along the DNA chain (see Fig. 1 for a schematic representation of the total system) interacting with it through a phenomenological potential which depends on , the coordinate of the Brownian particle along the DNA molecule, and the DNA instantaneous configuration (1)
This potential creates a classical field composed by a sum of gaussian wells centered at each base () and whose amplitude depends on the opening of the base pair. The term allows a linear dependence for low saturating the interaction for large in order to avoid self-trapping. In this sense, the particle interacts more intensely with open regions of the sequence. In addition, the base pairs are also affected by the particle, so that they will be more likely to be opened if the particle is within its range of interaction. The model introduces only three new parameters, as the longitudinal scale over which the particle slides is adimensional (). The interaction intensity and width are set so that bubbles span around base pairs, an adequate value for the kind of processes studied here [50]. The parameter saturates the interaction around , typical value for open base-pairs [50]–[52].
The model is simulated by integrating numerically the Langevin equations for the chain base pairs and the particle using the stochastic Runge-Kutta algorithm of fourth order [53] (see Text S1 for explicit formulation of the equations of motion). Each of the DNA sequences we study is simulated in five different realizations, each one covering , with a preheating time of . For sequences up to base pairs, these times are enough to ensure equilibrium and ergodicity. In addition, since one-dimensional diffusion times of binding proteins are in the range of milliseconds, our simulation times are reasonable from a biological perspective. The simulation temperature is . We use periodic boundary conditions for the diffusing particle and fixed boundary conditions for the sequence, adding base pair clamps at the end of each sequence to provide “hard-boundaries” and avoid undesirable end effects. Relevant observables from the trajectories can be obtained, mainly the base pairs mean position , where is the number of realizations and the simulation time of each realization, and the particle's trajectory histogram.
The large dimensionality of the system requires a method to reduce the number of coordinates while keeping the relevant information of study. PCA [54] is one the most popular methods to reduce systematically the dimensionality of a complex system. PCA performs a linear transformation by diagonalizing the covariance matrix , and thus removing all internal correlations. It has been proved that, by ordering the eigenvalues decreasingly, the few first principal components contain most of the fluctuations of the system, and thus can be chosen as convenient reaction coordinates [35], [55], [56].
We project the base pair trajectories into the first five eigenspaces, describing thus the system in terms of the first five principal components and the particle trajectory. With this choice we keep over the of the fluctuations.
The Conformational Markov Network has been proven to be a useful and powerful tool to analyze trajectories from high dimensional systems, such as those from Molecular Dynamics simulations [45], [57]–[59]. This representation is obtained by discretizing the conformational space explored by the system in order to build a complex network. Each node in the network represents a discretized region of the conformational space, a conformational microstate, weighted according to the fraction of trajectory visiting such microstate. The links of the network coincide with the observed transitions between microstates, and are thus directed and weighted. We build the Conformational Markov Network of our system by considering the posible positions of the particle along the chain, and binning each of the five principal components into bins.
Typically, the Conformational Markov Network is formed by a large number of nodes which prevent a direct interpretation of the results. In order to extract relevant information about the physical states of the system and its relevance in the dynamics, we split the network into its basins of attraction, i.e. regions in which the probability fluxes () converge to a common state (attractor) of the network. To do so, we apply the stochastic steepest descent algorithm, developed in [45], building a coarse grained representation of the former network. From this basin network, the Free Energy Landscape (FEL) can be represented as a hierarchical tree diagram (dendrogram or disconnectivity graph) [60], [61], by assigning to each node a free energy according to its weight where is the weight of the heaviest basin. This magnitude is used as a control parameter, increasing it step by step from the weightiest node, so that new nodes arise, together with their links (see Text S1 for a more explicit exposition of the algorithm). The disconnectivity graph represents each basin of attraction hierarchically ordered according to its free energy, while the connections among them stand for the barriers needed to jump from to another (see below and Text S1 for plots of the disconnectivity graphs or dendrograms).
We define now the macrostates of the system by clustering every basin separated by a free energy barrier lower than , as the system transits among them within short waiting times. In fact, we can check how they represent qualitatively similar physical configurations. Each macrostate has an assigned weight . We want to calculate free energy differences between specific and non-specific states. The basin network contains a huge number of low populated states, see [35], that constitute transitionary states between well defined attractors of the system. Physically, they are short-lived transitionary states where the particle diffuses until it binds to a target site. We determine these non-specific states as every basin with a population and calculate free energy differences between specific and non-specific states as , where is the total weight of all non-specific states. In addition, we define the entropy of a macrostate as .
We have analyzed nine promoter sequences from Anabaena PCC 7120 which exhibit different features. Using our computational approach, we have to identify the TSSs in the promoter sequences as sites where bubbles form with high probability. Within the frame of our model, this is reflected in larger openings of the chain at these sites and higher probability of the particle to visit them. Next, we apply the analysis algorithm to define the macrostates of the system and extract the FEL as a dendrogram or disconnectivity graph [59], [60]. This procedure allows us to characterize these states in order to extract solid conclusions about each sequence. The strength of each TSS can be determined and, if the sequence presents more than one TSS, their relative strength can be compared, obtaining useful biological conclusions.
Up to our knowledge, most works concerning PBD model limit themselves to the study of short promoter sequences, without justifying the study of this region alone, or how would the model behave in coding regions. In order to cover this gap, we have simulated the behavior of three complete genes from Anabaena PCC 7120. We use here the PBD model without including the interacting particle, as we wish just to check in which regions from a whole gene bubbles form more easily. The results allow us to compare the occurrence and intensities of the fluctuations detected in the promoter and the coding regions, validating our further analyses restricted to the promoter sequences.
Figure 2 shows the first four PCA eigenvectors for the analyzed genes with the promoter and codifying regions highlighted. Very localized eigenvectors indicate strong fluctuations in the region of maximal amplitude. As we can see in Fig. 2, the first eigenvector is delocalized, with small amplitude, accounting for the overall fluctuations of the whole sequence. Nevertheless, the three next eigenvectors are highly localized in specific spots of the sequence. Remarkably, these sites appear in the promoter sequence. Thus, when considering a complete gene within PBD model, most of the system fluctuations occur in the promoter sequence; this is, bubbles form with higher probability there, while the codifying region remains on average closed. This reveals the role of the DNA sequence in the DNA dynamics, and its influence on the DNA-protein interaction problems, supporting strongly that some binding sites in the promoter sequence can be characterized as regions where bubbles form easily, enhancing protein interaction.
We have used the complete model (chain and particle) to analyze nine promoter sequences comprising to base pairs. In addition, we have chosen promoters with different features, five with a single well characterized TSS (alr0750, argC, conR, furA and nifB), while four of them exhibit multiple TSSs (furB, ntcA, petF and petH) [62]–[69]. Figure 3 shows the base pair opening profile for each promoter sequence with the TSSs highlighted. The particle trajectory histograms are also plotted. In any case, a peak appears close to the TSS, meaning that, on average, bubbles form with high probability around it. In turn, the particle is attracted by this site, as it dwells with high probability around the TSS.
As it has been pointed out in several studies, the PBD model by itself has been successfully used to analyze promoter sequences, finding protein binding sites where bubbles form with high probability, so allowing the identification of TSSs or the TATA-box [30], [32]. Nonetheless, introducing this additional degree of freedom appears as a key feature for our purposes. We are mimicking an hypothetical searching mechanism that indeed affects the dynamics of the system. In the PBD model alone, opening events appear as rare excitations of the unique ground state, where the whole chain is closed. The particle enhances chain opening, stabilizing the bubbles, that last for longer times (around two orders of magnitude longer), enriching the free energy landscape. In addition, bubbles span over a larger number of base pairs, typically around , which is a consistent number if we attend to those that form in transcriptional processes [51], [52].
It is also remarkable that the opening probability is not strictly related with the A-T content of the local sequence. Although it is clear that long A-T stretches form “softer” regions in the sequence that can open easier, this intuitive argument does not necessarily applies always. The interplay between the sequence and the dynamics is much more complex. The nonlinearity in the Hamiltonian, the long-range cooperativity of the model and the disorder of the sequence revealed in its heterogeneity affects directly the equilibrium and dynamical behavior of the model, being essential to understand the actual breathing dynamics of DNA, as it has been pointed out in previous studies [30], [31], [40].
Interestingly, besides the peaks centered on the TSSs, other regions exhibit high probability to form bubbles. Many of these peaks correspond to typical regulation sites of bacteria, such as those located at or from the TSS, also claimed to be related with bubble formation [30], [40]. These regions appear thus as candidates for possible binding sites of other TFs that are known to be influenced by the physical properties of the DNA chain. Nonetheless, we focus our discussion just on the TSS, as they have been systematically identified in the genome of Anabaena PCC 1720.
In order to analyze the sequences in a more systematic way we apply the FEL analysis described in the methods section. This algorithm allows us to define the most relevant states in the dynamics characterizing them from a quantitative point of view. So far, we have shown which regions in the promoter sequences exhibit a higher probability to form bubbles and to be visited by the particle. Nonetheless, these magnitudes give just qualitative information, as the average do not inform about the importance of opening events in the system. The real interest of our model and method is the possibility of giving quantitative measures about the “strength” of the different sites in the sequences, specially interesting in those promoters with several TSSs. Each site can be characterized by the thermodynamical magnitudes calculated from the FEL landscape analysis.
We present together the data extracted from the simulation and analysis methods in Table 1. For each of the nine analyzed sequences we show the weight, free energy difference with respect to the non-specific states and the entropy of the TSSs state, all previously defined. We include also other non-identified states in case they appear relevant in the dynamics. Most populated states suppose most stable states, giving rise to high free energies differences. The entropy is the multiplicity of such macro states. Even if the free energy is high, a low entropy would indicate that this macro state is made up of few, yet very populated, basins, physically meaning that the state is very localized (narrow bubbles). The opposite case would indicate that the algorithm finds many, less populated basins that represent the same macrostate. This duality could indicate different regulation behaviors that are further addressed in the Discussion section.
To illustrate the FEL, Fig. 4 shows the free energy dendrograms of three chosen promoters (see Text S1 for the six remaining dendrograms). For the sake of clarity, we do not show the region corresponding to non-specific basins (where , defined above). The position of each basin on the vertical axis informs about its stability, while their hierarchical arrangement about the barrier needed to jump between each state. The dendrogram or disconnectivity graphs provides thus valuable and intuitive information about the thermodynamic and kinetic properties of the FEL of each promoter.
Groups of basins separated by barriers lower than are highlighted by a color circle, defining the macrostates of the system according to the criterion detailed in Methods section. We plot together the physical state associated with it, showing also the fraction of trajectory they occupy. Such states correspond to a large bubble located on the target site, with the particle centered there.
In most cases, the most populated macrostate, and thus the most stable one, coincides with an excitation in the TSS region. Other non identified sites also suppose very populated macrostates, suggesting the possibility of additional regulation sites as it is discussed in next section. Our method arises thus as a powerful tool to complement experimental results, providing additional physical information about the relative importance of these sites in regulation processes.
In this work, we propose the use of a coarse-grained model for protein-DNA interaction to analyze promoter sequences, allowing the detection and characterization of protein-binding sites (we focus on the TSS). The proposed model is based on physical principles and inspired on a relatively simple idea: certain DNA-interacting proteins (as RNA polymerase) couple their binding to DNA bubble dynamics. Due to this, we base our model on a PDB representation of the DNA chain -having been proven to reproduce DNA bubble dynamics successfully- and couple it to an additional degree of freedom representing the protein. In the framework of this model and by using a free energy landscape analysis, we have studied promoters of Anabaena PCC7120, allowing the detection and characterization of the TSSs.
Upon genome analysis and TSSs detection, high-throughput approaches, such as proteomics, are commonly used, resulting in an enormous amount of data in a relatively short period of time. However, analysis of raw data to end up in genome annotation or TSSs mapping is a demanding, time-consuming task, necessary for taking advantage of this information that may delay a more detailed analysis of specific issues. Among the large variety of these methods (see [70], [71] for review of most existing methods) a great amount of valuable information is obtained, resulting in highly efficient analysis of genome that, nonetheless, generally lacks a base on the physical mechanism of protein-DNA interaction. In this sense, our model and analysis method adopt a different strategy, not willing to compete in time performance with statistical-based techniques, but allowing a deeper understanding on the driving processes of protein binding. As a consequence of that, we are able not only to identify the TSSs, but also to characterize them in terms of physical magnitudes, allowing discussions about the strength of each site.
The nine promoters of cyanobacterium Anabaena PCC7120 studied in this work have been chosen in order to make the most of our model, without forgetting about its limitations. The genome of Anabaena PCC 7120 is well-known and the positions of TSSs have been defined under different metabolic conditions [72]. Firstly, it is remarkable how the different TSSs in the analyzed genes coincide with relevant states in the dynamics of the model, characterized as the heavier basins. In order to relate the information obtained with possible biological interpretation, we have analyzed a set of genes exhibiting several TSSs and whose regulation has been well characterized [67], [68], [73]–[77]. This choice allows us to assess directly the potential relation between the binding free energy values displayed in Table 1 for each of the located sites, and the relative strength of different TSSs associated to the same gene.
Among them, it is worth to mention the case of the ntcA promoter. The average opening shown in Fig. 3 reveals how the three existing TSSs in this base pairs sequence [78] are clearly identified, agreeing also as sites which the particle visits with high probability. As displayed in Table 1, the relative free energy (with respect to the NS states) of the three TSSs is quite different. Indeed these values are in very good agreement with the occurrence and behavior of the three TSSs experimentally determined [78], [78]–[80]. TSS2, located at position , produces a constitutive transcript regardless of the culture conditions, while TSS1 (position ) is only used in the absence of nitrogen. Finally TSS3 (position ) is also active under all conditions, but its use is highly induced under nitrogen deprivation. Table 1 displays a remarkably low free energy for TSS1, indicating that the presence of this macrostate is low in the dynamics, suggesting that its expression might be enhanced under more restrictive conditions. On the other hand, TSS2 and TSS3 appear as strong binding sites, covering both a large fraction of the total dynamics. These values are in good agreement with the ntcA transcription level at these sites under the correspondent conditions of nitrogen availability.
FurB, petF or petH show also consistent results. The TSSs of the three promotores are clearly identified, coinciding with the experimental positions [66], [72], [81]. Determination of TSSs for FurB promoter using the primer extension technique unravels revealing two TSSs at positions and from the ATG, both with similar intensities ([66]). Our in silico analysis is in good concordance with such conclusions, as we find two major macrostates with very similar weight ( and ) with an excitation just on these positions. The resulting profiles when the promoters of petF and petH are analyzed also display several preferred macrostates. Primer extension assays revealed a single TSS for the petF gene located at 100 bp upstream the translation start site [82]. More recently, high throughput analysis showed two TSSs for petF, at and , bp, in a better agreement with our predictions. Transcription of petH, encoding ferredoxin-NADP+ reductase takes place from a constitutive promoter at bp from the ATG and a NtcA activated promoter (TSS at bp). According to the proposed model, both TSSs are found as relevant macrostates in the basin network, although not as high peaks in Fig. 3. Indeed, the constitutive TSS () exhibits a higher probability () than the non-constitutive one (Table 1), indicating that the model is consistent with the experimental observations.
Concerning the five remaining promoters, high peaks are found around their single TSS, coinciding with the most (or one of the most) populated macrostates as we have defined them (Table 1). The case of conR is where our model works worse, as a significantly more relevant state appears in the dynamics. It should be noted that most experimentally determined TSSs have been obtained under standard culture conditions or under nitrogen deprivation, and the existence of additional TSSs under different conditions -impossible to account explicitly in our model- cannot be discarded. In addition, it must be noted that the model is not considering exclusively DNA-RNA polymerase interaction, but the influence of DNA breathing dynamics on protein binding. In such sense, additional binding sites for other proteins which are influenced by mechanical changes in the DNA conformation may also be detected.
We have compared our numerical results to the existing experimental ones on TSSs positions and intensities. Nonetheless, it is important to note that our method identifies additional relevant regions of the promoters that have not been experimentally probed yet. We shall mention the cases of promoters furA, conR or nifB where very populated macrostates appear aside from the discussed TSSs. Although we do not exclude the possibility of false positives, these macrostates may be related with unknown regulatory regions. Thus, our results suggest further experiments to search possible new relevant activity regions. Moreover, additional TSSs might appear if studied under different culture conditions, revealing the complexity of transcriptome profiles even in the case of simple organisms such as bacteria. To finish, we have already mentioned studies discussing the influence of bubble formation on certain DNA-binding proteins aside from RNA-polymerase [33], [34], [41], [42]. Being our model based on general physical features, additional macrostates found through our method might indicate the existence of binding sites for further regulatory proteins which participate in transcriptome processes of Anabaena PCC 7120.
Anabaena PCC 7120 has been shown to be an ideal experimental system to probe our numerical method. As it has been displayed, our results agree current experimental knowledge and propose possible new relevant activity regions. However, the model can be applied to the study of promoter sequences in many other organisms. Being the identification of protein binding sites in promoter sequences a key problem to understand and control regulation in biochemical and biotechnological processes, our methods appears as a powerful complementary tool in this scientific endeavor.
|
10.1371/journal.pntd.0002509 | Infection with Usutu Virus Induces an Autophagic Response in Mammalian Cells | Usutu virus (USUV) is an African mosquito-borne flavivirus closely related to West Nile virus and Japanese encephalitis virus, which host range includes mainly mosquitoes and birds, although infections in humans have been also documented, thus warning about USUV as a potential health threat. Circulation of USUV in Africa was documented more than 50 years ago, but it was not until the last decade that it emerged in Europe causing episodes of avian mortality and some human severe cases. Since autophagy is a cellular pathway that can play important roles on different aspects of viral infections and pathogenesis, the possible implication of this pathway in USUV infection has been examined using Vero cells and two viral strains of different origin. USUV infection induced the unfolded protein response, revealed by the splicing of Xbp-1 mRNA. Infection with USUV also stimulated the autophagic process, which was demonstrated by an increase in the cytoplasmic aggregation of microtubule-associated protein 1 light chain 3 (LC3), a marker of autophagosome formation. In addition to this, an increase in the lipidated form of LC3, that is associated with autophagosome formation, was noticed following infection. Pharmacological modulation of the autophagic pathway with the inductor of autophagy rapamycin resulted in an increase in virus yield. On the other hand, treatment with 3-methyladenine or wortmannin, two distinct inhibitors of phosphatidylinositol 3-kinases involved in autophagy, resulted in a decrease in virus yield. These results indicate that USUV virus infection upregulates the cellular autophagic pathway and that drugs that target this pathway can modulate the infection of this virus, thus identifying a potential druggable pathway in USUV-infection.
| The identification of cellular components and metabolic pathways involved in virus replication provides valuable information for the development of new antiviral strategies. Autophagy is one of these metabolic pathways with multiple implications during viral replication. Autophagy literally means self-digestion and constitutes a cellular process by which intracellular components are enclosed by membrane structures and degraded. Interestingly autophagy can contribute either positively or negatively to viral infections. For instance, several viruses hijack these autophagic membranes to build their replication complexes or take advantage on metabolic rearrangements induced following autophagy, while in other cases autophagy contributes to viral clearance and innate immunity. In this study, we explored the possible implication of the autophagic pathway during Usutu virus infection (USUV). USUV is an African mosquito-borne flavivirus that mainly infects mosquitoes and birds, although infections in humans have been also documented, thus warning about USUV as a potential health threat. Our results indicate that infection by USUV of different origins triggers an autophagic response within infected cells. Even more, drugs that target components from the autophagic pathway modulate USUV-infection. These results provide the basis for the design of new antiviral research lines against this pathogen.
| The variety of factors that have contributed to the emergence of the flavivirus West Nile virus (WNV) in the Americas and its re-emergence in other parts of the world could also provide a suitable scenario for the emergence of other arboviruses [1], [2], [3]. These potential threats for human and animal health include other related mosquito-borne viruses such as Usutu virus (USUV) [4]. USUV is an enveloped single-stranded positive polarity RNA virus that belongs to the Flavivirus genus in the Flaviviridae family. USUV was first described in South Africa in 1959, and since then, it has been reported in several African countries including Senegal, Central African Republic, Nigeria, Uganda, Burkina Faso, Cote d'Ivore, and Morocco [5]. The host range of USUV in Africa mainly comprises ornitophilic Culex mosquitoes and birds, although two isolations of USUV from human serum, including one severe case, have been documented [5]. USUV was reported to be circulating only in Africa until 2001, when it emerged in Central Europe [6]. From that time-point, USUV has been detected in several European countries often associated to episodes of avian mortality [4], [6]. There is also increasing evidence of virus circulation among horses and humans in Europe [7], [8], [9], [10] and recently two cases of neuroinvasive disease in humans have been documented [11], [12]. This current scenario reinforces the notion that USUV can infect humans and play a role as a pathogen capable to induce a broad spectrum of symptoms that range from fever, rash or jaundice to meningoencephalitis [5], [11], [12]. Albeit the number of cases of human USUV infections is rather limited, the similarities of USUV ecology with that of WNV emphasize the need to be cautious about its potential threat to human health [4], [5]. Even more, the observed symptoms of human USUV infections are not very specific, which could have probably led to an underestimation of the infections in endemic areas, which mainly comprise developing countries.
A detailed knowledge of the cellular processes involved in pathogen and host cell interactions is desirable to design effective strategies to combat arboviral diseases. In the case of USUV, the role of many aspects of the interaction between the virus and the host cell, for instance its relationship with the autophagic pathway, remains to be explored. Macroautophagy (thereafter referred as ‘autophagy’) is a cellular process by which cytoplasmic components are sequestered into double-membrane vesicles and degraded to maintain cellular homeostasis. In addition to this, autophagy constitutes an evolutionarily ancient process for survival during different forms of cellular stress, including infection with viruses [13], [14]. As a first line of defence against intracellular pathogens, autophagy can contribute to viral clearance through the degradation of viral components located in the cell cytoplasm [13]. But antiviral aspects of autophagy go beyond, and this catabolic route has been also implicated in both innate and adaptive immunity, i.e. by promoting the delivery of Toll-like receptor (TLR) ligands to endosomes, or by feeding antigens to MHC class II pathway [13], [15], [16]. Autophagy can also extend the survival of infected cells by limitation of apoptosis [17]. Conversely, some viruses can take advantage on the induction of autophagy by co-opting components from the autophagic machinery in their own benefit to provide the adequate cellular platforms for replication [18], [19], [20] or by rearrangement of cellular lipid metabolism in order to support strong viral replication [21]. All these features make of autophagy a relevant druggable metabolic pathway during multiple human disorders, including viral infections, so interventions on this route could constitute potential therapies [14], [16], [22].
Regarding the Flaviviridae, autophagy has been associated to different aspects of the replication and pathogenicity of some members of this virus family, including Dengue virus (DENV) [21], [23], [24], [25], Modoc virus [24], Japanese encephalitis virus (JEV) [26], and hepatitis C virus (HCV) [27], [28]. In the case of WNV, the induction or not of an autophagy response remains contentious. One recent report pointed that WNV infection induced an autophagic response [29], whereas another suggested that the autophagic pathway was not upregulated in WNV-infected cells [30]. Relative to USUV, to our knowledge, the involvement of the autophagic machinery during its replication has not been previously documented.
In this study we have analyzed the induction of autophagy following infection with a prototypic African strain of USUV and a recent European isolate [31]. The ability of both strains to provoke an autophagic response on infected cells was documented. Even more, pharmacological intervention at the autophagic pathway modulated USUV infection, thus identifying a cellular pathway for potential interventions on USUV infection.
All animals were handled in strict accordance with the guidelines of the European Community 86/609/CEE at the biosafety animal facilities of the Centro de Investigación en Sanidad Animal of the Instituto Nacional de Investigación Agraria y Alimentaria (CISA-INIA). The protocols were approved by the Committee on Ethics of Animal Experimentation of INIA (permit number 2013–015).
Mouse monoclonal antibody J2 against double-stranded RNA (dsRNA) was purchased from English & Scientific Consulting (Hungary). Rabbit monoclonal anti-LC3B, rabbit anti-p62/SQSTM1 and mouse monoclonal anti-β-actin antibodies were from Sigma-Aldrich (St. Louis, MO). Rabbit anti-calnexin antibody was from ECM Biosciences (Versailles, KY). Polyclonal serum from a mice experimentally infected with USUV SAAR 1776 (INIA permit number 2013–015) was also used to detect USUV proteins. Secondary antibodies against Mouse or Rabbit IgGs coupled to Alexa Fluor-488, -594 or -647 were purchased from Life Technologies (Molecular Probes, Eugene, O). Anti-rabbit and anti-mouse secondary antibodies coupled to horseradish peroxidase were from Dako and Sigma, respectively.
All manipulations of infectious virus were carried out in Biosafety level 3 (BSL-3) containment facilities. USUV strain SAAR 1776 (GenBank acc: AY453412.1), the reference South African strain of USUV, and the Austrian strain of USUV Vienna 2001-blackbird (USUV 939/01, GenBank acc: AY453411.1), a recent European isolate of USUV [31] were propagated five and three times, respectively, in Vero cells [32]. Viruses were used at a multiplicity of infection (MOI) of 5 PFU/cell in microscopy experiments and of 0.5 PFU/cell in the rest of experiments. Infections and virus titrations on semisolid agar medium were performed as previously described [33]. Cells were routinely tested for mycoplasma with Mycoalert Mycoplasma Detection Kit (Lonza, Rockland, ME).
Autophagy inhibitors 3-methyladenine (3-MA) and wortmannin, and autophagy inducer rapamycin, were purchased from Sigma and used at the concentrations of 2.5 mM, 0.5 µm and 100 ng/ml, respectively. Amonium chloride (NH4Cl, Merck) was used at 25 mM. Cells were infected, or mock-infected, and drugs were added to the medium after the first hour of infection. Stock solutions of wortmannin and rapamycin were prepared in dimethyl sulfoxide (DMSO), and DMSO was also used as control in non-treated cells (drug vehicle). Tunicamycin (Sigma), an inducer of unfolded protein response, was also dissolved in DMSO and used at 10 µg/ml. The viability of cells with or without treatment was tested with CellTiter-Glo Luminiscent Cell Viability Assay (Promega).
A plasmid encoding GFP-LC3 [34] was transfected to visualize autophagosome formation. Plasmid encoding mCherry-GFP-LC3 was used to detect acidified autophagosomal structures [35]. Fugene HD (Promega, Madison, WI) was used as transfection reagent according to the instructions provided by the manufacturer. Cells were infected or treated with the drugs 24 h post-transfection.
Assays were carried out as described [32]. Briefly, cells grown on glass cover slips were washed with PBS and fixed with 4% paraformaldehyde in PBS for 15 min at room temperature. Fixed cells were washed with PBS and permeabilized with BPTG (1% BSA, 0.1% TritonX-100, 1M glycine in PBS) for 15 min. Cells were incubated with primary antibody diluted in 1% BSA in PBS for 1 hour. After washing, cells were incubated with fluorescently conjugated secondary antibody for 45 min at room temperature. Samples were mounted with Fluoromount-G (SouthernBiotech, Birmingham, AL) and observed using a Leica TCS SPE confocal laser-scanning microscope. Images were acquired using Leica Advanced Fluorescence Software. Images were processed using ImageJ (http://rsbweb.nih.gov/ij/) and Adobe Photoshop CS2.
Vero cells infected with USUV (MOI of 5 PFU/cell) were washed and fixed 24 h p.i. (30 min at 37°C) in 4% paraformaldehyde-2% glutaraldehyde in 0.1 M phosphate buffer pH 7.4 plus 5 mM CaCl2. Cells were scrapped from the flasks and post-fixed in 1% osmium tetroxide-1% potassium ferricyanide (1 h at 4°C), washed three times with bidistilled water and treated with 0.15% tanic acid (1 min). Cells were washed with the buffer and with bidistilled water and stained with 2% uranyl acetate (1 h). Samples were washed and then dehydrated in ethanol and embedded in the resin. Samples were examined using a Jeol JEM-1010 electron microscope (Jeol, Japan) operated at 80 kV and images were acquired using a digital camera 4K×4K TemCam-F416 (Tietz Video and Image Processing Systems GmbH, Gauting, Germany).
Western blot were performed as reported [32]. Cells were lysed on ice in RIPA buffer (150 mM NaCl, 5 mM β-mercaptoethanol,1% NP-40, 0.1% sodium dodecyl sulfate [SDS], 50 mM Tris-HCl pH 8) supplemented with cOmplete protease inhibitor cocktail tablets (Roche, Indianapolis, IN) and Benzonase Nuclease (Novagen, EMD Chemicals, San Diego, CA). Protein concentration was determined by Bradford assay. Equal amounts of proteins were mixed with Laemmli sample buffer, subjected to SDS-PAGE and electrotransferred onto a nitrocellulose or a PVDF membrane. Membrane was blocked with 5% skimmed milk in PBS 0.05% Tween-20, incubated with primary antibodies (overnight at 4°C), washed three times with PBS-Tween, and subsequently incubated with secondary antibodies coupled to horseradish peroxidase (1 h at RT) diluted in 1% skimmed milk in PBS-Tween. Membrane was washed three times and proteins were detected by chemiluminiscence using a ChemiDocTM XRS+ System (Bio-Rad, Hercules, CA). Intensity of protein bands was quantified using ImageLab software 2.0.1 (Bio-Rad).
RNA was extracted from control and infected cells with TriPure Isolation Reagent (Roche, Mannheim, Germany) as indicated by the manufacturer. Reverse transcriptions PCR reactions (RT-PCR) were carried out using the SuperScript One Step RT-PCR (Life Technologies). Unspliced or spliced Xbp-1 mRNA was amplified as described [36]. Amplification of GAPDH mRNA was carried as a control for RNA extraction. PCR products were resolved by electrophoresis in a 2% agarose gel.
Data are presented as mean ± standard error of the mean (SEM). To test the significance of the differences, analysis of the variance (ANOVA) was performed with statistical package SPSS 15 (SPSS Inc, Chicago IL) applying Bonferroni's correction for multiple comparisons. Statistically significant differences were considered at P<0.05.
As a first approach, cells infected with USUV strain SAAR 1776 were analyzed by transmission electron microscopy. Infected cells exhibited morphological characteristics associated to flavivirus infection. These included electron dense virions located inside endoplasmic reticulum cisternae (Fig. 1A) and membrane enclosed groups of spherical vesicle-like structures (Fig. 1B). These clusters of vesicles correspond to vesicle packets (VPs) observed for WNV [32], [37], which have been also named double membrane vesicles (DMVs) in the case of DENV [38]. In addition to these classical features associated to flavivirus replication, an accumulation of cellular organelles morphologically related to those associated to degradative processes was patent in USUV-infected cells (Fig. 1C–F). Rearrangement of endoplasmic reticulum-derived structures wrapping around cytoplasmic material was observed (Fig. 1C), providing images typical of autophagic processes. Notably these membranes could be also observed continuous with the membrane of VPs. Membrane rearrangements included double membrane vacuoles (Fig. 1D, arrowheads) compatible with the morphology of autophagosomes. Double membranes engulfing cytoplasmic portions that resemble phagophore-like structures, which constitute the first stages of the formation of autophagosomes, were also observed (Fig. 1D, arrows). In addition to these, multi-lamellar structures constituted by smooth stacked membranes, which have been also associated to autophagic processes [39], were accumulated in the cytoplasm of USUV-infected cells (Fig. 1E), and even observed in association with double membrane vesicles (Fig. 1F, arrowheads). Taken together, the ultrastructural alterations observed in USUV-infected cells are compatible with the activation of an autophagic response.
Following induction of autophagy, microtubule-associated protein 1 light chain 3 (LC3), a mammalian homolog of yeast Atg8 (autophagy-related protein 8) is conjugated to phosphatidylethanolamine and targeted to autophagic membranes labelling autophagic vacuoles [34], [40]. This prompted us to analyze LC3 modification following infection with USUV. As controls, Vero cells treated in parallel with the inhibitor of autophagy 3-methyladenine (3-MA) or with the inductor of autophagy rapamycin [40] were analyzed by western blot (Fig. 2A). A significant increase in LC3-II was noticed on cells treated with rapamycin, whereas a significant decrease of LC3-II was observed in cells treated with 3-MA (Fig. 2A). Quantification of the LC3-II/actin ratio confirmed these observations (Fig. 2B). Interestingly, a significant increase in LC3-I was also noticed in cells treated with rapamycin, suggesting that rapamycin could not only promote LC3-I/II turnover whereas it could also result in accumulation of both LC3 species under our experimental conditions. As these results confirmed the reliability of western blot assays to detect alterations of the autophagic flux, the modification of LC3 was analyzed on cells infected with USUV (Fig. 2C). An increase in the amount of LC3-II was observed following infection with USUV when compared to mock-infected cells. This finding is compatible with alterations of the autophagic pathway in cells infected with USUV. Since PVDF membranes can result more sensitive to detect LC3-II than those of nitrocellulose [41], we performed similar analyses using these membranes. An increase in both LC3-I and LC3-II along time was observed in samples infected with USUV (Fig. 2D), which was confirmed by densitometry of protein bands (Fig. 2E), as commented above for cells treated with rapamycin. The accumulation of total LC3 (LC3-I and LC3-II) suggests that the formation of autophagosomes could be upregulated in USUV infected cells. Indeed, accumulation of total LC3 has been documented during DENV-induced autophagy [24]. As the increase in LC3-II was not accompanied by a decrease in LC3-I, this could indicate that the autophagic flux, and hence LC3-I/II turnover, is not upregulated in cells infected with USUV, or that expression of LC3 is stimulated by USUV infection. In fact, transcription of LC3 is increased in certain systems upon induction of autophagy [41]. To analyze if USUV infection altered the normal autophagic flux that involves degradation or turnover of autophagosomal proteins, the levels of other autophagy-related protein, p62/SQSTM1, were analyzed in USUV infected cells. The degradation of p62/SQSTM1, a polyubiquitin-binding protein that interacts with LC3 [35], has been described following upregulation of the autophagic flux under certain conditions [40]. However, no change of p62/SQSTM1 levels was found in cells infected with USUV (Fig. 2F), These results indicate that Vero cells infected with USUV presented an accumulation of both LC3-II and LC3-I, a finding that could suggest an accumulation of autophagosomes, apparently not associated to alteration of p62/SQSTM1 level.
As commented above, following induction of autophagy the lipidated form of LC3 is targeted to autophagic membranes labelling autophagic vacuoles [34], [40]. This feature can be monitored by fluorescence microscopy as an increase in LC3 puncta in the cell cytoplasm [34], [40]. Vero cells were transfected with a plasmid encoding GFP-LC3 for 24 h, and then infected with USUV (Fig. 3A). Cells were fixed at 24 h p.i. and samples were processed for immunofluorescence using an antibody specific for double-stranded RNA (dsRNA) -a well characterized marker of the flavivirus replication complex [32], [37], [38]- to verify that transfected cells were infected. As expected, no positive signal corresponding dsRNA was observed on mock-infected cells, whereas fluorescent spots could be detected in the cytoplasm of USUV-infected cells (Fig. 3A). As controls, transfected cells were also treated in parallel with 3-MA or with rapamycin. Cells treated with rapamycin displayed a spotted cytoplasm compatible with the formation of LC3 aggregates as a result of the upregulation of autophagy, whereas cells treated with 3-MA were apparently undistinguishable from control cells (Fig. 3A). The cytoplasmic aggregates displayed by cells infected with USUV suggest an upregulation of the autophagic pathway during USUV-infection. When the number of puncta corresponding to GFP-LC3 per cell was determined by fluorescence microscopy [40] (Fig. 3B), it was found that cells treated with 3-MA displayed a tendency to a reduction on the number of LC3 aggregates, although not statistically significant, whereas a statistically significant increase was observed in rapamycin treated cells, thus validating the reliability of this method to detect induction of autophagy. A statistically significant increase in the mean number LC3 aggregates in cells infected with USUV was also noticed. The extent of this increase was similar to that found for cells treated with rapamycin. Therefore, this analysis confirmed the accumulation of autophagosomes in cells infected with USUV.
It has been proposed that replication of several members of the Flaviviridae family may be based on autophagosome membranes labelled with LC3 [28], [42], although other studies do not support this view [21], [43]. Detailed observation of USUV-infected cells revealed that GFP-LC3 puncta did not colocalize with dsRNA positive spots that labelled USUV RNA replication sites (Fig. 4A), thus indicating that although USUV infection increased the formation of autophagosomes the viral replication did not take place directly on autophagic membranes. The localization of viral proteins in infected cells was analyzed using a specific polyclonal serum raised against USUV (Fig. S1). As commented for dsRNA, USUV proteins did not colocalize with GFP-LC3 aggregates (Fig. 4B), pointing that viral proteins were not associated with autophagic membranes. In contrast to this, dsRNA colocalized with calnexin (Fig. 4C), a maker from the endoplasmic reticulum, a finding consistent with previous observations pointing that replication of USUV is based on membrane structures derived from the endoplasmic reticulum [32]. The viral proteins stained with the polyclonal serum also partially colocalized with calreticulin (Fig. 4D), thus confirming the interaction of the endoplasmic reticulum and viral components during USUV infection.
Since GFP is acid-labile, it makes difficult to detect autophagosomal structures by fluorescence microscopy using GFP-LC3 once they have fused with endosomes or lysosomes [35]. To analyze if USUV-infected cells were enriched only on autophagosomal compartments that had not already fused with acidic compartments we used a tandem mCherry-GFP-tagged LC3 expression vector [35]. The mechanism of action of this construction is based on that GFP signal is reduced in an acidic environment, whereas mCherry is more stable [35], [40], [41]. In this way, colocalization of GFP and mCherry indicates a cellular compartment that has not fused with an acidic compartment (phagophore or autophagosome) whereas mCherry signal without GFP corresponds to an autophagosomal compartment that has fused with an endosome or lysosome (amphisome or autolysosome) [40]. To verify that this construction worked properly under the experimental settings, cells were transfected with mCherry-GFP-tagged LC3 plasmid and then treated with rapamycin (to promote autophagosome maturation) or with NH4Cl (to block the normal autophagic flux) (Fig. 5A) [41], [44]. When compared to control cells, an increase in the number of GFP puncta following rapamycin treatment was observed, however the increase in mCherry puncta was more marked (Fig. 5B), indicating that rapamycin promoted maturation of autophagosomal structures towards acidic organelles in which GFP fluorescence was lost although they retained mCherry fluorescence. In contrast, the impairment of organelle acidification and the blockage of the normal autophagic flux exerted by NH4Cl induced an accumulation of LC3 puncta positive for both GFP and mCherry (Fig. 5A and B) that indicated a reduction in the number of acidified autophagic structures (Fig. 5C). Cells infected with USUV displayed an accumulation of mCherry puncta that did not colocalize with GFP puncta (Fig. 5A, B and C). This indicates that at this time postinfection there is an accumulation of acidified autophagosomal structures in USUV-infected cells that correspond to amphisomes or autolysosomes. In addition to this, no colocalization between dsRNA and mCherry was found in these experiments indicating that neither autophagosomes nor autolysosomes are the places for replication of USUV (Fig. 5A). This is again consistent with the notion that replication of USUV is associated to the endoplasmic reticulum and not to structures of an autophagic origin.
The interaction of flavivirus with endoplasmic reticulum during viral replication can result in induction of endoplasmic reticulum stress [45], [46], [47]. To cope with this problem, flavivirus-infected cells can undergo a coordinated change in gene expression collectively known as unfolded protein response [45], [46], [47]. Since induction of the unfolded protein response can trigger an autophagic response [48], we analyzed if infection by USUV also activated the unfolded protein response. To this end, we monitored the splicing of Xbp-1 (X box binding protein 1) mRNA (Fig. 6A), which allows expression of the full length transcription factor Xbp-1 that upregulates transcription of multiple genes aimed to cope with endoplasmic reticulum stress and that has been detected as a common feature of unfolded protein response in flavivirus-infected cells [45]. Cells treated with tunicamycin, to pharmacologically induce the unfolded protein response, displayed an increase in the amount of spliced Xbp-1 not observed in control cells. Cells infected with USUV also displayed an increase in the amount of spliced Xbp-1 that was observed between 18 and 24 h p.i. The amount of spliced Xbp-1 detected in USUV-infected cells was comparable to that observed in tunicamycin treated cells (Fig. 6B). These results evidenced that infection with USUV shared in common with other flaviviruses the activation of the unfolded protein response.
Overall, our results pointed to an upregulation of the autophagic pathway in USUV-infected cells and, thus, the possibility of manipulating the autophagic pathway to modulate infection with USUV was addressed. Hence, to evaluate the potential of pharmacological modulation of autophagy as a candidate for antiviral approach design against USUV, the effect of two inhibitors of autophagy, 3-MA and wortmannin [49], and that of the inductor of autophagy rapamycin were analyzed [40]. First of all, the cellular viability under drug-treatments was determined (Fig. 7A). After 24 h of treatment with 3-MA, wortmannin, or rapamycin no major toxic effects on Vero cells were noticed, confirming the adequacy of these conditions for subsequent analyses. Treatment with either 3-MA or wortmannin resulted in a significant reduction of the virus yield of USUV virus (Fig. 7B). On the other hand, rapamycin induced a significant increase in the viral production of USUV. Taken together, these observations support an implication of the autophagic machinery on the replication of USUV and confirm that pharmacological intervention on the autophagic pathway modulates USUV infection.
The findings reported above were obtained using the prototypic USUV strain SAAR 1776 whose pathogenic capability has not been conclusively proven even in the birds. To analyze if the USUV circulating in Europe displayed similar interactions with the autophagic pathway, the USUV strain Vienna 2001 (a recent isolate of USUV which pathogenicity has been extensively proven at least in birds) was included in the study [6], [50]. Infection with USUV Vienna 2001 induced an increase in the levels of LC3-II and also LC3-I, as described for USUV SAAR 1776 (Fig. 8A and B). In addition to this, cells transfected with the plasmid encoding GFP-LC3 and infected with USUV Vienna 2001 also displayed a significant accumulation of GFP-LC3 aggregates throughout the cytoplasm compared to mock-infected cells (Fig. 8C and D). These fluorescent GFP-LC3 aggregates did not colocalize with dsRNA (Fig. 8C) as commented for cells infected with USUV SAAR 1776 (Fig. 4A). However, dsRNA colocalized with the endoplasmic reticulum marker calnexin (Fig. 8E) as described for USUV SAAR 1776 (Fig. 4C). Even more, USUV proteins detected using a specific mouse serum colocalized with calnexin (Fig. 8F) confirming the association of viral antigens of USUV Vienna 2001 with endoplasmic reticulum, as described for USUV SAAR 1776 (Fig. 4D). The effect of the inhibitors of autophagy 3-MA and wortmannin, and that of the inductor of autophagy rapamycin was also analyzed in parallel for USUV Vienna 2001 and SAAR 1776 (Fig. 8G). Treatment with 3-MA or wortmannin resulted in a significant reduction of the virus yield of USUV Vienna 2001 as well as USUV SAAR 1776. The extent of inhibition exerted by 3-MA was similar for both viral strains, whereas USUV Vienna 2001 was slightly less inhibited by wortmannin than USUV SAAR 1776. In contrast, rapamycin induced a significant increase in the viral production of both USUV strains. Taken together, these results indicate that the findings observed for USUV SAAR 1776 were shared by USUV Vienna 2001, a strain of USUV that is currently circulating in Europe with documented pathogenecity in birds.
The Flavivirus genus comprises more than 50 viral species that include well long known arthropod-borne pathogens as DENV, WNV, JEV, St. Louis encephalitis virus, Murray Valley encephalitis virus, Yellow fever virus or tick-borne encephalitis virus (http://www.ictvonline.org/virusTaxonomy.asp?version=2012). But this viral genus also contains other neglected viral pathogens of currently increasing interest. Among these recently considered potential threats is USUV, a flavivirus endemic from Africa that emerged in Europe during the last decade (see Introduction).
In addition to basic knowledge, characterization of cellular pathways involved in virus replication could help to identify novel therapeutic targets. In this regard, the interaction of USUV with the host cell almost remains as an unexplored field, so at this point the identification of cellular pathways that regulate USUV infection is desirable. In this study we have explored the possible interaction of USUV with the autophagic pathway during infection. Due to the availability of more suitable reagents to analyze the autophagic pathway in mammalian cells, we selected Vero cells for the analysis. In fact, Vero cells constitute a cell line widely used for the cultivation and titration of USUV. In this way, the interaction of USUV with the autophagic pathway in cells derived from bird or mosquito, the main natural hosts for USUV, remains to be further evaluated. Our results showed that infection in mammalian cells by either the reference South African strain of USUV (SAAR 1776) or a recent European strain (Vienna 2001) triggered an autophagic response in the host cell. This is consistent with findings obtained for other flaviviruses [21], [23], [24], [25], [26]. The autophagic response was characterized by an increase in the levels of both LC3-II and LC3-I, which correlated with the accumulation of autophagic structures in the cytoplasm of infected cells. Our results also showed an induction of Xbp-1 mRNA splicing following USUV infection. Xbp-1 mRNA splicing constitutes a marker of the induction of the unfolded protein response, a finding that has been related to autophagic process during the viral infection [51]. Regarding the characteristics of the autophagic response induced by USUV, no p62/SQSTM1 degradation was found in infected cells, a feature that has been also described for other flaviviruses [26], [29]. This finding together with the accumulation of both LC3-I and LC3-II could suggest an incomplete autophagic response that takes place without autophagosome maturation. However, the experiments performed with mCherry-GFP-LC3 plasmid revealed that USUV-infected cells were enriched in acidified autophagosomal structures, suggesting that at least a significant proportion of autophagic structures can maturate and fuse with acidified organelles in USUV-infected cells. These structures could include amphisomes or autolysosomes, whose morphology is compatible with those of multi-lamellar organelles observed by transmission electron microscopy [39]. Regarding the accumulation of both LC3-I and LC3-II, this has been described as a feature of DENV-induced autophagy [24], and was also observed for cells treated with rapamycin. Although this could result of reduced autophagosomal degradation, the detection of acidified autophagosomal structures in USUV-infected suggests that autophagosomes can maturate in USUV-infected cells. However, another non-excluding possibility is an increase of expression of LC3 following infection of USUV, since such increase by a mechanism involving the unfolded protein response has been documented [41], and this response is also activated following USUV infection. According with this possibility, the increase in other cellular proteins involved in autophagy, as p62/SQSTM1, during infection with the related flavivirus WNV has also been proposed [29].
There is a controversy related to the autophagic origin or not of the structures that provide the platform for replication of distinct members of the Flaviviridae family as DENV or HCV. While several studies have been pointed that viral replication may be based on membranes of autophagosomal origin that contain LC3 [28], [42] other studies clearly contradict these results [21], [43]. In USUV-infected cells, no major colocalization between LC3 containing structures and dsRNA (a well characterized marker of the flavivirus replication complex) was found. In addition to this, no colocalization was found between USUV proteins and LC3, indicating that viral proteins were not associated with autophagic structures, and suggesting that these structures did not provide the main platform for viral replication. In fact, these results agree with data pointing that USUV replication, as well as those of WNV or DENV [37], [38], is mainly based on modified endoplasmic reticulum structures [37], [38]. Even more, dsRNA in USUV infected cells colocalized with calnexin, a marker of the endoplasmic reticulum. In this way, USUV replication most probably would take place associated to the vesicle packets (VPs), which in other viral models have been shown to contain the dsRNA intermediates and have been probed to be constituted by invaginations of endoplasmic reticulum-derived membranes [37], [38]. Even more, the observed colocalization between calnexin and USUV proteins confirms the interaction of viral components with the endoplasmic reticulum.
Autophagy constitutes a major metabolic pathway that is currently being explored for treatment of multiple human disorders that include certain types of cancer and metabolic diseases, neurological disorders or viral infections [14], [16], [22]. In this regard, drugs that interfere with the autophagic pathway were assayed. The inductor of autophagy rapamycin increased virus yield of both USUV strains analyzed. In contrast to this, two structurally unrelated inhibitors of phosphatidylinositol 3-kinases (PI3Ks) involved in the induction of autophagy (3-MA and wortmannin) decreased virus yield of both USUV strains here analyzed, including the European USUV isolate Vienna 2001, representative of the USUV that is currently circulating in Europe. Interestingly, whereas 3-MA inhibited both viral strains in a similar manner, infection by USUV Vienna 2001 was less inhibited by wortmannin than that of USUV SAAR 1776. These observations point to the autophagic pathway as a novel partner of USUV infection and specifically point to PI3Ks as valid antiviral targets. Having in mind that different PI3Ks are under strict consideration as cellular targets for treatment of human disorders [52], the results here presented set a starting point for antiviral development to combat USUV based on inhibition of these cellular enzymes. Other previously identified cellular pathways as regulators of USUV infection have been the synthesis of fatty acids [32], the innate immune response induced in infected cells [53], and preliminary data also point to the induction of apoptosis in infected cells [32]. In fact, connections between autophagy and these other metabolic pathways involved on USUV infection have been also documented for other members of the Flaviviridae family [21], [24], [54]. In this way, the identification of the involvement of autophagy during USUV infection will help to decipher the puzzle of the interaction of USUV with host cells.
Overall this study provides the first evidence for a role of autophagy during the infection of the mosquito-borne USUV. Our results indicate that pharmacological inhibition of the autophagic pathway can reduce infection by this virus in cultured cells. These observations identify autophagy as a metabolic pathway involved on USUV-infection, thus opening a potential new research line for the design of antiviral therapies against this pathogen.
|
10.1371/journal.pmed.1002193 | Predictors of Chemosensitivity in Triple Negative Breast Cancer: An Integrated Genomic Analysis | Triple negative breast cancer (TNBC) is a highly heterogeneous and aggressive disease, and although no effective targeted therapies are available to date, about one-third of patients with TNBC achieve pathologic complete response (pCR) from standard-of-care anthracycline/taxane (ACT) chemotherapy. The heterogeneity of these tumors, however, has hindered the discovery of effective biomarkers to identify such patients.
We performed whole exome sequencing on 29 TNBC cases from the MD Anderson Cancer Center (MDACC) selected because they had either pCR (n = 18) or extensive residual disease (n = 11) after neoadjuvant chemotherapy, with cases from The Cancer Genome Atlas (TCGA; n = 144) and METABRIC (n = 278) cohorts serving as validation cohorts. Our analysis revealed that mutations in the AR- and FOXA1-regulated networks, in which BRCA1 plays a key role, are associated with significantly higher sensitivity to ACT chemotherapy in the MDACC cohort (pCR rate of 94.1% compared to 16.6% in tumors without mutations in AR/FOXA1 pathway, adjusted p = 0.02) and significantly better survival outcome in the TCGA TNBC cohort (log-rank test, p = 0.05). Combined analysis of DNA sequencing, DNA methylation, and RNA sequencing identified tumors of a distinct BRCA-deficient (BRCA-D) TNBC subtype characterized by low levels of wild-type BRCA1/2 expression. Patients with functionally BRCA-D tumors had significantly better survival with standard-of-care chemotherapy than patients whose tumors were not BRCA-D (log-rank test, p = 0.021), and they had significantly higher mutation burden (p < 0.001) and presented clonal neoantigens that were associated with increased immune cell activity. A transcriptional signature of BRCA-D TNBC tumors was independently validated to be significantly associated with improved survival in the METABRIC dataset (log-rank test, p = 0.009). As a retrospective study, limitations include the small size and potential selection bias in the discovery cohort.
The comprehensive molecular analysis presented in this study directly links BRCA deficiency with increased clonal mutation burden and significantly enhanced chemosensitivity in TNBC and suggests that functional RNA-based BRCA deficiency needs to be further examined in TNBC.
| Identifying chemosensitive triple negative breast cancers (TNBCs) could significantly impact the survival of patients with these difficult to treat cancers until novel targeted therapies become available.
We hypothesized that genomic somatic aberrations may provide important molecular clues about chemosensitivity in TNBC.
Our study used a carefully selected cohort of 29 uniformly treated TNBC patients who either achieved pathologic complete response (pCR) or had extensive residual disease after neoadjuvant anthracycline/taxane chemotherapy.
We sequenced the coding genomic DNA of TNBC tumors and compared the somatic mutations found in the two groups at the two extremes of the chemosensitivity spectrum.
Our analysis revealed that, although mutations in single genes were not individually predictive, TNBC tumors bearing mutations in genes involved in the androgen receptor (AR) and FOXA1 pathways were much more sensitive to chemotherapy.
We also found that mutations that lowered the levels of functional BRCA1 or BRCA2 RNA were associated with significantly better survival outcomes; we derived a BRCA deficiency signature to define this new, highly chemosensitive subtype of TNBC.
BRCA-deficient TNBC tumors have a higher rate of clonal mutation burden, defined as more clonal tumors with a higher number of mutations per clone, and are also associated with a higher level of immune activation, which may explain their greater chemosensitivity.
Mutations in the AR/FOXA1 pathway provide a novel marker for identifying chemosensitive TNBC patients who may benefit from current standard-of-care chemotherapy regimens.
The newly defined RNA-based BRCA-deficient subtype includes up to 50% of the TNBC tumors that appear to be immune primed, and it would be of interest to investigate combinations of chemotherapy with immunotherapies, which could provide clinical benefit for these patients.
Although our study showed concordant results in three different datasets, our key findings need to be further validated in a larger, prospectively designed study with archival samples.
| Triple negative breast cancer (TNBC) disproportionately affects younger women and women of African ancestry, contributing to health disparities. In the era of personalized cancer therapy, patients with TNBC remain at considerably higher risk of relapse and death than patients with other breast cancer subtypes, due to the aggressive nature of TNBC and the lack of newer targeted therapies [1,2]. TNBC patients typically receive chemotherapy with anthracycline and cyclophosphamide followed by taxane (anthracycline/taxane [ACT]) as standard-of-care treatment. Approximately one-third of patients achieve pathologic complete response (pCR) and have excellent survival, but the remaining patients relapse and eventually die of the disease [3–5]. Identifying those TNBC patients who might benefit from ACT chemotherapy and directing the remaining patients to novel targeted therapies may be an effective strategy with near-term clinical impact for managing TNBC.
Transcriptional signatures developed in the past decade to predict sensitivity to ACT chemotherapy in TNBC have had only partial success [6–8], in part owing to the extensive molecular heterogeneity of the disease [9,10]. A few studies have evaluated whether predictability can be improved by considering a tumor’s somatic genetic aberrations alone [11,12] or in combination with gene expression [13]. Recent next-generation sequencing efforts have identified genes recurrently mutated in TNBC, including TP53 and PIK3CA, but unfortunately have not yielded any new predictive or prognostic clues [12,14]. Among existing markers of chemosensitivity, BRCA germline mutation carriers are known to receive greater benefit from platinum-based chemotherapy [15] and poly(ADP-ribose) polymerase (PARP) inhibitors in TNBC and ovarian cancers [16–18], but it is unclear whether patients with these tumors also benefit from ACT chemotherapy. Furthermore, higher prevalence of tumor-infiltrating lymphocytes (TILs) has been associated with better prognosis [19–22], irrespective of the chemotherapy administered, and also with a higher rate of pCR to neoadjuvant anthracycline-based chemotherapy in TNBC [23]. Gene expression signatures that capture immune or stromal characteristics have shown promising prognostic performance [24,25].
Chemosensitive or chemoresistant phenotypes can arise in genomically heterogeneous cancers through diverse molecular mechanisms, resulting in weaker biomarker–phenotype associations at the population level and confounding the discovery of prognostic and predictive biomarkers associated with response. Moreover, mutations alone may not be generally predictive, as gene expression levels are also modulated through non-genetic mechanisms, such as epigenetic silencing, aberrant transcription, and allele-specific expression [26–28]. Broader tumor genomic metrics that capture the extent and diversity of genetic heterogeneity within single tumors, such as overall mutation load and clonality, could be promising biomarkers and have been reported to be associated with patients’ clinical outcome in melanoma and in head and neck cancers [29,30]. There is also renewed interest in the enhanced innate immune response triggered by neoantigens from mutated cancer cell DNA, especially in tumors with mismatch repair deficiency [31], a mechanism that may provide a potential link between BRCA-deficiency-related chemosensitivity and the protective effect of TILs in TNBC.
In this study, we present a comprehensive assessment of chemosensitivity and resistance to ACT in TNBC using whole exome sequencing (WES) to identify tumor genomic aberrations that are potentially predictive of response. We used a TNBC cohort from the MD Anderson Cancer Center (MDACC) consisting of both extremely sensitive and highly recalcitrant tumors to identify mutations in specific genes or pathways that may indicate response or resistance to ACT chemotherapy in TNBC, and validated our findings in a larger TNBC cohort from The Cancer Genome Atlas (TCGA) project. We used integrated whole exome, DNA methylation, copy number variation, and RNA sequencing (RNAseq) data to expand the definition of TNBC subgroups associated with better outcome.
After DNA extraction from the fine needle aspiration biopsies, 1 μg of genomic DNA was sheared to a mean fragment length of about 140 bp. A NimbleGen human solution-capture exome array (SeqCap EZ v2) was used to capture the exomes of tumor samples, using a procedure modified from the manufacturer’s instructions. The library was sequenced on an Illumina HiSeq 2000 platform in paired-end 75-cycles mode at the Yale Center for Genome Analysis.
Reads were filtered by Illumina CASAVA 1.8.2 software, trimmed at the 3′ end using FASTX v0.0.13, and aligned to the human reference genome (GRCh37) by Burrows-Wheeler Aligner v0.7.5a, and PCR duplicates were removed using the MarkDuplicates (Picard) algorithm. Local realignment around putative and known insertion/deletion (INDEL) sites and base quality recalibration were performed using RealignerTargetCreator (Genome Analysis Toolkit v3.1.1). MuTect v1.1.4 and Strelka v1.0.14 were used to call somatic single nucleotide variants and INDELs, respectively, with an in-house pooled normal reference obtained from ten normal blood DNA samples sequenced using the same protocol. Mutations in the Single Nucleotide Polymorphism Database (dbSNP build 138; https://www.ncbi.nlm.nih.gov/SNP/; variants not flagged as somatic or clinical or as having a minor allele frequency of <1%), the NHLBI Exome Sequencing Project (ESP6500; http://evs.gs.washington.edu/evs_bulk_data/ESP6500SI-V2-SSA137.GRCh38-liftover.snps_indels.vcf.tar.gz), the 1000 Genomes Project (http://www.1000genomes.org/), and Exome Aggregation Consortium dataset (ExAC release 0.1; ftp://ftp.broadinstitute.org/pub/ExAC_release/release0.1) were excluded as putative germline sequence alterations. Furthermore, mutations were also excluded if the ratio of the mutant allele frequency (MAF) in tumor versus the pooled normal was less than five or if the MAF was 0.45–0.55. Recurrent COSMIC (v64) variants (n ≥ 5) and ClinVar annotated variants (http://www.ncbi.nlm.nih.gov/clinvar/) were whitelisted. All mutations found to have significant association with clinical outcomes were manually visualized in the Integrative Genomics Viewer (https://www.broadinstitute.org/igv/) to filter potential false positive calls. Raw WES data and mutation calls from the MDACC TNBC cohort are deposited in the Sequence Read Archive (accession ID SRP063902; http://trace.ncbi.nlm.nih.gov/Traces/sra/?study=SRP063902).
The R package SciClone [35] was used to infer tumor clonality by clustering variants of similar MAF from a tumor sample. We only selected mutations with at least 10-fold coverage and MAF < 0.6 to exclude mutations that may involve copy number loss. Variants were clustered using a Bayesian binomial mixture model in SciClone with each cluster representing a separate clone in the tumor. The average number of mutations per clone in each sample was calculated as the weighted sum of the number of mutations in each clone multiplied by the clonal proportion estimated by SciClone.
The trinucleotide loadings for four mutational signatures previously identified in breast cancer, Signature.1B (age associated), Signature.2 (APOBEC), Signature.3 (BRCA), and Signature.6 (mismatch), were downloaded from a previous study [32]. We applied non-negative least squares to estimate the proportion of each signature in each sample, and the signature with the greatest estimated coefficient was designated the dominant signature.
The association between mutational status (at the gene or pathway level) and pCR rate or overall survival was assessed using the Fisher exact test and the Kaplan-Meier survival estimator, respectively. Wilcoxon rank tests were used to compare characteristics (MR, MATH score, number of neoantigens, and mutation signature) between groups (pCR versus RD, BRCA-D versus BRCA-N).
We sequenced the genomic DNA from 29 cases in two response groups (pCR, n = 18; RD, i.e., chemoresistant with moderate or extensive residual cancer burden [38], n = 11) using WES (mean nucleotide coverage 150×; more than 90% of target bases had >20× coverage in all samples). Most detected somatic mutations were not recurrent, and only the MR of TP53 was significantly above background across all samples (false discovery rate < 0.1 using MutSigCV [34]). Twenty-two of 29 tumors (76%) carried non-silent TP53 mutations (S1 Fig), but there was no evidence of association with chemosensitivity (Fisher exact test, p > 0.5).
Functional mutations in nine canonical biological pathways were associated with chemotherapy response (Fisher exact test, p < 0.05; S1 Table). Due to the small size of the response groups, we applied a bootstrap strategy to evaluate the robustness of these associations under resampling (S2 Fig), and a permutation strategy to assess their significance (S1 Table). The top two pathways, “regulation of androgen receptor activity” and “FOXA1 transcription factor network,” remained significant, with mutations in both pathways being associated with pCR. Considering their substantial overlap (14 genes in common out of 59 in the androgen receptor [AR] pathway and 58 in the FOXA1 pathway), we merged the two pathways into the “AR- and FOXA1-regulated network.” Tumors carrying mutations in the AR/FOXA1 pathway had a significantly higher pCR rate (94.1% compared to 16.6% in tumors without such mutations, q = 0.02 after Bonferroni correction; Fig 2B). Furthermore, functional mutations occurred almost exclusively in chemosensitive tumors (21/22, or 95% of mutations) except for a single truncating BRCA1 mutation found in one RD tumor with AJCC stage IIIB cancer but minimal residual cancer burden (Fig 2A; Table 1, sample id 757_004_004). At the gene level, 13 genes had at least one mutation affecting 17 chemosensitive patients (58.6%) across the entire cohort (Fig 2A). In most cases, different genes were mutated in individual tumors, and despite the strong association observed at the pathway level, none of these genes individually had a significant association with pCR. Among them, BRCA1 was the most frequently mutated gene (17%), and four of the mutations observed in the pCR cohort were associated with hereditary breast cancers in the ClinVar database (S3 Fig). The BRCA1 mutation found in the tumor with RD (Fig 2A; see also S3 Fig) is a stop-gain mutation also reported in ClinVar as pathogenic, but due to lack of RNAseq data we could not assess the relative expression of mutant and WT BRCA1 transcripts.
To assess whether broad genomic measures that capture the overall burden and heterogeneity of somatic mutations are predictive of chemosensitivity, we calculated the MR and MATH score in each TNBC tumor. We also estimated tumor clonality by applying clonal decomposition to the somatic mutation profile of each tumor, and confirmed a positive correlation between the MATH score and the estimated number of clones in each tumor (S4 Fig). We sought to combine MR and MATH into a composite score, which we called CMB, that captures both the clonality of a tumor and the number of somatic mutations per clone. MR and MATH scores were median-dichotomized in the cohort, and CMB categories were defined as low (low MR, high MATH), high (high MR, low MATH), or intermediate (all others). There was an increasing tendency in the average number of mutations per clone from low to high CMB (S5 Fig). The pCR rate in the low, intermediate, and high CMB tumors was 33%, 64%, and 89%, respectively (Fig 2C), suggesting that tumors with a high number of mutations per clone (small number of clones, high MR) have significantly better response (p = 0.05) than low CMB tumors, which may be subclonal or have an overall low MR. The CMB categories were not significantly associated with other clinical or pathologic characteristics such as tumor stage or grade, or with patient’s age.
To validate our findings, we selected TCGA TNBC samples based on IHC status and PAM50 subtype, excluded stage IV cancers and cases with short follow-up, and downloaded available processed data of WES, RNAseq, DNA methylation, and copy number variation (Fig 1; S2 Table).
Among the 102 TNBC cases from the TCGA cohort with available exome sequencing data, 19 had at least one functional somatic mutation in the AR/FOXA1 pathway and 35 had at least one functional somatic mutation in the AR/FOXA1 pathway or at least one germline BRCA1 mutation. As with the MDACC cohort, BRCA1 was most frequently mutated among this gene set, with 21 (20%) patients carrying germline mutations and two (2%) carrying somatic mutations (Fig 3A). Patients with at least one mutation in one of the genes in this pathway had excellent overall survival, while those with WT genes had significantly worse outcomes (p = 0.028; Fig 3B). We observed a similar trend when we considered only somatic mutations (S6 Fig).
We computed the MR and MATH score in 101 TCGA TNBC cases with available exome sequence alignment data. Also in this cohort, the MATH score was positively correlated with the estimated number of clones (R = 0.44, p < 0.001; S7 Fig). Cases were stratified into three categories of CMB using the same criteria as with the MDACC cohort. Tumors with high CMB harbored a significantly higher average number of mutations per clone than tumors with low CMB (S8 Fig), and, consistent with the MDACC cohort, patients with these tumors had a significantly improved overall survival rate (p = 0.029) compared to patients with tumors with low CMB (Fig 3C).
BRCA1 was the most frequently mutated gene among the gene set that we found to be associated with chemosensitivity: it was mutated in about 20% of TNBC cases in both cohorts. Loss of function variants in BRCA1 or BRCA2 genes lead to homologous recombination defects and may contribute to sensitivity to platinum-based chemotherapy, but it is yet unclear whether this deficiency is associated with improved benefit from standard-of-care ACT chemotherapy. Besides germline and somatic mutations, deletions or epigenetic silencing can also result in DNA repair deficiency in a large proportion of TNBC cases. Here we systematically evaluated BRCA deficiency in TNBC using integrated analysis of DNA sequencing, DNA methylation, and RNAseq data from TCGA and assessed its association with clinical outcome.
Among the 101 TNBCs from the TCGA dataset, 21 tumors (20.5%) had inactivating germline SNP or somatic mutations in BRCA1, four (3.9%) in BRCA2, and two (1.9%) in both (Fig 4A and 4B). One hotspot germline SNP (rs1799950; Q356R in exon 10) was particularly common in this cohort, appearing in 15 cases (S3 Fig). Furthermore, we assessed the copy number variation in BRCA1/2 and promoter hypermethylation in BRCA1. Although different types of abnormalities are associated with BRCA1/2 inactivation, they all result in low expression of functional WT BRCA transcript. In mutation carriers, WT transcript abundance was determined from the overall expression, and MAF from RNAseq data. As expected, the abundance of WT BRCA1/2 transcripts was significantly lower in mutation carriers than in non-carriers (p < 10−16; Fig 4A and 4B). We therefore defined the BRCA1/2 deficiency threshold as the maximum level of WT BRCA1/2 transcript expressed in BRCA1/2 mutation carriers. That threshold is indicated by an arrow in Fig 4A and 4B. Tumors with WT BRCA1/2 transcript abundance below the corresponding threshold were classified as BRCA-D. This includes all the cases with germline or somatic mutations, or deep loss, or epigenetic silencing of the BRCA genes.
Based on this expanded definition of BRCA deficiency, 48 TNBC cases (47%) were characterized as BRCA-D. Specifically, 43 cases (42%) were BRCA1 deficient, nine (9%) were BRCA2 deficient, and four (4%) were deficient in both (Fig 4C). This definition captured all the cases with BRCA mutation, BRCA1 promoter methylation, or BRCA deletion, but also an additional nine cases (9%) that expressed low levels of WT BRCA transcripts for unknown reasons. We applied non-negative linear regression to 76 tumors with available mutation context data to estimate the proportion of mutations explained by the BRCA, age, APOBEC, and mismatch repair mutational signatures in each sample. The BRCA signature was dominant in 78.8% BRCA-D tumors (p < 0.001), the age signature was dominant in BRCA-N tumors (p = 0.004), and the APOBEC signature was more often present in BRCA-N tumors (14%) than in BRCA-D tumors (p = 0.03; Fig 4C). Furthermore, BRCA-D tumors were associated with significantly higher MRs based on non-silent (p < 0.001) or silent (p = 0.02) mutations, but had clonal heterogeneity (p = 0.55) similar to that of BRCA-N TNBC tumors (S9 Fig).
We found that expression of WT BRCA transcripts defines a new TNBC subtype (BRCA-D) characterized by high MR but typical clonal heterogeneity. Moreover, TNBC tumors with high CMB were highly enriched in BRCA-D tumors (67%) compared to tumors of low CMB (7%) (p < 0.001; Fig 5A), suggesting that BRCA1/2-mediated homologous recombination deficiency is associated with tumors of high CMB that were extremely sensitive to ACT chemotherapy in the MDACC cohort (Fig 2C).
Patients with BRCA-D tumors had 100% 4-y overall survival, compared to 79.5% (95% CI 66.6%–94.9%) for BRCA-N tumors (log-rank test, p = 0.018; Fig 5B), in the TCGA TNBC cohort. To further validate this finding in an independent dataset of TNBC cases without requiring RNA and DNA sequencing data, we developed a gene expression signature that predicts BRCA-D status in the TCGA dataset. We identified 24 genes that were strongly overexpressed and 26 genes that were underexpressed in BRCA-D compared to BRCA-N cases (arbitrary cutoff, unadjusted p < 0.002; S10 Fig). Among the 50 genes identified (S3 Table), BRCA1 was the gene most strongly associated with BRCA deficiency (p < 0.001). For each sample, a BRCA deficiency signature score was computed as the mean expression of the 24 overexpressed genes minus the mean expression of the 26 underexpressed genes. The score was median-dichotomized to predict BRCA-D status if high or BRCA-N status if low. The predictor of BRCA deficiency status had a sensitivity of 85.4% and specificity of 81.5% for predicting BRCA deficiency status in the TCGA cohort where it was developed. Furthermore, the patients with predicted BRCA-D tumors had significantly better overall survival than patients with predicted BRCA-N tumors (log-rank test, p = 0.013; Fig 5C). We applied the BRCA deficiency signature to an independent cohort of 278 chemotherapy-treated TNBC cases from the METABRIC cohort, which validated that patients with tumors predicted to be BRCA-D by the gene signature had significantly better overall survival compared to patients with tumors predicted to be BRCA-N (log-rank test, p = 0.009; Fig 5D).
In our analysis we discovered that BRCA deficiency is associated with higher CMB and that both BRCA deficiency and high CMB are predictive of chemosensitivity in TNBC. Given that a high prevalence of TILs has often been associated with better response in TNBC, we wanted to further investigate whether high CMB and BRCA deficiency are associated with higher immune activation. In the TCGA TNBC cohort, we observed a strong correlation between the overall MR and the number of predicted neoantigens (Spearman correlation = 0.76; Fig 6A). Due to higher overall MR, BRCA-D tumors had a significantly greater number of predicted neoantigens compared to BRCA-N tumors (p = 0.003; S11 Fig). Additionally, to estimate immune cell prevalence we used the average expression of the lymphocyte-specific genes GZMB, PRF1, CXCL13, IRF1, IKZF1, and HLA-E in each tumor sample [39]. Interestingly, tumors with predominantly clonal mutations were associated with higher immune presence compared to those with subclonal mutations (p = 0.003; Fig 6B), implying a negative association between clonal heterogeneity and immune response (S12 Fig). Therefore, tumors with high CMB harbor a greater number of predicted neoantigens per clone (Fig 6C), which elicit a higher immune response (Fig 6D). In summary, our analysis suggests a connection between BRCA deficiency status and high CMB, which results in a greater number of clonal neoantigens, leading to immune activation and potentially mediating enhanced response to ACT chemotherapy in TNBC.
We report results from an integrated genomic analysis of a TNBC cohort deliberately selected to represent extremely chemosensitive tumors and tumors highly resistant to standard-of-care ACT chemotherapy. Although no significant associations were identified between recurrent functional somatic mutations in specific genes and chemotherapy response, aggregating at the pathway level revealed that mutations occurring in two pathways, “regulation of androgen receptor activity” and “FOXA1 transcription factor network,” were significantly associated with pCR in TNBC (94% pCR rate in tumors with mutated pathways versus 17% in tumors without such mutations). Furthermore, TNBC patients from the TCGA cohort whose tumors had at least one mutation in the above pathways had excellent survival, with no deaths observed in 4 y when treated with ACT-containing regimens.
TNBC is highly heterogeneous, and up to six different subtypes have been recognized by transcriptional profiling, each associated with different clinical outcomes and responses to therapy [40,41]. One subtype, the luminal androgen receptor (LAR) subtype, is characterized by luminal gene expression driven by the AR and is generally associated with low response to chemotherapy [40]. The AR is expressed in about 10%–40% of TNBCs, but its role in prognosis or as a potential therapeutic target in TNBC has remained controversial [42]. In TNBC, signaling through the AR is hypothesized to mimic ER signaling, initiating transcriptional activation that promotes cell growth through the involvement of the transcription factor FOXA1 [43]. This has provided a justification for targeting AR in AR-positive TNBC. Recent single-arm phase II studies that evaluated the effect of the AR antagonists bicalutamide and enzalutamide in metastatic AR+ TNBC reported 6-mo clinical benefit rates of 19% and 29%, respectively [44,45], suggesting a direct role of AR in this TNBC subtype. Our results are consistent with the above observations, suggesting that mutations in the AR/FOXA1 pathway could result in abrogation of AR-related signaling, resulting in improved sensitivity to standard chemotherapy and better overall survival.
TNBC tumors are characterized by broad genomic and transcriptional heterogeneity [10,14]. The extent of genomic and transcriptional heterogeneity in tumors appears to be associated with resistance to chemotherapy in TNBC [10] and with worse prognosis in head and neck cancer [30]. Furthermore, high somatic mutation load has been linked to favorable outcomes in pancreatic cancer [46] but to worse prognosis in ER-positive breast cancer [47]. In our assessment of broad genomic measures as potential predictors of chemosensitivity, we found that patients with tumors with high CMB, defined as high mutational load but low clonality or a high number of somatic mutations per clone, have a significantly higher pCR rate (89%) and excellent survival (no deaths in 4 y in the TCGA cohort) compared to patients with tumors with low CMB (pCR rate of 33%). Therefore, the clonality of a tumor appears be critical in determining chemotherapy sensitivity and survival outcome. Tumors with high mutation load, for instance due to defective DNA damage response pathways, are sensitive to chemotherapy provided that they are not subclonal, that is, they do not contain mutations of lower variant allele frequencies that would have originated from subclones arising later in the tumor’s clonal expansion. Subclonal tumors contain not only the clonal mutations that were present in the founding cell but also subclonal mutations that emerged in subsequent clones during clonal expansion and thus exhibit broader genetic heterogeneity, which contributes to resistance to chemotherapy [48–50].
BRCA1 and BRCA2 are critical for the process of DNA repair by homologous recombination repair (HRR), and deficient HRR makes cancers more susceptible to DNA-damaging agents. Familial BRCA1 or BRCA2 mutant breast tumors tend to have a TNBC phenotype and often exhibit extreme levels of genomic instability [51]. The “BRCAness” phenotype is more broadly defined as defective HRR, driven not only by germline BRCA1 and BRCA2 mutations but also by somatic mutations or other alterations in these or other genes involved in HRR [52]. Indeed, ovarian tumors with BRCA1 or BRCA2 mutations, either germline or somatic, are associated with higher mutational burden and better survival outcomes following treatment with platinum-based chemotherapy [53] or PARP inhibitors [54]. In TNBC, germline BRCA1 mutation carriers were found to have higher pCR rates to neoadjuvant ACT chemotherapy compared to non-carriers (46% versus 22%) and significantly better survival outcomes [55]. Similar results were reported in TNBC with promoter methylation of BRCA1, where BRCA1 methylation was associated with better survival outcomes following adjuvant ACT chemotherapy [56]. Our broad definition of BRCA deficiency from RNAseq data based on the level of WT BRCA transcripts incorporates the effect of all genomic aberrations leading to inadequate levels of functional BRCA. This broader definition of BRCA deficiency included 46% of the TNBC tumors in the TCGA cohort; this subset had an excellent survival outcome (100% 4-y survival) following adjuvant ACT chemotherapy. Although the evidence provided by the survival data in the TCGA cohort is somewhat limited due to shorter follow-up and a lower number of deaths observed than expected for TNBC, the effect of BRCA deficiency was confirmed in the METABRIC cohort, which has 10 y of follow-up and more representative overall survival for TNBC. Our analysis therefore suggests that a definition of BRCA deficiency based on RNAseq could be a clinically useful biomarker of chemosensitivity for TNBC.
Immunotherapy is now emerging as a potentially viable therapeutic option for TNBC patients [57], but this treatment is expected to be effective only for a subset of the patients. Antibodies against programmed death 1 (PD-1) were significantly more effective in mismatch-repair-deficient colorectal cancers, most likely due to an increased number of neoantigens in these tumors [31]. We observed that BRCA-D TNBC tumors are characterized by high CMB and carry a greater number of predicted neoantigens that tend to be clonal. This could be the reason for the higher level of immune infiltration observed in these tumors [58] and may have contributed to the improved response to ACT chemotherapy that we observed. These results suggest that the combination of immunotherapies with ACT chemotherapy or PARP inhibitors might be an effective strategy for treating BRCA-D tumors. This is currently being evaluated in a phase I/II study in BRCA-D ovarian cancer (ClinicalTrials.gov; study ID NCT02571725).
In summary, we have provided an integrated characterization of the chemotherapy response phenotypes in TNBC. The strong connection of ACT chemosensitivity and immune activity with a new transcriptionally defined BRCA-D phenotype could help inform future therapeutic strategies for TNBC patients.
Limitations of our single-institution retrospective study include the small size of the discovery cohort and potential selection bias as samples were included based on both chemotherapy response and availability of residual biopsy materials for DNA isolation. Given the genomic heterogeneity of TNBC, this might limit the generalizability of our results. Another limitation is the lack of tumor-matched normal DNA for these tumors, which may result in reduced sensitivity and specificity for detecting somatic mutations in this cohort. Although our observation that mutations in AR/FOXA1 genes are associated with better outcomes in ACT-treated TNBC patients was validated in the TCGA cohort, the low number of events observed in this cohort limits the power of the analysis. Yet, our key finding that BRCA-D TNBC tumors identified by the BRCA deficiency signature are indeed associated with better outcomes after chemotherapy was confirmed in both the TCGA and METABRIC datasets. Although these findings will require validation in larger multi-institutional datasets, preferably originating from prospective clinical studies, they could provide the impetus for examining BRCA deficiency in TNBC in the context of increased CMB, with potentially improved response to immunotherapies.
|
10.1371/journal.ppat.1002385 | Two Novel Transcriptional Regulators Are Essential for Infection-related Morphogenesis and Pathogenicity of the Rice Blast Fungus Magnaporthe oryzae | The cyclic AMP-dependent protein kinase A signaling pathway plays a major role in regulating plant infection by the rice blast fungus Magnaporthe oryzae. Here, we report the identification of two novel genes, MoSOM1 and MoCDTF1, which were discovered in an insertional mutagenesis screen for non-pathogenic mutants of M. oryzae. MoSOM1 or MoCDTF1 are both necessary for development of spores and appressoria by M. oryzae and play roles in cell wall differentiation, regulating melanin pigmentation and cell surface hydrophobicity during spore formation. MoSom1 strongly interacts with MoStu1 (Mstu1), an APSES transcription factor protein, and with MoCdtf1, while also interacting more weakly with the catalytic subunit of protein kinase A (CpkA) in yeast two hybrid assays. Furthermore, the expression levels of MoSOM1 and MoCDTF1 were significantly reduced in both Δmac1 and ΔcpkA mutants, consistent with regulation by the cAMP/PKA signaling pathway. MoSom1-GFP and MoCdtf1-GFP fusion proteins localized to the nucleus of fungal cells. Site-directed mutagenesis confirmed that nuclear localization signal sequences in MoSom1 and MoCdtf1 are essential for their sub-cellular localization and biological functions. Transcriptional profiling revealed major changes in gene expression associated with loss of MoSOM1 during infection-related development. We conclude that MoSom1 and MoCdtf1 functions downstream of the cAMP/PKA signaling pathway and are novel transcriptional regulators associated with cellular differentiation during plant infection by the rice blast fungus.
| Magnaporthe oryzae, the causal agent of rice blast disease, is an important model fungal pathogen for understanding the molecular basis of plant-fungus interactions. In M. oryzae, the conserved cAMP/PKA signaling pathway has been demonstrated to be crucial for regulating infection-related morphogenesis and pathogenicity, including the control of sporulation and appressorium formation. In this study, we report the identification of two novel pathogenicity-related genes, MoSOM1 and MoCDTF1, by T-DNA insertional mutagenesis. Our results show that MoSOM1 or MoCDTF1 are essential for sporulation, appressorium formatiom and pathogenicity, and also play a key role in hyphal growth, melanin pigmentation and cell surface hydrophobicity. Nuclear localization sequences and conserved domains of the MoSom1 and MoCdtf1 proteins are crucial for their biological function. MoSom1 interacts physically with the transcription factors MoCdtf1 and MoStu1. We also show evidence that MoSom1 has the capacity to interact with CpkA, suggesting that MoSom1 may act downstream of the cAMP/PKA signaling pathway to regulate infection-related morphogenesis and pathogenicity in M. oryzae. Our studies extend the current understanding of downstream components of the conserved cAMP/PKA pathway and its precise role in regulating infection-related development and cellular differentiation by M. oryzae.
| Eukaryotic organisms, including fungi, can sense and respond to extracellular cues via various signaling pathways for regulating a variety of developmental and differential cellular processes. Among these pathways, the conserved cyclic AMP-dependent protein kinase A (cAMP/PKA) signaling pathway has been well studied. The secondary messenger cAMP is universally produced through cyclization of ATP catalyzed by adenylate cyclases (ACs), and the level of cellular cAMP is regulated by cAMP phosphodiesterases [1], [2]. PKA consists of two catalytic subunits and two regulatory subunits. Binding of four cAMP molecules at two sites on each regulatory subunit causes conformational changes in PKA regulatory subunits, releasing activated PKA catalytic subunits which subsequently phosphorylate target proteins, including transcription factors, to control various physiological processes [3]–[6].
The cAMP/PKA response pathway plays an important role in fungal morphogenesis and virulence in plant pathogenic fungi [7]. During the last two decades, the function of several components of the cAMP/PKA pathway, in particular, AC and PKA, has been determined in a number of plant pathogenic fungi, including Colletotrichum trifolii [8], C. lagenarium [9], [10], Fusarium verticillioides [11], Magnaporthe oryzae [12]–[14], Sclerotinia sclerotiorum [15] and U. maydis [16], [17]. In yeasts, several downstream target proteins of PKA have also been identified and functionally characterized. In Saccharomyces cerevisiae for instance, the Flo8 transcription factor is critical for pseudohyphal growth in diploids, haploid invasive growth and flocculation and functions downstream of the cAMP/PKA pathway [18], [19]. A family of FLO genes, including FLO11 (also referred as MUC1) which encodes a cell surface flocculin critical for both pseudohyphal growth and invasive growth, are regulated or activated by Flo8 [19]–[22]. It has been shown that the binding of Flo8 to the promoter of FLO11 is regulated by Tpk2 (a catalytic subunit of PKA) in S. cerevisiae [23]. In both S. cerevisiae and Candida albicans, APSES (Asm1, Phd1, Sok2, Efg1, and StuA) transcription factors are targets for the cAMP/PKA pathway [24]–[27]. C. albicans Flo8 interacts with Efg1, a homolog of the Phd1/Sok2 and StuA proteins that regulate morphogenesis of S. cerevisiae and Aspergillus nidulans, respectively, and is essential for hyphal development and virulence [28]. In phytopathogenic fungi, several APSES transcription factors, including F. oxysporum FoStuA, Glomerella cingulata GcStuA and M. oryzae MoStu1 (Mstu1), have recently been identified [29]–[31]. Both GcStuA and MoStu1 are required for appressorium mediated plant infection [30], [31], while FoStuA is dispensable for pathogenicity by F. oxysporum [29]. In U. maydis, three transcription factors, Prf1, Hgl1 and Sql1, regulated by cAMP pathway have also been identified [32]–[34]. However, the downstream targets of the cAMP/PKA pathway still remain largely unknown in phytopathogenic fungi.
Magnaporthe oryzae is the causal agent of rice blast, the most destructive disease of rice worldwide [35], [36]. In the last two decades, M. oryzae has arisen as a model fungal pathogen for understanding the molecular basis of plant-fungus interactions [36]–[39]. It is now clear that infection-related morphogenesis is controlled by the cAMP response pathway and activation of the mitogen-activated protein kinase (MAPK) cascade in M. oryzae [12], [40]–[42]. Appressorium formation of M. oryzae requires the cAMP-response pathway, which responds to inductive signals from the rice leaf, including surface hydrophobicity and wax monomers from the plant [12]–[14], [43]–[45]. Deletion of the M. oryzae MAC1 gene encoding adenylate cyclase resulted in mutants that cannot form appressoria and were defective in the growth of aerial hyphae and conidiation [12]. However, these defects in Δmac1 mutants could be complemented by adding exogenous cAMP or by spontaneous mutations in the regulatory subunit of PKA gene SUM1 [44]. Consistent with this, M. oryzae CPKA, which encodes the catalytic subunits of PKA, is dispensable for appressorium formation, but is required for appressorial penetration [13], [14]. Additionally, the role of the M. oryzae Pmk1 MAPK pathway in regulating appressorium development has been clearly established [40], [42], [46]–[49]. Therefore, the cAMP/PKA pathway and Pmk1 MAPK cascade are essential for regulation of appressorium development and pathogenicity in the rice blast fungus.
In M. oryzae, the upstream activation of adenylate cyclase appears to be mediated by G-proteins in response to physical and chemical properties of the rice leaf surface. The M. oryzae genome contains three Gα (MagA, MagB, and MagC), one Gβ (Mgb1), and one Gγ (Mgg1) subunits. For the three Gα subunits, only disruption of MAGB can significantly reduce vegetative growth, conidiation, appressorium formation, and pathogenicity, although the ΔmagC mutants are also reduced in conidiation [50]. MagB may respond to surface cues to stimulate Mac1 activity and cAMP synthesis, because expression of a dominant active allele of MAGB causes appressoria to form on hydophilic hard surfaces [51]. Rgs1, a regulator of G-protein signaling, interacts with all the three Gα subunits and functions as a negative regulator of G-proteins in M. oryzae [52]. Additionally, both MGB1 and MGG1 are essential for appressorium formation and plant infection [53], [54]. M. oryzae PTH11 which encodes a putative G-protein-coupled receptor may be involved in regulating Mac1 activities, because PTH11 is required for surface recognition and virulence and exogenous cAMP restores appressorium formation and pathogenicity in PTH11 deletion mutants [55]. Recently, we reported that MoRic8 interacts with MagB and acts upstream of the cAMP/PKA pathway to regulate multiple stages of infection-related morphogenesis in M. oryzae [56]. However, downstream targets of the cAMP/PKA pathway are not well studied in M. oryzae.
Here, we present the identification and functional characterization of two novel pathogenicity-related genes identified by insertional mutagenesis, MoSOM1 and MoCDTF1, which are required for morphogenesis and virulence. Our results have provided evidence that MoSOM1 and MoCDTF1 are regulated by the cAMP/PKA pathway. Deletion of either MoSOM1 or MoCDTF1 resulted in defects in hyphal growth, sporulation, appressorium formation and virulence. MoSom1 strongly interacted with the transcription factors, MoCdtf1 and MoStu1, and also weakly interacted with CpkA in yeast two hybrid assays performed in the presence of cAMP. Moreover, MoSOM1 can complement the defects of S. cerevisiae flo8 in haploid invasive growth and diploid pseudohyphal development. When considered together, these data suggest that MoSom1 is an important regulator of infection-related development in M. oryzae which interacts with the transcription factors, MoCdtf1 and MoStu1, and acts downstream of the cAMP/PKA signaling pathway.
To investigate the molecular basis of plant infection by M. oryzae, a large T-DNA insertional mutagenesis library (∼20,000 transformants) was constructed. All of the transformants were first screened for impairment in pathogenesis by inoculating barley leaves (cv. Golden Promise) with conidia or hyphae (if conidia were not available) using a barley cut-leaf assay. The mutants obtained from the first round screening were subsequently verified by inoculating rice leaves. Among them, YX-145, YX-1303 and YX-864 (Figure 1A; Table S1) were identified as mutants, which were incapable of causing disease on barley or rice leaves (CO-39) following inoculation with hyphae (Figure 1B). To identify the T-DNA integration sites in the mutants, genomic DNA flanking the integrated T-DNAs was obtained from the third round PCR products (Figure S1) and sequenced, respectively. By amplifying the genomic DNAs flanking the left border of the integrated T-DNA, the patterns of T-DNA integrated into these mutants were determined (Figure 1C).
The T-DNA insertion in YX-145 was found at position 593835+, which is 2457 bp downstream of the translational start site, in the seventh exon of a hypothetical gene MGG_04708 (GenBank XP_362263) located on supercontig 16 of chromosome IV. We named the T-DNA tagged gene MoSOM1, because it putatively encodes a predicted protein which is homologous with Som1 proteins, which may be involved in the cAMP-dependent protein kinase pathway controlling growth polarity in related fungal species. MoSom1 showed 47.54, 36.66, 36.84, 37.29, 51.47 and 47.96% amino acid identity with Neurospora crassa Som1 (AAF75278), Aspergillus nidulans OefA (AAW55626), A. niger Som1 (XP_001395127), A. fumigatus Som1 (XP_746706), Metarhizium acridum Som1 (EFY91592) and Verticillium albo-atrum Som1 (XP_003006356), respectively. However, MoSom1 showed only 14.76% and 14.93% amino acid identity with Saccharomyces cerevisiae Flo8 (DAA07769) and Candida albicans Flo8 (AAQ03244 ), respectively. Phylogenetic analysis of the putative homologs of MoSom1 was shown in Figure S2A.
The T-DNA integration site in YX-1303 was at position 1126131-, which is 544 bp downstream of the translational start site, in the first exon of a hypothetical gene MGG_11346 (GenBank XP_001413674) located on supercontig 27 of chromosome I. The T-DNA tagged gene putatively encodes a protein with no known function. We named the gene MoCDTF1 (for Magnaporthe oryzae cAMP-dependent transcription factor gene). MoCdtf1 showed 21.44, 24.23, 18.73 and 27.37% amino acid identity with N. crassa NCU00124 (XP_957248), Sclerotinia sclerotiorum SS1G_07310 (XP_001591864), A. nidulans AN4210 (XP_661814) and Gibberella zeae FG06653 (XP_386829). However, no homolog of MoCdtf1 exists in the genomes of the yeasts Saccharomyces cerevisiae and C. albicans. Phylogenetic analysis of the putative homologs of MoCdtf1 was shown in Figure S2B.
In the YX-864 mutant, MoMSB2 (MGG_06033) was disrupted by T-DNA integration (Figure 1C). To verify the non-pathogenic phenotype of YX-864, we performed a targeted gene deletion of MoMSB2 (Figure S3A). The resulting Δmomsb2 null mutants, MK9 and MK12 (Table S1), were selected by Southern blot analysis (Figure S3B); and were also confirmed by the lack of MoMSB2 transcript using RT-PCR amplification with 864Q-F and 864Q-R (Table S2). Deletion of MoMSB2 had no obvious effect on vegetative growth, conidial germination and sexual development, but caused defects in conidiation, appressorium formation and virulence (Figure S4). The defect in appressorium formation could not be restored by adding exogenous 1,16-hexadecanediol (Diol), cyclic adenosine 3′,5′-cyclophosphate (cAMP), and 3-iso-butyl-1-methylxanthine (IBMX). In S. cerevisae, it has been shown that Msb2 interacts with Sho1 and Cdc42 to promote their function in the filamentous growth pathway [57]. However, no direct interactions between MoMsb2 and MoSho1 (MGG_09125) and MoCdc42 (MGG_00466) were detected in yeast two hybrid assays (data not shown). Taken together, our data provide evidence that MoMSB2 is required for plant infection-related morphogenesis and virulence in M. oryzae, which is consistent with a very recent study in which the gene was independently identified [49].
To determine the role of MoSOM1 in plant infection and confirm the predicted role based on phenotypic analysis of YX-145, we performed targeted gene deletion of MoSOM1 using the gene replacement vectors pMoSOM1-KO (Figure S3C). The gene replacement was analyzed by PCR amplification with primers 145-F and 145-R (Table S2) from transformants. The resulting Δmosom1 null mutants, SK5, SK21 and SK27 (Table S1), were selected based on Southern blot analysis (Figure S3D) and also confirmed by RT-PCR amplification using primers 145Q-F and 145Q-R. One of the transformants resulting from ectopically integrated pMoSOM1-KO, ES16, was used as a control strain. To complement the mutant, the 2.8 kb MoSOM1 gene-coding sequence and a 1.5 kb promoter region was re-introduced into SK27 (Δmosom1) to obtain two complemented strains, SC1 and SC3 (Table S1). Similarly, the Δmocdtf1 null mutants, CTK2 and CTK15, were generated by a targeted gene deletion of MoCDTF1 (Figure S3E and F). The complemented strains, CTC1 and CTC5, were obtained by transforming the genomic DNA including 4.1 kb MoCDTF1 gene-coding sequence and a 1.6 kb promoter region back to Δmocdtf1 (CTK15).
We then harvested the mycelium of Δmosom1 and Δmocdtf1 mutants from liquid CM cultures to inoculate susceptible barley and rice using the cut leaf assay. Our results showed that the wild-type strain Guy11, ectopic (ES16) or complementation (SC1 and CTC1) transformants caused typical rice blast lesions on both intact and abraded barley or rice leaves (Figure 2A). However, consistent with the original analysis of YX-145, the Δmosom1 (SK27) mutant was non-pathogenic on both susceptible barley and rice leaves, even when they were abraded to remove the surface cuticle (Figure 2A). The Δmocdtf1 (CTK15) mutant was non-pathogenic on both barley and rice leaves, but was still able to cause some disease symptoms when leaf surfaces were abraded (Figure 2A). We were unable to carry out a pathogenicity assay using spray inoculation, because these mutants were completely defective in sporulation in culture (see below). Furthermore, the Δmosom1 (SK27) mutant was non-pathogenic when inoculated onto rice roots, but the Δmocdtf1 (CTK15) mutant was still able to cause some disease symptom (Figure 2B). These results therefore demonstrated that the non-pathogenic phenotype of YX-145 and YX-1303 mutants was caused by T-DNA integration and that both MoSOM1 and MoCDTF1 are crucial for plant infection in M. oryzae.
Deletion of MoSOM1 caused significant defects in hyphal growth and colony pigmentation (Figure 3A). The Δmosom1 mutant formed colonies that were less pigmented and which formed less aerial hyphae (Figure 3A). All Δmosom1 mutants (SK5, SK21 and SK27) showed the same phenotypes and only data for mutant SK27 are therefore presented here. When the Δmosom1 mutant (SK27) was grown in CM liquid culture, it formed very small compact mycelium masses, in contrast to the bigger but less compact mycelium formed by the wild-type strain (Figure 3A). The growth rate of mycelium from each strain was determined (Figure 3B). The Δmosom1 mutant and YX-145 were significantly reduced in vegetative growth, forming colonies with diameters of 3.6±0.09 cm and 3.7±0.08 cm after 10-day incubation on CM at 25°C, respectively, compared with 6.8±0.1 cm colony diameter of wild-type strain Guy11 (P<0.01) (Figure 3B). We also carried out mycelial dry weight assays. The results showed that the Δmosom1 mutant was significantly reduced in mycelial dry weight with 0.151±0.007 g compared with 0.330±0.015 g of the wild-type strain Guy11 (P<0.01) after 2-day incubation in liquid CM at 25°C.
Deletion of MoCDTF1 also caused defects in vegetative growth and colony pigmentation on CM plate cultures compared with the wild-type strain, although the affected degree was not as severe as in Δmosom1 mutants (Figure 3A). The Δmocdtf1 mutant (CTK15) formed mycelium that was not well pigmented compared with the wild-type strain and formed smaller mycelium masses in liquid culture (Figure 3A). The Δmocdtf1 mutant and YX-1303 were reduced in vegetative growth, forming colonies with diameters of 5.0±0.08 cm and 5.1±0.1 cm after 10-day incubation on CM at 25°C, respectively, compared with 6.8±0.1 cm colony diameter of wild-type strain Guy11 (P<0.01) (Figure 3B). The other Δmocdtf1 mutant (CTK2) had the same phenotypes as CTK15 (data not shown). To further investigate the roles of MoSOM1 and MoCDTF1, two Δmosom1Δmocdtf1 mutants D-3 and D-9 were created by transformation of pMoSOM1-DK (Figure S3G) into the strain CTK15 (Δmocdtf1) and selected by PCR and confirmed by RT-PCR with the primers 145-F and 145-R (Figure S3H and I), respectively. The Δmosom1Δmocdtf1 mutant D-3 grew more slowly than both the Δmosom1 (SK27) and Δmocdtf1 mutants (CTK15) in culture (Figure 3A and B). Additionally, when the Δmosom1 mutant (SK27) and Δmocdtf1 mutant (CTK15) were inoculated on various media, including MM, PDA and OMA, their vegetative growth and colony pigmented were also impaired (Figure S5). We conclude that MoSOM1 and MoCDTF1 are required for vegetative growth and mycelium pigmentation.
The ability to form spores was evaluated by carefully washing the surface of 10-day-old cultures on CM plates. YX-145, SK27, YX-1303 and CTK15 were unable to form conidia, while the wild-type strain Guy11 produced numerous conidia with 21.0±2.0×106 spores per plate (Figure 4A). When these mutants were grown on different growth media, including MM, PDA, OMA, sporulation was also not observed. These results showed that asexual sporulation was completely blocked by the deletion/disruption of either MoSOM1 or MoCDTF1, indicating that each of the two genes is essential for conidiation in M. oryzae. Furthermore, no conidiophores were observed from the cultures of the mutants, while Guy11 formed normal conidiophores and conidia (Figure 4B). The phenotypes were also observed from other targeted gene replacement mutants, such as SK5, SK21 and CTK2. These results suggest that the defect in conidiation of the Δmosom1 and Δmocdtf1 mutants may be caused by the lack of aerial conidiophore development.
To determine the role of MoSOM1 and MoCDTF1 in sexual reproduction, the wild type Guy11 (MAT1-2), SK27 and CTK15 were crossed with a standard tester strain TH3 (MAT1-1) of M. oryzae to allow perithecium production. After three weeks, the junctions between mated individuals were examined for the presence of perithecia. We observed numerous perithecia at the junctions of the wild type strains Guy11 and TH3, but no perithecia were formed after crossing SK27 with TH3 or CTK15 with TH3 (Figure 4C), even when the incubation time was extended to six weeks. Similarly, crossing of TH3 with the T-DNA insertional mutants (YX-145 and YX-1303), SK5, SK21 and CTK2 did not produce any perithecia, indicating that MoSOM1 or MoCDTF1 are essential for fertility and development of fruiting bodies by M. oryzae. The Δmosom1Δmocdtf1 mutant D-3 was also unable to produce conidiophores, conidia and was completely impaired in sexually development (Figure 4A–C). We conclude that MoSOM1 and MoCDTF1 are both essential for production of asexual and sexual spores by M. oryzae.
Since the Δmosom1 and Δmocdtf1 mutants were unable to produce spores, we harvested mycelium of the mutants from liquid CM culture and appressorium formation was investigated by placing hyphae on hydrophobic surfaces. Numerous appressoria were formed from mycelium of the isogenic wild type strain Guy11, but no appressoria were observed at 24 h or even 48 h post inoculation with the Δmosom1 (SK27) and Δmocdtf1 (CTK15) mutants (Figure 4D). When mycelium of these mutants was placed on barley or rice leaf surfaces, no appressorium formation was induced and no penetration events were observed at 24 h post inoculation (data not shown), indicating the non-pathogenic phenotypes of Δmosom1 and Δmocdtf1 mutants on host leaves may be caused by the defect in appressorium formation. The Δmosom1Δmocdtf1 was also unable to form appressoria from mycelium (Figure 4D). These results suggest that MoSOM1 and MoCDTF1 are both required for appressorium formation and plant infection by M. oryzae.
An easily wettable phenotype can be observed when a fungal culture becomes easily water-logged, due to a loss of surface hydrophobicity, brought about by the absence of the rodlet layer associated with aerial hyphae and conidiospores [58]. We observed that colonies of YX-145 and Δmosom1 mutants were distinct from the wild-type strain Guy11 and formed less aerial hyphae. YX-1303 and Δmocdtf1 mutants were also reduced in aerial hypha formation. We therefore tested the surface hydrophobicity of these strains (Figure 5A). Drops of water and 0.2% gelatin remained on the surface of mycelium of Guy11 and older mycelium of the Δmocdtf1 mutant (CTK15) after 24–48 h incubation, and drops of detergent solution remained suspended on the surface of colonies of Guy11 for about 10–30 min before soaking into the mycelium. By contrast, drops of water and detergent solution immediately soaked into the cultures of the Δmosom1 mutant (SK27) and young mycelium of CTK15 (Figure 5A). Similar results were observed for the other Δmosom1 and Δmocdtf1 mutants. The surface hydrophobicity of the double knockout mutants D-3 and D-9 was similar to the Δmosom1 mutants. The results indicate that deletion of either MoSOM1 or MoCDTF1 affects cell surface hydrophobicity in M. oryzae.
As a consequence of the wettable phenotype of the mutants, we reasoned that M. oryzae hydrophobin genes might be down-regulated in the mutants. To test this idea, we investigated the expression of M. oryzae hydrophobin-encoding genes, including MPG1 and MHP1 and two MHP1 homologs (MGG_09134 and MGG_10105), by quantitative RT-PCR (qRT-PCR). We found that expression of hydrophobin encoding genes was significantly (P<0.01) down-regulated in both Δmosom1 and Δmocdtf1 mutants, particularly in the Δmosom1 mutant (Figure 5B).
To investigate the expression pattern of MoSOM1 during infection-related development, a 1.52 kb promoter fragment upstream of the gene and the entire MoSom1 protein-coding sequence were fused in-frame to the green fluorescent protein (GFP)-encoding gene, GFP (sGFP), and introduced into the Δmosom1 mutant SK27. Transformants carrying a single integration of the pMoSOM1-GFP were selected by DNA gel blot analysis. An independent single plasmid insertion event occurred in the transformants, SC1 and SC3 (Table S1). Punctate green fluorescence was observed in the two transformants. SC3 was used to investigate the spatial localization of the MoSom1 protein in detail. In this analysis, GFP fluorescence was observed both in mycelium and in conidia of SC3, and each cell contained one fluorescence punctum (Figure 6A), suggesting that MoSom1 may localize to the nucleus of each cell. To test this idea, mycelium and conidia of SC3 were stained with 4′-6-Diamidino-2-phenylinodle (DAPI) to stain nuclei specifically. The merged image of GFP and DAPI staining showed that MoSom1-GFP localizes to the nucleus and that each cell contains a single nucleus (Figure 6A). To observe MoSOM1 expression and nuclear division patterns during appressorium development in M. oryzae, conidia of the strain SC3 were allowed to germinate on hydrophobic GelBond film surfaces. During conidium germination, the nucleus in the germinating cell entered mitosis and then one of the daughter nuclei migrated to the incipient appressorium (Figure 6B). Three nuclei that remained in the conidium degenerated and could no longer be seen after approximately 18 hours post inoculation, consistent with previous observations of nuclear division in M. oryzae [59]. Bright green fluorescence of the strain SC3 during penetration on onion epidermis was also observed, as shown in Figure 6C. However, qRT-PCR analysis showed that the expression levels of MoSOM1 were similar at different developmental stages (data not shown), indicating expression throughout the life cycle of the fungus. The expression and localization of MoSom1-GFP was identical in the other transformant SC1. A similar strategy was used to investigate the expression pattern of MoCDTF1 and localization of the encoded protein during infection-related development. Green fluorescence was also observed in nuclei, both in mycelium and in conidia of the transformants CTC1 (Figure S6) and CTC5. However, weak GFP fluorescence was observed in mycelium and in conidia of CTC1 and CTC5 compared strong GFP fluorescence observed in SC1 and SC3. These results provide evidence that both MoSom1 and MoCdtf1 proteins are localized to the nucleus in M. oryzae.
To ensure that all phenotypes observed in the Δmosom1 and Δmocdtf1 mutants were associated with the gene replacement event, we carried out phenotypic analysis of complemented transformants SC1, SC3, CTC1 and CTC5. The GFP-expressing transformants SC1 and CTC1 exhibited full virulence to barley and rice by cut-leaf assay using mycelium inoculations (Figure 2A) or by seedling assays with conidial spray-inoculation. The other phenotypes of Δmosom1 and Δmocdtf1 mutants, including vegetative growth, conidiation and appressorium formation, were all fully complemented by re-introduction of the genes of MoSOM1 or MoCDTF1 (Figure 3B; Figure 4A, B and D; Figure S7). However, the mutants were not responsive to 10 mM exogenous cAMP (data not shown), indicating MoSom1 and MoCdtf1 may act downstream of the cAMP/PKA pathway. We conclude that MoSOM1 or MoCDTF1 are both essential for multiple steps of plant infection-related morphogenesis development and pathogenicity in M. oryzae.
To confirm the position and size of the introns of MoSOM1 and MoCDTF1, cDNA clones of the coding sequence were obtained by reverse transcription-PCR with primer pairs of SOM-E-F/SOM-Xh-R and P1303-F/1303H-Kpn-R (Table S2) and the resulting PCR products cloned into pGEM-T easy vectors and sequenced, respectively. Comparison of the cDNA and sequenced genomic DNA confirmed that MoCDTF1 has an open reading frame of 4,121 bp interrupted by one intron (62 bp) and putatively encodes a 1352 aa protein, which is identical to the protein sequence predicted by automated annotation of the M. oryzae genome sequence (ID: MGG_11346.6; Broad Institute). MoSOM1 has an open reading frame of 2,789 bp interrupted by seven introns (56 bp, 85 bp, 62 bp, 24 bp, 66 bp, 81 bp and 68 bp, respectively) and putatively encodes a 781 aa protein (ID: MGG_04708.6; Broad Institute). However, five splice variants of MoSom1 were also found, as shown in Figure S8. Furthermore, all of the alternatively spliced isoforms of MoSom1 could be detected in RNA extracted from mycelium cultured in liquid CM (1 d, 3 d and 5 d) or conidia from 10-day-old CM plates (data not shown). Three missed amino-acid fragments occurred in exons 4, 6 and 7, respectively, while the extra amino-acid fragment was in exon 4. These data suggested that there may be various forms of post-transcriptional modification of MoSOM1 in M. oryzae.
Both S. cerevisiae Flo8 and C. albicans Flo8 contain a LUFS (LUG/LUH, Flo8, Single-stranded DNA binding protein) domain and there is a LisH (Lissencephaly type 1-like homology) motif within the domain. Similarly, a LUFS domain harbored a LisH motif was also found at the N-terminal portion of the M. oryzae MoSom1 protein (Figure S9A). The amino acid alignment of LisH domains of MoSom1 homologs from related fungal species were shown in Figure 7A, indicating that the fungal LisH domain in fungi is conserved. In addition, MoCdtf1 has a C-terminal ZnF_C2H2 domain. The amino acid alignment of the putative zinc finger, ZnF_C2H2 domain in MoCdtf1 was shown in Figure S9B.
The position of the LisH domain of MoSom1 was shown in Figure 7B. To explore the role of the LisH domain of the MoSom1 protein, we generated a mutant allele of MoSOM1-GFP by deletion of the LisH domain. The resulting transformants (SL1 and SL7) expressing MoSOM1ΔLISH-GFP produced more aerial hyphae and formed more melanized colonies than the original Δmosom1 mutant, but they were still defective in conidiation, asexual/sexual development and pathogenicity (Figure 7C). In these strains, the GFP fluorescence was observed both in nucleus and cytoplasm of hypha (Figure 7C), indicating that protein localization was somewhat affected by the deletion of LisH domain of MoSOM1. Additionally, mutants carrying deletions in the ZnF_C2H2 domain of MoCdtf1 had the same phenotypes as the original strain CTK15 (data not shown), indicating that the domain is essential for the function of MoCdtf1 in M. oryzae. These results indicated that both the LisH domain of MoSom1 and the ZnF_C2H2 domain of MoCdtf1 are essential for infection related morphorgenesis and virulence in M. oryzae.
Consistent with their observed localization patterns (Figure 6; Figure S6), both M. oryzae MoSom1 and MoCdtf1 were predicted to be nuclear localized proteins. The positions of two predicted nuclear localization signals (NLSs) were shown in Figure 7B. To determine the role of the predicted NLSs of MoSom1, we generated mutant alleles of MoSOM1-GFP deleted of each individual putative NLS (PKKK or PSKRVRL) and transformed them into the Δmosom1 mutant (SK27). We found that transformants (SN1-2 and SN1-5) expressing the MoSOM1ΔPKKK -GFP grew normally on CM medium, produced numerous conidia and were fully pathogenic. Moreover, green fluorescence was still observed in the nucleus of these transformants (Figure 7C). However, like the original Δmosom1 mutant, strains (SN2-3 and SN2-4) expressing the MoSOM1ΔPSKRVRL-GFP were unable to produce asexual/sexual spores and were non-pathogenic. Interestingly, we observed green fluorescence of these strains in the cytoplasm of hypha (Figure 7C). These results suggest that PSKRVRL but not PKKK sequence is essential for the function and transportation of MoSom1 protein from cytoplasm to the nucleus. Using a similar strategy, we also demonstrated that the predicted NLS (PPKRKKP) of MoCdtf1 was crucial for the protein localized to the nucleus and its functions during differentiation and plant infection (Figure 7C).
In Saccharomyces cerevisae, Flo8 is critical for invasive growth and flocculation in haploids and pseudohyphal growth in diploids [18]. To determine if MoSOM1 can functionally complement the S. cerevisae flo8 defects, we carried out yeast complementation assays. Our results showed that a yeast strain expressing MoSOM1 in the haploid flo8 mutant HLY850 was restored in its ability to carry out invasive growth on YPD medium (Figure 8A). Consistently, the strain expressing MoSOM1 in the dipoliod flo8 mutant HLY852 recovered the ability to carry out pseudohyphal development on SLAD (synthetic low ammonium dextrose medium) (Figure 8B). These data suggest that MoSOM1 can functionally complement yeast flo8 defects in both haploid invasive growth and diploid pseudohyphal development.
To understand the regulation of MoSOM1 and MoCDTF1 by the cAMP/PKA pathway, the expression of both MoSOM1 and MoCDTF1 was determined by qRT-PCR in Δmac1, ΔcpkA, ΔmagA, ΔmagB and Δrgs1 mutants (Table S1). For comparison, other signaling mutants impaired in infection-related morphogenesis, such as Δpmk1 and Δmps1, were also used. Interestingly, we found that expression levels of MoSOM1 and MoCDTF1 were significantly reduced in Δmac1, ΔcpkA and ΔmagA mutants (P<0.01), but not in other mutants (Figure S10). However, qRT-PCR analysis showed that expression of MoCDTF1 was not significantly regulated by the deletion of MoSOM1 or vice versa (data not shown). These results indicate that expression of MoSOM1 and MoCDTF1 are down-regulated by impairment of the cAMP/PKA signaling pathway.
To understand whether over-expression of the MoSOM1 can restore the phenotypes of the Δmac1 or ΔcpkA mutants, we developed two strains (OM1 and OM4) expressing MoSOM1-GFP driven by the TrpC promoter from A. nidulans in the Δmac1 mutant, and similarly constructed two strains (OC2 and OC7) in the ΔcpkA mutant. Strong fluorescence was observed at the nucleus of these strains (Figure S11A). However, the phenotypes of the Δmac1 or ΔcpkA mutants, including appressorium formation and pathogenicity (Figure S11B and C), were not restored by over-expression of MoSOM1. Interestingly, treatment of SC3 with the adenylate cyclase inhibitor MDL-12,330A hydrochloride at high concentrations, led to some accumulation of MoSom1-GFP in the cytoplasm (Figure S12). These results provide evidence that phosphorylation of MoSom1 by activated CpkA may be important for its nuclear localization.
Our results showed that the phenotypes of Δmosom1 and Δmocdtf1 mutants are somewhat similar and that expression of both MoSOM1 and MoCDTF1 are regulated by the cAMP/PKA signaling pathway. In a previous report, MoStu1 (MGG_00692), an APSES protein of M. oryzae, was shown to be required for pathogenicity and sporulation [31]. To determine whether MoSom1 interacts with the two transcription factors, MoCdtf1 and MoStu1, we carried out yeast two hybrid (Y2H) experiments. The results provided evidence that MoSom1 physically interacts with MoCdtf1 and MoStu1 (Figure 9A), suggesting both MoCdtf1 and MoStu1 were regulated by a direct interaction with MoSom1. However, we did not observe interactions between MoSom1 and other tested proteins, including CpkA and MoLdb1, under these experimental conditions. Additionally, interactions between MoCdtf1 and MoStu1 or MoLdb1 were also not observed in Y2H.
Previously, an interaction between C. albicans Flo8 and Tpk2 was observed in a modified Y2H system [23]. To examine the interaction between MoSom1 and CpkA, we added 5 mM exogenous cAMP into yeast growing medium. Interestingly, a weak interaction between MoSom1 and CpkA was detected by addition of 5 mM exogenous cAMP which may potentially reduce the binding of PKA catalytic subunits with regulatory subunits, while no interaction was detected between the two proteins without adding exogenous cAMP (Figure 9B), presumably because the CpkA is inactive and tightly bound to the endogenous PKA regulatory subunit. These results further demonstrate that MoSom1 may act downstream of the cAMP/PKA pathway in M. oryzae.
To identify genes that are putatively regulated by MoSOM1, we generated serial analysis of gene expression (SAGE) libraries for the wild-type strain (Guy11, 3728956 tags) and the Δmosom1 mutant SK27 (3449284 tags) using mycelium grown in liquid CM medium. To confirm gene expression patterns derived from the SAGE libraries, 10 down-regulated genes in the Δmosom1 mutant were randomly selected and validated by qRT-PCR. The results showed that each gene expression pattern was consistent with that in the SAGE data (Figure 10A). To identify genes that were subjected to regulation by MoSom1, we compared the gene expression profiles between the wild-type strain and the MoSom1 mutant. In total, 719 genes were up-regulated with log2 ratio (Δmosom1/Guy11) >2 and 439 genes were down-regulated with log2 ratio (Δmosom1/Guy11)<−2 (Figure 10B). Genes regulated by deletion of MoSOM1 with log2 Ratio (Δmosom1/Guy11)>1.5 or <−1.5 were shown in Table S3. By analysis of the SAGE data, we found that several pathogenicity-related genes (MPG1, MoVPR1, MoAAT1, MSP1, MoSSADH, MoACT, and COS1) were significantly down-regulated, whereas some (MoRIC8, MAC1, CPKA, MgRAC1, BUF1 and TPS1) were up-regulated (Table 1). The expression patterns of these genes by SAGE were consistent with those by qRT-PCR analysis (Table 1). Interestingly, most genes involved in the cAMP/PKA pathway, including MAC1 and CPKA, were significantly up-regulated by deleting MoSOM1 (Table 1), suggesting that MoSom1 is a negative regulator of their transcription.
Recently, we have described two pathogenicity-related genes, MoRIC8 and MoLDB1 [56], [60]. MoRic8 interacts with Gα subunit MagB and acts upstream of the cAMP/PKA pathway to regulate infection-related morphogenesis. MoLdb1 is a morphogenetic regulator and the Δmoldb1 mutants are similar phenotypes to the Δmosom1 mutants. Therefore, we also generated SAGE libraries from the Δmoric8 mutant Q-10 (3636867 tags) and the Δmoldb1 mutant AK58 (3615472 tags). Sixty most up- or down-regulated genes in the SAGE library of Δmosom1, which were also detected in the SAGE libraries of the Δmoric8 and Δmoldb1 mutants, were presented in Table S4. As expected, the profile of gene expression in the Δmoric8 mutant was very consistent with that in the Δmosom1 mutant SK27, because both MoRic8 and MoSom1 proteins appear to be involved in the cAMP/PKA signaling pathway. Interestingly, the gene expression profiling of the Δmoldb1 mutant was also consistent with that in Δmoric8 or Δmosom1 mutants, although there were interesting differences such as the expression of CPKA, TPS1 and MoACT, as shown in Table 1. These data suggest that there may be a potential link between MoSom1 and MoLdb1 in regulating infection-associated gene expression in M. oryzae.
In this study we identified three T-DNA insertional mutants, YX-145, YX-1303 and YX-864, which were defective in multiple steps of plant infection and morphogenesis by the rice blast fungus Magnaporthe oryzae. HiTAIL-PCR analysis revealed the integrated T-DNA in the mutants disrupted genomic regions corresponding to genes of MoSOM1, MoCDTF1 and MoMSB2, respectively. Targeted deletion of MoSOM1 or MoCDTF1 caused severe defects in both fungal morphogenesis and virulence, which were consistent with the corresponding T-DNA insertional mutants (Figure 2–4). To our knowledge, both MoSOM1 and MoCDTF1 genes have not been functionally characterized previously in phytopathogenic fungi. In addition, our results also showed that MoMSB2 was required for plant infection-related morphogenesis and virulence in M. oryzae, which is consistent with a very recent study in which the gene was independently identified [49]. However, we also observed that deletion of MoMSB2 resulted in a significant reduction in conidiation (Figure S4A), which was distinct from the previous report.
MoSom1 and MoCdtf1 are key morphogenetic regulators. Like most fungal pathogens, asexual reproduction and infection-related development play key roles in the disease cycle in M. oryzae [39]. Molecular genetic analysis of conidiation reveals several conidiation-associated genes that have distinct effects on control of conidiation and conidial morphology. The con7 mutant, for instance, produces a mixture of normal and aberrantly shaped conidia unable to form appressorium, and is non-pathogenic [61]. However, very few mutants have been identified that have completely lost the ability to form conidia in M. oryzae. The MoHOX2 gene encodes a putative homeobox transcription factor. Deletion mutants of MoHOX2 completely abolished asexual sporulation, but the mutants were still pathogenic through hypha-driven appressoria [62], [63]. Recently, we have reported that MoLDB1 gene encoding a protein with a putative LIM binding domain is necessary for fungal morphogenesis [60]. Deletion mutants of MoLDB1 completely lost the ability to differentiate spores, including meiotically generated ascospores, and were non-pathogenic. The mutants were also unable to differentiate conidiophores or appressoria from mycelium [60]. One of the most interesting findings we report here is that deletion either MoSOM1 or MoCDTF1 completely blocked asexual/sexual sporulation and appressorium development from mycelium and the mutants were non-pathogenic. Interestingly, similar to MoLDB1, both MoSOM1 and MoCDTF1 are also required for efficient hyphal growth, melanization and hydrophobicity. Furthermore, we did not observe conidiophores in the mutants, indicating that the defect in conidiation of the mutants is associated with lack of conidiophore formation rather than subsequent conidiogenesis.
M. oryzae MoSom1 is homologous with Aspergillus nidulans OefA and the hypothetical proteins from other related fungal species. Among these proteins in filamentous fungi, only A. nidulans OefA has been investigated [64]. OEFA has been identified by gene silencing and over-expression approaches and targeted deletion of OEFA causes a “fluffy” growth phenotype due to its development of undifferentiated aerial hyphae [64]. However, the detailed role of OefA in signaling pathways has not been characterized. In yeasts, previous studies have shown that Saccharomyces cerevisiae Flo8 is critical for filamentous growth and functions downstream of the cAMP-PKA pathway [18], [19], [23]. Similarly, Candida albicans Flo8 is also essential for hyphal development and virulence and functions downstream of the cAMP-PKA pathway [28]. Since MoSom1 showed only 14.76% and 14.93% amino acid identity with S. cerevisiae Flo8 and C. albicans Flo8, respectively, this makes it difficult to find orthologs of Flo8 from the genomes of filamentous fungi by BLAST search. As a consequence of this, a recent report mentioned that the M. oryzae genome, including many other filamentous ascomycetes, may lack distinct orthologs of Flo8 [65]. However, we have shown in this report that MoSom1 functions downstream of the cAMP/PKA pathway, in a similar manner to yeast Flo8. Several lines of evidence support such a view. First, MoSOM1 can complement a S. cerevisiae flo8 mutant in its ability to carry out haploid invasive and diploid pseudohyphal growth. Second, a strong interaction between MoSom1 and MoStu1 and a weak interaction between MoSom1 and CpkA was detected by yeast two-hybrid analysis. Thirdly, MoSOM1 expression was significantly down-regulated by deletion of MAC1 or CPKA, the two key components of the cAMP/PKA pathway, and finally, the defects of Δmosom1 mutants could not be restored by supplementation with exogenous cAMP.
MoSom1 directly interacted with MoStu1 in a yeast two-hybrid assay, and might therefore act as a regulator of MoStu1 to regulate fungal morphogenesis in M. oryzae. In C. albicans, Efg1, an APSES transcription factor, is essential for regulating morphogenesis [66]. A previous report has demonstrated that C. albicans Flo8 interacts with Efg1 to regulate expression of hypha-specific genes and genes important for virulence [28]. In M. oryzae, MoStu1 is also an APSES transcription factor [31]. Deletion of MoSTU1 results in a reduction of mycelial growth and conidiation and a delay in appressorium formation, and deletion mutants are non-pathogenic [31]. Consistently, we also found that a strong interaction between M. oryzae MoSom1 and MoStu1 in a yeast two-hybrid assay, indicating that MoSom1 may act as a regulator of MoStu1 to regulate fungal morphogenesis. However, because of the different phenotypes of Δmosom1 and Δmostu1 mutants, it seems reasonable to predict that MoSom1 also interacts with other transcription factors in addition to MoStu1.
In a previous study, a direct interaction between S. cerevisae Flo8 and Tpk2 proteins was observed using a modified yeast two-hybrid system carried out in the presence of exogenous cAMP [23]. We found a weak interaction between M. oryzae MoSom1 and CpkA but only when the selection medium was supplemented with 5 mM exogenous cAMP (Figure 9B). This analysis makes a prediction possible that directly places MoSom1 downstream of the cAMP/PKA signaling. In S. cerevisiae, phosphorylation of Flo8 by Tpk2 is required for Flo8 interaction with the FLO11 promoter both in vivo and in vitro [23]. Since multiple PKA phosphorylation sites were also predicted in the MoSom1 protein (see Figure S8B), therefore, in addition to transcriptional regulation, it is possible that MoSom1 is activated by serine/threonine phosphorylation by CpkA to regulate genes required for fungal morphogenesis and pathogenicity. Additionally, we noted that there were obvious different phenotypes between Δmosom1 and ΔcpkA. It is therefore also possible that MoSom1 may be activated by additional regulators from different signaling pathways.
LisH domains exist in various eukaryotic proteins and are required for regulating microtubule dynamics, either by mediating dimerization, or by binding cytoplasmic dynein heavy chain or microtubules directly [67]. Like yeast Flo8, M. oryzae MoSom1 has a LUFS domain with a conserved LisH motif at its N-terminus (Figure S9A). Multiple alignment analyses indicated that the LisH domain is highly conserved in fungi (Figure 7A). We found that the LisH domain is required for the function of MoSom1 in M. oryzae, because deletion of the LisH domain in MoSOM1 partially impaired protein localization to the nucleus and resulted in similar phenotypes to the Δmosom1 mutant (Figure 7C). It is possible that the LisH domain may therefore mediate cytoskeletal interactions necessary for transport of MoSom1 to the nucleus. In S. cerevisiae Flo8 has been localized to the nucleus [18]. Consistent with this, our results also showed that MoSom1 localized to the nucleus and that the predicted NLS of PSKRVRL is important for the function and transportation of MoSom1 protein from the cytoplasm to the nucleus.
In this study, we also found the expression of M. oryzae MoSOM1 was significantly down-regulated by deletion of either MAC1 or CPKA (Figure S10) , which encode the key components of the cAMP/PKA pathway and, interestingly, several genes involved in the cAMP/PKA pathway were significantly up-regulated after deletion of MoSOM1 (Table 1). These data are also consistent with MoSom1 acting downstream of the cAMP/PKA pathway.
When considering these results together, we conclude that MoSom1 is likely to act as a transcriptional regulator that functions downstream of the cAMP/PKA pathway to regulate fungal morphogenesis and pathogenicity.
M. oryzae appears to possess over 400 transcription factor genes, but only a minority of them have so far been characterized, including MST12 [68], CON7 [61], MIG1 [69], MoHOX8 [62], COM1 [70], MoAP1 [71] and MoMCM1 [72], which are required for fungal morphogenesis or plant infection by M. oryzae. In this study, we identified a novel transcription factor, MoCdtf1, which is essential for sporulation, apressorium formation and virulence. However, ΔmoCdtf1 mutants were able to cause some disease on wounded leaves or roots, although the disease severity was significantly reduced compared with the isogenic wild-type strain or complemented strains (Figure 2). These results were consistent with a recent report, in which an insertional mutant M558 was presented in which the T-DNA was integrated into the promoter of MoCDTF1 and also showed impairment in conidiation and pathogenicty, but still infected rice roots [73]. MoCdtf1 has a putative NLS sequence and a conserved zinc finger structure, which are important for MoCdtf1 protein localized to nucleus and for regulating plant infection-related mophorgenesis. Like MoSOM1, expression of M. oryzae MoCDTF1 was significantly down-regulated by deletion of either MAC1 or CPKA (Figure S10). More interestingly, we found that MoCdtf1 physically interacts with MoSom1 in a yeast two hybrid assay (Figure 9A). These data suggest that M. oryzae MoCdtf1 may function as a transcription factor that acts downstream of the cAMP/PKA pathway.
The importance of MoSom1 to infection-related development was underlined by transcriptional profile analysis using SAGE, which demonstrated that a large set of genes are differentially regulated in a Δmosom1 mutant compared to a wild type M. oryzae strain. Significantly, morphogenetic genes, such as the MPG1 hydrophobin gene and the BUF1 melanin biosynthesis gene, as well as physiological regulators such as the TPS1 trehalose-6-phosphate synthase gene were among those differentially regulated. This is consistent with MoSom1 affecting processes pivotal to the formation and function of appressoria and acting downstream of the cyclic AMP signaling pathway, which is necessary for infection-related development in rice blast. The pleiotropic effects of the Δmosom1 mutation on mycelial growth rate do, however, suggest that some of the observed major changes in gene expression may be a consequence of the slower growth rate and aberrant mycelial morphology of Δmosom1 mutants. Dissecting specific families of genes regulated by the moSom1 pathway during appressorium development will therefore be important in elucidating the underlying biological processes regulated by this signaling mechanism.
In summary, based on results from this report, we have developed a model of the cAMP/PKA signaling pathway in M. oryzae that is shown in Figure 11. Surface recognition and initiation of appressorium formation is regulated by the pathway. Moreover, the cAMP/PKA pathway is also involved in regulation of hyphal growth, asexual/sexual sporulation and invasive growth in host tissues. Free CpkA may activate MoSom1 protein to regulate appressium turgor generation through MoStu1 and to control sporulation and appressorium formation through MoCdtf1. However, it is also possible that additional transcription factors are regulated by MoSom1 to control these developmental processes. The model will allow us to test the wider roles of the cAMP/PKA pathway in regulating fungal morphogenesis and plant infection in M. oryzae in future.
All mutants described in the present study were generated from the Magnaporthe oryzae wild-type strain Guy11 [74], and are listed in Table S1. Standard growth and storage procedures for fungal strains were performed, as described previously [58]. A. tumefaciens AGL1 was used for T-DNA insertional transformation. Escherichia coli strain DH-5α was used for routine bacterial transformations and maintenance of various plasmids in this study. Southern blot analysis was performed by the digoxigenin (DIG) high prime DNA labeling and detection starter Kit I (Roche, Mannheim, Germany). General procedures for nucleic acid analysis followed standard protocols [75].Total RNA was extracted from mycelium of M. oryzae using the SV Total RNA Isolation System (Z3100; Promega Corp.) according to the manufacturer's instructions.
For construction of the gene replacement vector pMoSOM1-KO (Figure S3C), 1.4 kb (left border) and 1.2 kb (right border) flanking sequences of the MoSOM1 gene locus were amplified using primer pairs of 3F/4R and 5F/6R (Table S2; Figure S3C) and cloned sequentially into pGEM-T easy vectors to generate pGEM-145L and pGEM-145R, respectively. The 1.4 kb HPH gene cassette, which encodes hygromycin phosphotransferase under control of the A. nidulans TrpC promoter [76], was amplified with primers HPH-Kpn-F and HPH-Xba-R (Table S2) using pCB1003 as a template and clone into pGEM-T easy vectors to give pGEM-HPH. The pGEM-HPH was digested with KpnI and ApaI and inserted the fragment from pGEM-145R with the same digestions to generate pGEM-HPH-R. The pMoSOM1-KO was constructed by insertion SpeI and XbaI fragment from pGEM-145L into corresponding site of pGEM-HPH-R. To construct complementation vector pMoSOM1-GFP, a 4.3 kb fragment including 2.8 kb MoSOM1 gene-coding sequence and a 1.5 kb promoter region were amplified using primers 145H-Nde-F and 145H-Hind-R (Table S2) and then cloned into pGEM-T easy vectors to produce pGEM-SOM. The pMoSOM1-GFP was generated by ligation of pGEM-SOM with the 1.5 kb GFP allele, which was amplified using primers GFP-Hind-F and GFP-Xho-R (Table S2). The pMoSOM1-DKO vector was constructed by replacing the HPH of pMoSOM1-KO with a 0.94 kb bar gene cassette encoding phosphinothricin acetyl transferase under control of the A. nidulans TrpC promoter, which was amplified with primers Bar-Xba-F and Bar-Kpn-R (Table S2) using pMLH21-bar [77] as a template. The MoSOM1 over-expression vector, pOE-MoSOM1, was constructed by insertion the 4.3 kb fragment (2.8 kb MoSOM1 gene-coding sequence and 1.5 kb GFP cassette) , which was amplified with the primers 145OE-Xho-F and GFP-Xho-R (Table S2) using the pMoSOM1-GFP as a template, into the corresponding site of pCB1532 with the A. nidulans trpC promoter.
A similar strategy was used to construct the gene replacement vector pMoCDTF1-KO. About 1.2 kb (left border) and 1.5 kb (right border) flanking sequences MoCDTF1 gene locus were amplified using primer pairs of 7F/8R and 9F/10R (Table S2; Figure S3E) and cloned sequentially into pGEM-T easy vectors to generate pGEM-1303L and pGEM-1303R, respectively. The pGEM-1303R was digested with SacI and XbaI and the released fragment was inserted into the corresponding site of pGEM-HPH to produce pGEM-1303HR. The pGEM-1303HR was digested with KpnI and ApaI and then inserted with the fragment liberated from pGEM-1303L to generate pMoCDTF1-KO. To construct complementation vector pMoCDTF1-GFP, a 5.7 kb fragment including 4.1 kb MoCDTF1 gene-coding sequence and a 1.6 kb promoter region were amplified using primers 1303H-Aat-F and 1303H-Kpn-R (Table S2) and then cloned into pGEM-T easy vectors to produce pGEM-CDTF. The pMoCDTF1-GFP was generated by ligation of pGEM-CDTF with the 1.5 kb GFP allele, which was amplified using primers GFP-Kpn-F and GFP-Xho-R (Table S2).
To construct the MoMSB2 gene replacement vector pMoMSB2-KO (Figure S3A), a 4.2 kb fragment spanning the MoMSB2 locus was amplified with primers 1F and 2R (Table S2) and cloned into pGEM-T easy vector (Promega, Madison, WI, U.S.A.), and a 1.7 kb Xho I and Spl I fragment containing the majority of the MoMSB2 ORF was removed and replaced sequentially with the 1.4 kb HPH gene cassette amplified with primers HPH-Spl-F and HPH-Xho-R (Table S2) using pCB1003 as a template. For construction of complementation vector pMoMSB2-HB, a 4.2 kb fragment including 2.4 kb MoMSB2 gene-coding sequence and a 1.8 kb promoter region were amplified using primers 864H-Sal-F and 864H-Spe-R (Table S2) and then cloned into pGEM-T easy vectors to produce pMoMSB2-HB.
For deletion of the MAC1 gene, the gene deletion vector pMoMAC1-KO was generated using a similar strategy to pMoMSB2-KO. A 4.8 kb fragment spanning the MoMAC1 locus was amplified with primers MAC-KO-FP/MAC-KO-RP (Table S2) and cloned into pGEM-T easy vector to give pGEM-MAC1. The HPH gene cassette was amplified with the primers HPH-Hind-F and HPH-Hind-R (Table S2) using PCB1003 as a template. The pMoMAC1-KO was constructed by insertion HPH gene cassette with HindIII ends into the corresponding restriction site of pGEM-MAC1. The vector for deletion of MAGA gene was kindly provided by professor Hao Liu, Tianjin University of Science and Technology.
The resulting vectors were linearized and transformed into M. oryzae Guy11 protoplasts to generate gene null mutants, respectively, as previously described [58]. Together with pCB1532 [78] vectors, the complementation vectors, pMoSOM1-GFP, pMoCDTF1-GFP and pMoMSB2-HB, were used to co-transform into their corresponding mutants, respectively. The vector pOE-MoSOM1 was used to transform Δmac1 and ΔcpkA mutants to generate strains that MoSOM1 was over-expressed, respectively. GFP fluorescence was observed using a Leica TCS SP5 inverted confocal laser scanning microscope (Leica, Wetzlar, Germany).
Three rounds of PCR amplification were carried out for the construction of pMoSOM1ΔLisH-GFP described as follows. First, 1.6 kb and 4.0 kb fragments were amplified with the primer pairs of 145H-Nde-F/LisH-R and LisH-F/GFP-Xho-R (Table S2) using pMoSOM1-GFP as a template, respectively. Second, the two PCR products were mixed and performed PCR reaction (10 reaction cycles) without adding primers. Third, a 5.6 kb fragment containing 1.5 kb native MoSOM1 promoter, 2.6 kb MoSOM1 gene-coding sequence (without Lish domain) and 1.5 kb GFP cassette was amplified by the primers 145H-Nde-F and GFP-Xho-R (Table S2) using the mixture as a template. Finally, the pMoSOM1ΔLisH-GFP was generated by insertion of the 5.6 kb fragment into pGEM-T easy vector.
A similar strategy was used to construct pMoSOM1ΔPKKK-GFP and pMoCDTF1ΔPPKRKKP-GFP vectors. The pMoSOM1ΔPKKK-GFP was generated from pMoSOM1-GFP using primer pairs of 145H-Nde-F/PKKK-R and PKKK-F/GFP-Xho-R (3.7 kb and 2.1 kb PCR products, respectively), whereas the pMoCDTF1ΔPPKRKKP-GFP was generated from pMoCDTF1-GFP using primer pairs of 1303H-Aat-F/1303CD-R and 1303CD-F/GFP-Xho-R (4.8 kb and 2.4 kb PCR products, respectively). The pMoSOM1ΔPSKRVRL-GFP was constructed by self-ligation of the PCR products amplified with primers PSK-F and PSK-R (Table S2) using pMoSOM1-GFP as a template. The primers used for the constructions were listed in Table S2.
The pMoSOM1ΔLisH-GFP, pMoSOM1ΔPKKK-GFP, pMoSOM1ΔPSKRVRL-GFP were used to transform the Δmosom1 mutants to generate MoSOM1ΔLisH, MoSOM1ΔPKKK, MoSOM1ΔPSKRVRL, respectively. The pMoCDTF1ΔPPKRKKP-GFP was used to transform Δmocdtf1 to produce MoCDTF1ΔPPKRKKP.
For cut-leaf assays, fragments were cut from the leaves of 10-day-old barley cv Golden Promise and 14-day-old rice cv CO-39 seedlings, both highly susceptible toward M. oryzae, and placed in plastic plates containing wetted filters. Mycelium from 2-day-old liquid CM cultures at 25°C was placed onto leaf sections and the plates were incubated in a cycle of 12 h of light and 12 h of dark at 25°C. Wounded leaves were prepared by removing the surface cuticle by abrasion with an emery board as described previously [79]. For spray-inoculation assays, conidial suspensions were diluted in 0.2% gelatin to 1×105 conidia ml−1 for rice infections using rice cv. CO-39. Conidia were spray-inoculated using an artist's airbrush onto 14-day-old plants. Rice seedlings were incubated in plastic bags for 24 h to maintain high humidity and then transferred to controlled environment chambers at 25°C and 90% relative humidity with illumination and 14 h light periods. For root infection assays, rice seeds were germinated for 3 days at 28°C and then transferred to plates contained 2% water agar. Mycelial plugs were carefully placed rice roots. Each test was repeated three times. Disease lesions were examined and photographed after 5 days of incubation.
Vegetative growth was assessed by measurement of colony diameter on plate cultures of M. oryzae grown on CM. For mycelium dry weight assays, the same size blocks (1×1.5 cm2) cut from 7-day-old CM cultures were blended and inoculated in flasks containing 150 ml liquid CM medium. The flasks were incubated at 25°C for 2 days (150 rpm). After incubation, the mycelia produced in liquid cultures were filtered and washed. The dry weight of each mycelium was determined after drying at 60°C for 24 h. Three replicates of each treatment were performed, and the experiment was repeated three times.
Conidial development was assessed by harvesting conidia from the surface of 10-day-old plate cultures and by determining the concentration of the resulting conidial suspension using a haemocytometer. Appressorium development was assessed by allowing conidia at a concentration of 1×104 conidia ml−1 to germinate on hydrophobic GelBond films or onion epidermis and incubating them in a humid environment at 25°C. For appressorium formation from the tips of mycelia, mycelia of the wild-type strain Guy11 and mutant strains were harvested from 48 h liquid CM cultures, and the mycelium fragment suspensions were placed on hydrophobic GelBond film surfaces to allow appressorium development. Appressorium formation was observed after 24 h incubation at 25°C in darkness. Fertility assays were carried out by pairing Guy11 (MAT1-2) and tested strains with standard tester strain TH3 (MAT1-1) on oatmeal agar (OMA) plates, as described previously [60]. Each test was repeated three times.
Total RNA was utilized for synthesis of the first strand cDNA using the PrimeScript™ 1st Strand cDNA Synthesis Kit (D6110A, TaKaRa, Tokyo). The resultant cDNA was used as a template for quantitative RT-PCR (qRT-PCR). qRT-PCR was performed with a SYBR Green Realtime PCR Master Mix Kit (QPK-201, TOYOBO, Osaka, Japan) using an iCycler iQ™ Multicolor Real-Time PCR Detection System (Bio-Rad, Munich, Germany). All qRT-PCR reactions were conducted in triplicates for each sample and the experiment was repeated three times. M. oryzae beta-tubulin gene (MGG_00604) amplified with the primer pairs of BT-F/BT-R was used as an endogenous reference. The abundance of the gene transcripts was calculated relative to this control using the 2−ΔΔCT method [80]. All the primers used for qRT-PCR were listed in Table S2.
Yeast complementation was carried out as described previously [28]. The full length cDNA of MoSOM1 was amplified with primers SOM-E-F and SOM-Xh-R and cloned into pYES2 vector to generate pYES2-SOM1. The yeast expression vector pYES2-SOM1 was transformed into the haploid mutant HLY850 and the diploid mutant HLY852 of S. cerevisiae, respectively. The transformants grown on SD-Ura plates were selected to test the ability of invasive growth on YPD plate and the pseudohyphal growth on SLAD plate supplemented with galactose. The yeast strains, MY1384 (MATa wild type), HLY850 (MATa flo8::hisG ura3-52), CGx68 (MATa/α wild type) and HLY852 (MATa/α flo8::hisG/flo8::hisG ura3-52/ura3-52), were kindly provided by Professor Jiangye Chen of Shanghai Institute for Biological Sciences, Chinese Academy of Sciences.
The Y2H assay was conducted according to the BD Matchmaker Library Construction & Screening Kits instructions (Clontech, PaloAlto, CA, U.S.A.). The full-length cDNA of MoSOM1, MoCDTF1, MoSTU1, MoLDB1 and CPKA was amplified with the primer pairs SOM-E-F/SOM-Xh-R, 1303YTH-E-F/1303YTH-E-R, STU-E-F/STU-E-R, LDB-E-F/LDB-S-R and CPK-F/CPK-R (Table S2), respectively. The cDNA of MoSOM1 was cloned into pGADT7 as the prey vector pGADT7-MoSOM1 and the other cDNAs were cloned into pGBKT7 as the bait vector, respectively. The resulting pGADT7-MoSOM1 and each bait vector were co-transformed into yeast strain AH109. The Leu+ and Trp+ yeast transformants were isolated and assayed for growth on SD-Trp-Leu-His-Ade medium. Yeast strains for positive and negative controls were from the Kit.
The M. grisea wild-type strain Guy11 and the mutants, SK27 (Δmosom1), AK58 (Δmoldb1) [60] and Q-10 (Δmoric8) [56], were cultured in liquid CM medium at 28°C for 48 h in the dark (at 200 rpm). The mycelium of these strains was harvested, and total RNA was extracted using the SV Total RNA Isolation System (Z3100; Promega) according to the manufacturer's instructions. The RNA samples were then sent to Beijing Genomics Institute (BGI; Huada) for serial analysis of gene expression (SAGE).
|
10.1371/journal.pntd.0007457 | “It’s just a fever”: Gender based barriers to care-seeking for visceral leishmaniasis in highly endemic districts of India: A qualitative study | Diagnosis and treatment for visceral leishmaniasis (VL) is considered to be delayed amongst poor, rural women in highly endemic districts of Bihar and Jharkhand. The objective of this study was to assess and understand barriers to VL diagnosis and treatment for women in endemic districts with a high burden of VL.
The study used a stratified and purposive sample of 33 female patients with VL, 11 health staff, 11 local (unqualified) health providers and 12 groups of community elders drawn from ten districts in Bihar and four in Jharkhand with high burdens of VL. The study was conducted within an exploratory and inductive framework, using semi-structured in-depth interviews and discussions.
Women accessing treatment more quickly tended to move faster from treating their symptoms on their own to seeking care from local providers. Perception among female patients of the illness being not serious (owing to initially non-specific and mild symptoms), lack of money, prioritisation of household chores over their need to seek care and the absence of a male guardian to accompany them in seeking care at facilities worked together to drive these choices. Most patients and their families did not suspect VL as the cause for their non-specific symptoms, but when VL was suspected, treatment shopping ended. Lack of prioritization of women’s health issues appears to be a pervasive underlying factor. Public health facilities were not an early treatment choice for the majority, but where it was, the diagnosis of VL was often not considered when presenting with under 2 weeks of symptoms, nor were appropriate follow-up plans instituted.
The insidious presentation of VL and the low prioritisation of women’s health need to be jointly addressed through messages that emphasise the importance of early diagnosis and treatment of disease, which is low-cost in time and money when managed in public health facilities. Clear messages that project prioritising women’s care-seeking over household work as a smart choice and the need for rallying male support are needed. Additionally, efforts to reduce missed opportunities through early case suspicion and engaging private providers to better counsel women with suspected VL could close critical gaps in the continuum of care.
| India bears the greatest burden of a fatal parasitic disease called visceral leishmaniasis (VL), popularly known as Kala Azar. The disease is confined mostly to hot spots in Bihar and Jharkhand in the eastern part of the country. Amongst factors hampering efforts to eliminate VL are delays from the onset of symptoms to the time patients are diagnosed and treated, where disease transmission is thought to be greatest and longer delays result in poorer outcomes; this is considered to be a particular problem in poorer women living in rural areas. This study found patterns in careseeking: initial self-medication followed by multiple visits to local (unqualified) providers and visits to private facilities, before ultimately reaching the public health facilities where treatment was freely available. Those who did attend public health facilities early in their illness were not tested for VL nor followed up with a possible diagnosis of VL in mind. Female patients tended to under-estimate the severity of their illness, while social and economic reasons also influenced care-seeking behaviour–particularly the lack of a male relative in the house, and a reluctance to utilise already meagre resources. The conclusions of the study were the need to encourage women with persistent fever to seek care without delay, while ensuring that factors that serve as barriers, such as low prioritisation of women’s health by households (and the women themselves) is countered by messages emphasizing the danger of delayed treatment for VL and the low time-cost burden of availing care within the public health sector. Improving awareness of VL amongst informal and formal health providers remains key in this process.
| Visceral leishmaniasis (VL) or Kala azar, is a vector-borne disease caused by the protozoan Leishmania donovani and transmitted through the bite of the phlebotomine sand fly. Up to 100,000 cases are estimated to occur globally every year [1], and the disease is normally fatal within two years if untreated [2]. The pathogenesis of VL is complex and clinical presentations vary from asymptomatic infection to fatal disease [3,4].
While VL is highly endemic in the Indian subcontinent and East Africa, India accounts for nearly half of all cases worldwide [1]. The epidemiological features of the disease in the Indian subcontinent together with the availability of effective diagnostic, treatment and vector control measures make it amenable to its elimination as a public health problem [5]. As such, since 2005, India has been part of a regional initiative to achieve this elimination target [6], defined as less than 1 case per 10,000 population at the sub-district level. In the Indian context, these units are called blocks, with populations ranging from 80,000 to 300,000. As a result of the elimination efforts, and possibly also due to the cyclical epidemic nature of VL, there has been a steady decline in reported cases, with an 82% decline between 2011 and 2017 [7].
Despite the encouraging trend, much remains to be done. India remains the only country in the regional elimination initiative that has not yet reached the threshold target, with the disease continuing to be endemic in the states of Bihar, Jharkhand, West Bengal and Uttar Pradesh. By 2018, 49 blocks in Bihar and 25 blocks in Jharkhand have not yet achieved the elimination threshold, with Bihar contributing to nearly 70% of national cases [7]. These highly endemic blocks are spread across 10 districts in Bihar and four districts in Jharkhand.
Under the aegis of the National Roadmap for Kala azar Elimination and with support from a range of stakeholders, over recent years there has been an accelerated effort to control the disease, with rapid diagnostic tests made available at all, and treatment with single-dose Liposomal Amphotericin B at selected nodal public health care facilities in endemic blocks [8]. Kala azar technical supervisors (KTS) at each block oversee vector management and active surveillance, while frontline health workers called Accredited Social Health Activists (ASHA) support prevention and case detection activities at the community level. Development partners have also been conducting a wide range of community education efforts for VL.
Timely detection and treatment of those with acute symptomatic VL is the mainstay of elimination efforts, as this results in the interruption and shortening of transmission from human hosts, which are currently the only known reservoir in the Indian subcontinent. Results from recent modelling studies suggested that decreasing the time between the onset of symptoms to treatment (OT) from 40 to 20 days could bring the elimination target forward by a year [9], with an even larger potential in Bihar [10].
Women with VL appear to access care later than men. A retrospective study of VL patient data from 2012–13 in Bihar found significantly lower proportion of women among reported cases compared to the background population. The paper concluded that this is likely due to under-reporting (as a result of poorer access to healthcare for women), and not necessarily due to lower incidence amongst women, as there was no corresponding difference between the sexes among those under 15 years of age [11]. This is supported by a 2014 study that showed a marked decrease in the proportion of female patients with rising age [12], while an earlier 2006 study that showed that 60–80% of VL patients in facilities were men [13]. A 2003 study from Bangladesh reported that in one highly affected village, reproductive-age women were three times as likely to die from VL compared to men or children; where VL accounted for 23% of all deaths, 80% of these were adult women. Qualitative data from the same study suggest that women experienced substantial barriers to seeking care [14].
VL remains a disease of the poor, with 83% households belonged to the lowest two wealth quintiles in the state’s wealth distribution, while 70% live in mud adobe houses, consistent with the breeding preferences of the sand fly [15]. Caste is an important and reliable indicator of socio-economic status and is used in national surveys as a measure of economic inequality and access to services. It appears that those belonging to the caste categories of Scheduled Caste (SC), and Scheduled Tribes (ST) are disproportionately impacted by VL: a 2012 study of VL patients in Bihar, found that patients from SC had twice the odds of presenting late at treatment centres than others [16].
The objective of this study was to assess and understand barriers to VL diagnosis and care for women in endemic districts with a high burden of VL.
The study was designed to be exploratory and inductive within the post-positivist paradigm, using qualitative inquiry techniques. A purposive and stratified sample was used, with the overall intent of providing a rich and contextualized understanding of care-seeking patterns and drivers of those choices of women patients, but also to serve as an ex-ante strategy for extrapolating results towards providing evidence for practice. The days between onset of symptoms to treatment (OT) was used to stratify the sample into:
Other participants in the study were:
In addition, a smaller sample of male patients with VL with OT of 28–50 days and > 50 days was selected and investigated as a form of quality control, to compare the findings related to female patients and thus remain open and alert to alternative explanations for themes that emerge for female patients.
The total sample size was a trade-off between that which would sufficiently answer the research question and capture a range of experiences, without being too repetitive, and what was feasible. Sampling for patients with VL was done in three stages. In the first stage, 12 blocks were selected as study sites using the Kala Azar Management Information System (KAMIS) national surveillance data from 2015–17, based on the relative distribution of the disease burden between Bihar and Jharkhand. Two blocks each from four of the 10 highly endemic districts in Bihar and one each from all four endemic districts in Jharkhand were selected, using the following criteria:
Of the 12 blocks selected, all met the above criteria aside from one, which met the first three but did not have cases with OT>50, which was selected for logistical reasons. S1 Table gives the details of selection criteria listed above for the blocks that were selected for the study.
In the second stage, a list of patients with VL aged 18 years and above were obtained from the selected blocks and organized by age, caste and OT, and samples were chosen for each of the OT categories of female patients, ensuring those belonging to SC (in Bihar) and those belonging to ST (in Jharkhand) were included. Additional samples were selected as backup. One male patient–with either moderate or late access–was selected from nine of the twelve blocks chosen.
In the third stage, field teams obtained identifying information of these pre-selected patients from the surveillance register maintained at the block level by the KTS, along with details of those with confirmed VL who had died before treatment. Where the pre-selected patients were not traced, new cases from September–December 2017 were included. The final sample included 45 patients: 10 women with early access, 10 women with moderate access, 13 women with late access, 3 men with moderate access, 6 men with late access and 3 who died from confirmed VL. The deceased include two men and one male child, and interviews were conducted out with the wives of the male patients and the mother of the child. In addition, the following non-patient participants were selected: 11 KTS, 11 local providers and groups of community members in 12 communities. Table 1 gives the distribution of the sample across study units:
Field teams carried out in-depth interviews using field-tested semi-structured guides that probed for factors affecting care-seeking, circumstances around each decision and sought to construct a timeline for care-seeking. Similar tools were used to carry out in-depth interviews of the KTS, local providers and focus group discussions in communities to explore trends in care-seeking patterns and their perspectives on factors affecting decisions about care-seeking. Data collection teams consisted of an interviewer/facilitator, note-taker and a manager. Data was collected between December 2017 and January 2018.
Transcribed data and field notes were translated to English and coded in NViVo 8, along with analytic reflective notes of the principal investigator, forming and refining categories as the coding progressed. Data was further compared and contrasted to discern conceptual similarities as well as outliers and refine the distinctions between categories until a theoretical framework emerged.
The study was approved by the Ethics Review Committee of the Foundation for Research in Health Systems, Bangalore, India. Written informed consent was obtained from all participants in the study. Standard procedures for maintaining patient confidentiality and data privacy were followed. All human subjects were 18 years of age or older.
Of the 33 female patients sampled, 23 were from Bihar, with equal numbers having early, moderate and late access; while 10 were from Jharkhand, of which half were those with late access. Women from SC communities was predominantly from Bihar and those from ST communities from Jharkhand, in line with the overall distribution of communities in the two states. Seven out of ten women with late access were from ST communities while two of the three deceased were from ST communities and one from OBC community. Of the nine male patients sampled, four were from OBC communities, three from SC and two from ST communities. About a third of the female patients lived 15 km or farther from the nearest primary health centre (PHC), while two-thirds of the female patients were either landless labourers or worked on non-irrigated land of their own. All women with early access, nine out of ten women with moderate access and nine out of thirteen women with late access lived in houses partly or fully made of mud. A conspicuous observation was the abject poverty of most of the participants and their living conditions, such as inadequate warm clothing to protect against the cold weather. Fig 1 gives the distribution of patients by OT.
A striking finding in the study was that many patients and their family members did not consider VL as a possibility, even after weeks and months of having the symptoms, and after having rounds of unsuccessful treatment. This appeared to be despite having experienced a case of VL in the immediate or extended family, within the neighbourhood or even having been treated for VL themselves in the past. Others reported that they learnt about VL through awareness activities that were held in their neighbourhood, and could recite the salient features of the disease, and yet did not suspect VL in themselves. In an extreme instance, the husband and mother-in-law of a woman with late access had suffered from VL in the previous 2 years and yet she did not suspect VL in herself, but went from self-medicating to several private facilities and a traditional healer before ending up at the local government facility.
Others who had prior knowledge of VL, including those who have had VL in the family or in their neighbourhood, stated that they would have gone to the government facility earlier and “not run here and there for treatment”, had they gotten a clear indication earlier that they were suffering from VL.
Levels of awareness in communities ranged from comprehensive knowledge of causation, transmission, treatment and prognosis, to just knowing that VL is a killer disease. Misconceptions were few and not widely reported, and they include the belief that VL is caused by evil spirits and that it spreads through coughing. Several patient-participants stated that they had not heard about VL prior to falling ill, but picked up the basics from staff at the government facility during treatment. Several non-patient participants stressed the need for creating more awareness, while others felt that public awareness was already high. All patient-participants were confident to encourage others with similar symptoms to seek care at the government facility if symptoms persisted.
The overall care-seeking pattern shows similarities across male and female patients and those with early, moderate and late access: Initial self-medication, followed by repeated visits to or by the local provider, and visits to one or more private facilities, the last of which diagnosed VL and encouraged them to reach a government facility. Perceptions about the severity of the illness, lack of money, not having a male relative in the home and having the male of the household migrating for work drove early treatment choices. These factors also decided how long each of episode of care-seeking would last. This pattern appears to have been more drawn out for women and men with late access, but not for those with early and moderate access, depending on which of those factors were present. Some went to a government facility early on, but were not tested for VL. Fig 2 below depicts this typical pattern, along with some variations.
The government facility is not typically an early choice, but those who did attend one within two weeks of the onset of symptoms were not tested or had a negative test. Without a clear follow-up plan, they went back to treatment shopping. It is unfortunate that patients that reach the public health system in highly endemic districts for VL are not offered advice about the possibility of persistent fever being VL and the important to return for follow up, and it represents significant missed opportunity.
At least three options emerge as changeable drivers to expedite care seeking behaviour of women in the short to medium term, as indicated in Fig 4 below.
The study has some important limitations that need to be considered. Primarily, the sampling frame was taken from the national VL surveillance register, which means that patients that were either not reported or not recorded in the register, such as those receiving diagnosis and treatment entirely in the private sector, were not considered in the sampling frame–although recent studies show that underreporting is low [2]. Secondly, the study was limited to high endemic blocks, where presumably the knowledge and awareness of VL is higher than in moderately and lower endemic area, where the results of this study cannot be extrapolated. Finally the small sample of male patients makes comparative analysis and thematic saturation for male patients difficult.
|
10.1371/journal.pcbi.1002466 | Coordinated Optimization of Visual Cortical Maps (I) Symmetry-based Analysis | In the primary visual cortex of primates and carnivores, functional architecture can be characterized by maps of various stimulus features such as orientation preference (OP), ocular dominance (OD), and spatial frequency. It is a long-standing question in theoretical neuroscience whether the observed maps should be interpreted as optima of a specific energy functional that summarizes the design principles of cortical functional architecture. A rigorous evaluation of this optimization hypothesis is particularly demanded by recent evidence that the functional architecture of orientation columns precisely follows species invariant quantitative laws. Because it would be desirable to infer the form of such an optimization principle from the biological data, the optimization approach to explain cortical functional architecture raises the following questions: i) What are the genuine ground states of candidate energy functionals and how can they be calculated with precision and rigor? ii) How do differences in candidate optimization principles impact on the predicted map structure and conversely what can be learned about a hypothetical underlying optimization principle from observations on map structure? iii) Is there a way to analyze the coordinated organization of cortical maps predicted by optimization principles in general? To answer these questions we developed a general dynamical systems approach to the combined optimization of visual cortical maps of OP and another scalar feature such as OD or spatial frequency preference. From basic symmetry assumptions we obtain a comprehensive phenomenological classification of possible inter-map coupling energies and examine representative examples. We show that each individual coupling energy leads to a different class of OP solutions with different correlations among the maps such that inferences about the optimization principle from map layout appear viable. We systematically assess whether quantitative laws resembling experimental observations can result from the coordinated optimization of orientation columns with other feature maps.
| Neurons in the visual cortex form spatial representations or maps of several stimulus features. How are different spatial representations of visual information coordinated in the brain? In this paper, we study the hypothesis that the coordinated organization of several visual cortical maps can be explained by joint optimization. Previous attempts to explain the spatial layout of functional maps in the visual cortex proposed specific optimization principles ad hoc. Here, we systematically analyze how optimization principles in a general class of models impact on the spatial layout of visual cortical maps. For each considered optimization principle we identify the corresponding optima and analyze their spatial layout. This directly demonstrates that by studying map layout and geometric inter-map correlations one can substantially constrain the underlying optimization principle. In particular, we study whether such optimization principles can lead to spatially complex patterns and to geometric correlations among cortical maps as observed in imaging experiments.
| Neurons in the primary visual cortex are selective to a multidimensional set of visual stimulus features, including visual field position, contour orientation, ocular dominance, direction of motion, and spatial frequency [1], [2]. In many mammals, these response properties form spatially complex, two-dimensional patterns called visual cortical maps [3]–[25]. The functional advantage of a two dimensional mapping of stimulus selectivities is currently unknown [26]–[28]. What determines the precise spatial organization of these maps? It is a plausible hypothesis that natural selection should shape visual cortical maps to build efficient representations of visual information improving the ‘fitness’ of the organism. Cortical maps are therefore often viewed as optima of some cost function. For instance, it has been proposed that cortical maps optimize the cortical wiring length [29], [30] or represent an optimal compromise between stimulus coverage and map continuity [31]–[44]. If map structure was largely genetically determined, map structure might be optimized through genetic variation and Darwinian selection on an evolutionary timescale. Optimization may, however, also occur during the ontogenetic maturation of the individual organism for instance by the activity-dependent refinement of neuronal circuits. If such an activity-dependent refinement of cortical architecture realizes an optimization strategy its outcome should be interpreted as the convergence towards a ground state of a specific energy functional. This hypothesized optimized functional, however, remains currently unknown. As several different functional maps coexist in the visual cortex candidate energy functionals are expected to reflect the multiple response properties of neurons in the visual cortex. In fact, consistent with the idea of joint optimization of different feature maps cortical maps are not independent of each other [8], [10], [19], [23], [42], [45]–[48]. Various studies proposed a coordinated optimization of different feature maps [31], [33], [34], [37], [38], [40]–[42], [44], [49]–[51]. Coordinated optimization appears consistent with the observed distinct spatial relationships between different maps such as the tendency of iso-orientation lines to intersect OD borders perpendicularly or the preferential positioning of orientation pinwheels at locations of maximal eye dominance [8], [10], [19], [23], [42], [45], [47]. Specifically these geometric correlations have thus been proposed to indicate the optimization of a cost function given by a compromise between stimulus coverage and continuity [33], [35], [38], [40], [42], [44], a conclusion that was questioned by Carreira-Perpinan and Goodhill [52].
Visual cortical maps are often spatially complex patterns that contain defect structures such as point singularities (pinwheels) [6], [12], [53], [54], [55] or line discontinuities (fractures) [13], [56] and that never exactly repeat [3]–[10], [12]–[25], [57]. It is conceivable that this spatial complexity arises from geometric frustration due to a coordinated optimization of multiple feature maps in which not all inter-map interactions can be simultaneously satisfied [51], [58]–[61]. In many optimization models, however, the resulting map layout is spatially not complex or lacks some of the basic features such as topological defects [29], [51], [58], [62], [63]. In other studies coordinated optimization was reported to preserve defects that would otherwise decay [51], [58]. An attempt to rigorously study the hypothesis that the structure of cortical maps is explained by an optimization process thus raises a number of questions: i) What are the genuine ground states of candidate energy functionals and how can they be calculated with precision and rigor? ii) How do differences in candidate optimization principles impact on the predicted map structure and conversely what can be learned about an hypothetical underlying optimization principle from observations on map structure? iii) Is there a way to analyze the coordinated organization of cortical maps predicted by optimization principles in general? If theoretical neuroscience was able to answer these questions with greater confidence, the interpretation and explanation of visual cortical architecture could build on a more solid foundation than currently available. To start laying such a foundation, we examined how symmetry principles in general constrain the form of optimization models and developed a formalism for analyzing map optimization independent of the specific energy functional assumed.
Minima of a given energy functional can be found by gradient descent which is naturally represented by a dynamical system describing a formal time evolution of the maps. Response properties in visual cortical maps are arranged in repetitive modules of a typical spatial length called hypercolumn. Optimization models that reproduce this typical length scale are therefore effectively pattern forming systems with a so-called ‘cellular’ or finite wavelength instability, see [64]–[66]. In the theory of pattern formation, it is well understood that symmetries play a crucial role [64]–[66]. Some symmetries are widely considered biologically plausible for cortical maps, for instance the invariance under spatial translations and rotations or a global shift of orientation preference [51], [63], [67]–[71]. In this paper we argue that such symmetries and an approach that utilizes the analogy between map optimization and pattern forming systems can open up a novel and systematic approach to the coordinated optimization of visual cortical representations.
A recent study found strong evidence for a common design in the functional architecture of orientation columns [3]. Three species, galagos, ferrets, and tree shrews, widely separated in evolution of modern mammals, share an apparently universal set of quantitative properties. The average pinwheel density as well as the spatial organization of pinwheels within orientation hypercolumns, expressed in the statistics of nearest neighbors as well as the local variability of the pinwheel densities in cortical subregions ranging from 1 to 30 hypercolumns, are found to be virtually identical in the analyzed species. However, these quantities are different from random maps. Intriguingly, the average pinwheel density was found to be statistical indistinguishable from the mathematical constant up to a precision of 2%. Such apparently universal laws can be reproduced in relatively simple self-organization models if long-range neuronal interactions are dominant 3,70–72. As pointed out by Kaschube and coworkers, these findings pose strong constraints on models of cortical functional architecture [3]. Many models exhibiting pinwheel annihilation [51], [58] or pinwheel crystallization [62], [63], [73] were found to violate the experimentally observed layout rules. In [3] it was shown that the common design is correctly predicted in models that describe long-range interactions within the OP map but no coupling to other maps. Alternatively, however, it is conceivable that they result from geometric frustration due to inter-map interactions and joint optimization. In the current study we therefore in particular examined whether the coordinated optimization of the OP map and another feature map can reproduce the quantitative laws defining the common design.
The presentation of our results is organized as follows. First we introduce a formalism to model the coordinated optimization of complex and real valued scalar fields. Complex valued fields can represent for instance orientation preference (OP) or direction preference maps [14], [24]. Real valued fields may represent for instance ocular dominance (OD) [1], spatial frequency maps [20], [45] or ON-OFF segregation [74]. We construct several optimization models such that an independent optimization of each map in isolation results in a regular OP stripe pattern and, depending on the relative representations of the two eyes, OD patterns with a regular hexagonal or stripe layout. A model-free, symmetry-based analysis of potential optimization principles that couple the real and complex valued fields provides a comprehensive classification and parametrization of conceivable coordinated optimization models and identifies representative forms of coupling energies. For analytical treatment of the optimization problem we adapt a perturbation method from pattern formation theory called weakly nonlinear analysis [64]–[66], [75]–[78]. This method is applicable to models in which the spatial pattern of columns branches off continuously from an unselective homogeneous state. It reduces the dimensionality of the system and leads to amplitude equations as an approximate description of the system near the symmetry breaking transition at which the homogeneous state becomes unstable. We identify a limit in which inter-map interactions that are formally always bidirectional become effectively unidirectional. In this limit, one can neglect the backreaction of the complex map on the layout of the co-evolving scalar feature map. We show how to treat low and higher order versions of inter-map coupling energies which enter at different order in the perturbative expansion.
Second we apply the derived formalism by calculating optima of two representative low order examples of coordinated optimization models and examine how they impact on the resulting map layout. Two higher order optimization models are analyzed in Text S1. For concreteness and motivated by recent topical interest [3], [79], [80], we illustrate the coordinated optimization of visual cortical maps for the widely studied example of a complex OP map and a real feature map such as the OD map. OP maps are characterized by pinwheels, regions in which columns preferring all possible orientations are organized around a common center in a radial fashion [53], [55], [81], [82]. In particular, we address the problem of pinwheel stability in OP maps [51], [71] and calculate the pinwheel densities predicted by different models. As shown previously, many theoretical models of visual cortical development and optimization fail to predict OP maps possessing stable pinwheels [29], [51], [58], [62]. We show that in case of the low order energies, a strong inter-map coupling will typically lead to OP map suppression, causing the orientation selectivity of all neurons to vanish. For all considered optimization models, we identify stationary solutions of the resulting dynamics and mathematically demonstrate their stability. We further calculate phase diagrams as a function of the inter-map coupling strength and the amount of overrepresentation of certain stimuli of the co-evolving scalar feature map. We show that the optimization of any of the analyzed coupling energies can lead to spatially relatively complex patterns. Moreover, in case of OP maps, these patterns are typically pinwheel-rich. The phase diagrams, however, differ for each considered coupling energy, in particular leading to coupling energy specific ground states. We therefore thoroughly analyze the spatial layout of energetic ground states and in particular their geometric inter-map relationships. We find that none of the examined models reproduces the experimentally observed pinwheel density and spatially aperiodic arrangements. Our analysis identifies a seemingly general condition for interaction induced pinwheel-rich OP optima namely a substantial bias in the response properties of the co-evolving scalar feature map.
We model the response properties of neuronal populations in the visual cortex by two-dimensional scalar order parameter fields which are either complex valued or real valued [53], [83]. A complex valued field can for instance describe OP or direction preference of a neuron located at position . A real valued field can describe for instance OD or the spatial frequency preference. Although we introduce a model for the coordinated optimization of general real and complex valued order parameter fields we consider as the field of OP and as the field of OD throughout this article. In this case, the pattern of preferred stimulus orientation is obtained by(1)The modulus is a measure of the selectivity at cortical location .
OP maps are characterized by so-called pinwheels, regions in which columns preferring all possible orientations are organized around a common center in a radial fashion. The centers of pinwheels are point discontinuities of the field where the mean orientation preference of nearby columns changes by 90 degrees. Pinwheels can be characterized by a topological charge which indicates in particular whether the orientation preference increases clockwise or counterclockwise around the pinwheel center,(2)where is a closed curve around a single pinwheel center at . Since is a cyclic variable in the interval and up to isolated points is a continuous function of , can only have values(3)where is an integer number [84]. If its absolute value , each orientation is represented only once in the vicinity of a pinwheel center. In experiments, only pinwheels with a topological charge of are observed, which are simple zeros of the field .
OD maps can be described by a real valued two-dimensional field , where indicates ipsilateral eye dominance and contralateral eye dominance of the neuron located at position . The magnitude indicates the strength of the eye dominance and thus the zeros of the field correspond to the borders of OD.
In this article, we view visual cortical maps as optima of some energy functional . The time evolution of these maps can be described by the gradient descent of this energy functional. The field dynamics thus takes the form(4)where and are nonlinear operators given by , . The system then relaxes towards the minima of the energy . The convergence of this dynamics towards an attractor is assumed to represent the process of maturation and optimization of the cortical circuitry. Various biologically detailed models have been cast to this form [35], [51], [85].
All visual cortical maps are arranged in repetitive patterns of a typical wavelength . We splitted the energy functional into a part that ensures the emergence of such a typical wavelength for each map and into a part which describes the coupling among different maps. A well studied model reproducing the emergence of a typical wavelength by a pattern forming instability is of the Swift-Hohenberg type [65], [86]. Many other pattern forming systems occurring in different physical, chemical, and biological contexts (see for instance [75]–[78]) have been cast into a dynamics of this type. Its dynamics in case of the OP map is of the form(5)with the linear Swift-Hohenberg operator(6), and a nonlinear operator. The energy functional of this dynamics is given by(7)In Fourier representation, is diagonal with the spectrum(8)The spectrum exhibits a maximum at . For , all modes are damped since and only the homogeneous state is stable. This is no longer the case for when modes on the critical circle acquire a positive growth rate and now start to grow, resulting in patterns with a typical wavelength . Thus, this model exhibits a supercritical bifurcation where the homogeneous state looses its stability and spatial modulations start to grow.
The coupled dynamics we considered is of the form(9)where , and is a constant. To account for the species differences in the wavelengths of the pattern we chose two typical wavelengths and . The dynamics of and is coupled by interaction terms which can be derived from a coupling energy . In the uncoupled case this dynamics leads to pinwheel free OP stripe patterns.
How many inter-map coupling energies exist? Using a phenomenological approach the inclusion and exclusion of various terms has to be strictly justified. We did this by symmetry considerations. The constant breaks the inversion symmetry of inputs from the ipsilateral () or contralateral () eye. Such an inversion symmetry breaking could also arise from quadratic terms such as . In the methods section we detail how a constant shift in the field can eliminate the constant term and generate such a quadratic term. Including either a shift or a quadratic term thus already represents the most general case. The inter-map coupling energy was assumed to be invariant under this inversion. Otherwise orientation selective neurons would, for an equal representation of the two eyes, develop different layouts to inputs from the left or the right eye. The primary visual cortex shows no anatomical indication that there are any prominent regions or directions parallel to the cortical layers [67]. Besides invariance under translations and rotations of both maps we required that the dynamics should be invariant under orientation shifts . Note, that the assumption of shift symmetry is an idealization that uncouples the OP map from the map of visual space. Bressloff and coworkers have presented arguments that Euclidean symmetry that couples spatial locations to orientation shift represents a more plausible symmetry for visual cortical dynamics [68], [87], see also [88]. The existence of orientation shift symmetry, however, is not an all or none question. Recent evidence in fact indicates that shift symmetry is only weakly broken in the spatial organization of orientation maps [89], [90]. A general coupling energy term can be expressed by integral operators which can be written as a Volterra series(10)with an -th. order integral kernel . Inversion symmetry and orientation shift symmetry require to be even and that the number of fields equals the number of fields . The lowest order term, mediating an interaction between the fields and is given by i.e.(11)Next, we rewrite Eq. (11) as an integral over an energy density . We use the invariance under translations to introduce new coordinates(12)This leads to(13)The kernel may contain local and non-local contributions. Map interactions were assumed to be local. For local interactions the integral kernel is independent of the locations . We expanded both fields in a Taylor series around (14)For a local energy density we could truncate this expansion at the first order in the derivatives. The energy density can thus be written(15)Due to rotation symmetry this energy density should be invariant under a simultaneous rotation of both fields. From all possible combinations of Eq. (15) only those are invariant in which the gradients of the fields appear as scalar products. The energy density can thus be written as(16)where we suppress the argument . All combinations can also enter via their complex conjugate. The general expression for is therefore(17)From all possible combinations we selected those which are invariant under orientation shifts and eye inversions. This leads to(18)The energy densities with prefactor to do not mediate a coupling between OD and OP fields and can be absorbed into the single field energy functionals. The densities with prefactors and (also with and ) are complex and can occur only together with () to be real. These energy densities, however, are not bounded from below as their real and imaginary parts can have arbitrary positive and negative values. The lowest order terms which are real and positive definite are thus given by(19)The next higher order energy terms are given by(20)Here the fields and enter with an unequal power. In the corresponding field equations these interaction terms enter either in the linear part or in the cubic nonlinearity. We will show in this article that interaction terms that enter in the linear part of the dynamics can lead to a suppression of the pattern and possibly to an instability of the pattern solution. Therefore we considered also higher order interaction terms.
These higher order terms contain combinations of terms in Eq. (19) and are given by(21)As we will show below examples of coupling energies(22)form a representative set that can be expected to reproduce experimentally observed map relationships. For this choice of energy the corresponding interaction terms in the dynamics Eq. (9) are given by(23)with and denoting the complex conjugate. In general, all coupling energies in , and can occur in the dynamics and we restrict to those energies which are expected to reproduce the observed geometric relationships between OP and OD maps. It is important to note that with this restriction we did not miss any essential parts of the model. When using weakly nonlinear analysis the general form of the near threshold dynamics is insensitive to the used type of coupling energy and we therefore expect similar results also for the remaining coupling energies.
Numerical simulations of the dynamics Eq. (9), see [63], [91], with the coupling energy Eq. (22) and are shown in Fig. 1. The initial conditions and final states are shown for different bias terms and inter-map coupling strengths . We observed that for a substantial contralateral bias and above a critical inter-map coupling pinwheels are preserved from random initial conditions or are generated if the initial condition is pinwheel free. Without a contralateral bias the final states were pinwheel free stripe solutions irrespective of the strength of the inter-map coupling.
We studied Eq. (9) with the low order inter-map coupling energies in Eq. (22) using weakly nonlinear analysis. We therefore rewrite Eq. (9) as(24)where we shifted both linear operators as , . The constant term in Eq. (9) is replaced by a quadratic interaction term with , see Methods. The uncoupled nonlinearities are given by , while and are the nonlinearities of the low order inter-map coupling energy Eq. (23). We study Eq. (24) close to the pattern forming bifurcation where and are small. We therefore expand both control parameters in powers of the small expansion parameter (25)Close to the bifurcation the fields are small and thus nonlinearities are weak. We therefore expand both fields as(26)We further introduced a common slow timescale and insert the expansions in Eq. (24) and get(27)and(28)We consider amplitude equations up to third order as this is the order where the nonlinearity of the low order inter-map coupling energy enters first. For Eq. (27) and Eq. (28) to be fulfilled each individual order in has to be zero. At linear order in we get the two homogeneous equations(29)Thus and are elements of the kernel of and . Both kernels contain linear combinations of modes with a wavevector on the corresponding critical circle i.e.(30)with the complex amplitudes , and , . In view of the hexagonal or stripe layout of the OD pattern shown in Fig. 1, is an appropriate choice. In the following sections we assume i.e. the Fourier components of the emerging pattern are located on a common circle. To account for species differences we also analyzed models with detuned OP and OD wavelengths in part (II) of this study.
At second order in we get(31)As and are elements of the kernel . At third order, when applying the solvability condition (see Methods), we get(32)We insert the leading order fields Eq. (30) and obtain the amplitude equations(33)For simplicity we have written only the simplest inter-map coupling terms. Depending on the configuration of active modes additional contributions may enter the amplitude equations. In addition, for the product-type coupling energy, there are coupling terms which contain the constant , see Methods and Eq. (40). The coupling coefficients are given by(34)From Eq. (33) we see that inter-map coupling has two effects on the modes of the OP pattern. First, inter-map coupling shifts the bifurcation point from to . This can cause a potential destabilization of pattern solutions for large inter-map coupling strength. Second, inter-map coupling introduces additional resonant interactions that for instance couple the modes and their opposite modes . In case of the inter-map coupling terms in dynamics of the modes are small. In this limit the dynamics of the modes decouples from the modes and we can use the uncoupled OD dynamics, see Methods. When we scale back to the fast time variable and set , we obtain(35)The amplitude equations are truncated at third order. If pattern formation takes place somewhat further above threshold fifth order, seventh order, or even higher order corrections are expected to become significant and can induce quantitative modifications of the low order solutions. If third order approximate solutions exhibit degeneracies or marginal stability, higher orders of perturbation theory will qualitatively change the solutions. However, none of the solutions found in the studied models was only marginally stable. This suggests that the obtained solutions are in general structurally stable. A derivation of amplitude equation with higher order inter-map coupling energies is presented in Text S1.
Using symmetry considerations we derived inter-map coupling energies up to eighth order in the fields, see Eq. (19), Eq.(20), and Eq.(21). Which of these various optimization principles could reproduce realistic inter-map relationships such as a uniform coverage of all stimulus features? We identified two types of optimization principles that can be expected to reproduce realistic inter-map relationships and good stimulus coverage. First, product-type coupling energies of the form . These energies favor configurations in which regions of high gradients avoid each other and thus leading to high coverage. Second, gradient-type coupling energies of the form . In experimentally obtained maps, iso-orientation lines show the tendency to intersect the OD borders perpendicularly. Perpendicular intersection angles lead to high coverage as large changes of the field in one direction lead to small changes of the field in that direction. To see that the gradient-type coupling energy favors perpendicular intersection angles we decompose the complex field into the selectivity and the preferred orientation . We obtain(36)If the orientation selectivity is locally homogeneous, i.e. , then the energy is minimized if the direction of the iso-orientation lines () is perpendicular to the OD borders. In our symmetry-based analysis we further identified terms that are expected to lead to the opposite behavior for instance mixture terms such as .
Pinwheels are prominent features in OP maps. We therefore also analyze how different optimization principles impact on the pinwheel positions with respect to the co-evolving feature maps. The product-type coupling energies are expected to favor pinwheels at OD extrema. Pinwheels are zeros of and are thus expected to reduce this energy term. They will reduce energy the most when is maximal which should repel pinwheels from OD borders, where is zero. Also the gradient-type coupling energy is expected to couple the OD pattern with the position of pinwheels. To see this we decompose the field into its real and imaginary part(37)At pinwheel centers the zero contours of and cross. Since there and are almost constant and not parallel the energy can be minimized only if is small at the pinwheel centers, i.e. the extrema or saddle-points of .
From the previous considerations we assume all coupling coefficients of the energies to be positive. A negative coupling coefficient can be saturated by higher order inter-map coupling terms. In the following, we discuss the impact of the low order inter-map coupling energies on the resulting optima of the system using the derived amplitude equations. The corresponding analysis for higher order inter-map coupling energies is provided in Text S1.
As indicated by numerical simulations and weakly nonlinear analysis of the uncoupled OD dynamics, see Methods, we discussed the influence of the OD stripe, hexagon, and constant solutions on the OP map using the coupled amplitude equations derived in the previous section. A potential backreaction onto the dynamics of the OD map can be neglected if the modes of the OP map are much smaller than the modes of the OD map. This can be achieved if . We first give a brief description of the uncoupled OP solutions. Next, we study the impact of the low order coupling energies in Eq. (22) on these solutions. We demonstrate that these energies can lead to a complete suppression of orientation selectivity. In the uncoupled case there are for two stable stationary solutions to the amplitude equations Eq. (35), namely OP stripes(38)and OP rhombic solutions(39)with , , an arbitrary phase, and . In the uncoupled case the angle between the Fourier modes is arbitrary. The stripe solutions are pinwheel free while the pinwheel density for the rhombic solutions varies as and thus . For the rhombic solutions pinwheels are located on a regular lattice. We therefore refer to these and other pinwheel rich solutions which are spatially periodic as pinwheel crystals (PWC). In particular, we refer to pinwheel crystals with as rhombic spatial layout as rPWC solutions and pinwheel crystals with a hexagonal layout as hPWC solutions. Without inter-map coupling, the potential of the two solutions reads , thus the stripe solutions are always energetically preferred compared to rhombic solutions.
In the following we study three scenarios in which inter-map coupling can lead to pinwheel stabilization. First, a deformation of the OP stripe solution can lead to the creation of pinwheels in this solution. Second, inter-map coupling can energetically prefer the (deformed) OP rhombic solutions compared to the stripe solutions. Finally, inter-map coupling can lead to the stabilization of new PWC solutions.
For the low order interaction terms the amplitude equations are given by , with the potential(40)with the uncoupled contributions(41)The coupling coefficients read , , , , where is the angle between the wavevector and .
We first studied the impact of the low order product-type coupling energy. Here, the constant enters explicitly in the amplitude equations, see Eq. (40) and Eq. (68).
When using a gradient-type inter-map coupling energy the interaction terms are independent of the OD shift . In this case, the coupling strength can be rescaled as and is therefore independent of the bias . The bias in this case only determines the stability of OD stripes, hexagons or the constant solution.
In this study we presented a symmetry-based analysis of models formalizing that visual cortical architecture is shaped by the coordinated optimization of different functional maps. In particular, we focused on the question of whether and how different optimization principles specifically impact on the spatial layout of functional columns in the primary visual cortex. We identified different representative candidate optimization principles. We developed a dynamical systems approach for analyzing the simultaneous optimization of interacting maps and examined how their layout is influenced by coordinated optimization. In particular, we found that inter-map coupling can stabilize pinwheel-rich layouts even if pinwheels are intrinsically unstable in the weak coupling limit. We calculated and analyzed the stability properties of solutions forming spatially regular layouts with pinwheels arranged in a crystalline array. We analyzed the structure of these pinwheel crystals in terms of their stability properties, spatial layout, and geometric inter-map relationships. For all models, we calculated phase diagrams showing the stability of the pinwheel crystals depending on the OD bias and the inter-map coupling strength. Although differing in detail and exhibiting distinct pinwheel crystal phases for strong coupling, the phase diagrams exhibited many commonalities in their structure. These include the general fact that the hexagonal PWC phase is preceded by a phase of rhombic PWCs and that the range of OD biases over which pinwheel crystallization occurs is confined to the stability region of OD patch solutions.
Our analytical calculations of attractor and ground states close a fundamental gap in the theory of visual cortical architecture and its development. They rigorously establish that models of interacting OP and OD maps in principle offer a solution to the problem of pinwheel stability [51], [71]. This problem and other aspects of the influence of OD segregation on OP maps have previously been studied in a series of models such as elastic net models [34], [37], [41], [42], [51], [59], self-organizing map models [38], [40], [44], [49], [50], spin-like Hamiltonian models [58], [93], spectral filter models [94], correlation based models [95], [96], and evolving field models [97]. Several of these simulation studies found a higher number of pinwheels per hypercolumn if the OP map is influenced by strong OD segregation compared to the OP layout in isolation or the influence of weak OD segregation [51], [93], [97]. In such models, large gradients of OP and OD avoid each other [40], [49]. As a result, pinwheel centers tend to be located at centers of OD columns as seen in experiments [19], [45], [46], [48], [95]. By this mechanism, pinwheels are spatially trapped and pinwheel annihilation can be reduced [51]. Moreover, many models appear capable of reproducing realistic geometric inter-map relationships such as perpendicular intersection angles between OD borders and iso-orientation lines [59], [94], [95]. Tanaka et al. reported from numerical simulations that the relative positioning of orientation pinwheels and OD columns was dependent on model parameters [95]. Informative as they were, almost all of these previous studies entirely relied on simulation methodologies that do not easily permit to assess the progress and convergence of solutions. Whether the reported patterns were attractors or just snapshots of transient states and whether the solutions would further develop towards pinwheel-free solutions or other states thus remained unclear. Moreover, in almost all previous models, a continuous variation of the inter-map coupling strength was not possible which makes it hard to disentangle the contribution of inter-map interactions from intrinsic mechanisms. The only prior simulation study of a coordinated optimization model that tracked the number of pinwheels during the optimization process did not provide evidence that pinwheel annihilation could be stopped but only reported a modest reduction in annihilation efficiency [51]. From this perspective, the prior evidence for coordination induced pinwheel stabilization appears relatively limited. Our analytical results leave no room to doubt that map interactions can stabilize an intrinsically unstable pinwheel dynamics. They also reveal that interaction of orientation preference with a stripe pattern of OD is per se not capable of stabilizing pinwheels.
Independent of its predictions, our study clarifies the general mathematical structure of interaction dominated optimization models. To the best of our knowledge our study for the first time describes an analytical approach for examining the solutions of coordinated optimization models for OP and OD maps. Our symmetry-based phenomenological analysis of conceivable coupling terms provides a general classification and parametrization of biologically plausible coupling terms. To achieve this we mapped the optimization problem to a dynamical systems problem which allows for a perturbation expansion of fixed points, local minima, and optima. Using weakly nonlinear analysis, we derived amplitude equations as an approximate description near the symmetry breaking transition. We identified a limit in which inter-map coupling becomes effectively unidirectional enabling the use of the uncoupled OD patterns. We studied fixed points and calculated their stability properties for different types of inter-map coupling energies. This analysis revealed a fundamental difference between high and low order coupling energies. For the low order versions of these energies, a strong inter-map coupling typically leads to OP map suppression, causing the orientation selectivity of all neurons to vanish. In contrast, the higher order variants of the coupling energies do generally not cause map suppression but only influence pattern selection, see Text S1. We did not consider an interaction with the retinotopic map. Experimental results on geometric relationships between the retinotopic map and the OP map are ambiguous. In case of ferret visual cortex high gradient regions of both maps avoid each other [42]. In case of cat, however, high gradient regions overlap [18]. Such positive correlations cannot be easily treated with dimension reduction models, see [98]. It is noteworthy that our phenomenological analysis identified coupling terms that could induce an attraction of high gradient regions. Such terms contain the gradient of only one field and can thus be considered as a mixture of the gradient and the product-type energy.
Our results indicate that a patchy layout of a second visual map interacting with the OP map is important for the effectiveness of pinwheel stabilization by inter-map coupling. Such a patchy layout can be easily induced by an asymmetry in the representation of the corresponding stimulus feature such as eye dominance or spatial frequency preference. In spatial frequency maps, for instance, low spatial frequency patches tend to form islands in a sea of high spatial frequency preference [45]. Also in cat visual cortex the observed OD layout is patchy [99]–[103]. In our model, the patchy layout results from the overall dominance of one eye. In this case, OD domains form a system of hexagonal patches rather than stripes enabling the capture and stabilization of pinwheels by inter-map coupling. The results from all previous models did not support the view that OD stripes are capable of stabilizing pinwheels [51], [93], [97]. Our analysis shows that OD stripes are indeed not able to stabilize pinwheels, a result that appears to be independent of the specific type of map interaction. In line with this, several other theoretical studies, using numerical simulations [51], [93], [97], indicated that more banded OD patterns lead to less pinwheel rich OP maps. For instance, in simulations using an elastic net model, the average pinwheel density of OP maps interacting with a patchy OD layout was reported substantially higher (about 2.5 pinwheels per hypercolumn) than for OP maps interacting with a more stripe-like OD layout (about 2 pinwheels per hypercolumn) [51].
Several lines of biological evidence appear to support the picture of interaction induced pinwheel stabilization. Supporting the notion that pinwheels might be stabilized by the interaction with patchy OD columns, visual cortex is indeed dominated by one eye in early postnatal development and has a pronounced patchy layout of OD domains [104]–[106]. Further support for the potential relevance of this picture comes from experiments in which the OD map was artificially removed resulting apparently in a significantly smoother OP map [44]. In this context it is noteworthy that macaque visual cortex appears to exhibit all three fundamental solutions of our model for OD maps: stripes, hexagons, and a monocular solution, which are stable depending on the OD bias. In the visual cortex of macaque monkeys, all three types of patterns are found near the transition to the monocular segment, see [106] and Fig. (8). Here, OD domains form bands in the binocular region and a system of ipsilateral eye patches at the transition zone to the monocular region where the contralateral eye gradually becomes more dominant. If pinwheel stability depends on a geometric coupling to the system of OD columns one predicts systematic differences in pinwheel density between these three zones of macaque primary visual cortex. Because OD columns in the binocular region of macaque visual cortex are predominantly arranged in systems of OD stripes our analysis also indicates that pinwheels in these regions are either stabilized by other patchy columnar systems or intrinsically stable.
One important general observation from our results is that map organization was often not inferable by simple qualitative considerations on the energy functional. The organization of interaction induced hexagonal pinwheel crystals reveals that the relation between coupling energy and resulting map structure is quite complex and often counter intuitive. We analyzed the stationary patterns with respect to intersection angles and pinwheel positions. In all models, intersection angles of iso-orientation lines and OD borders have a tendency towards perpendicular angles whether the energy term mathematically depends on this angle, as for the gradient-type energies, or not, as for the product-type energies. Intersection angle statistics thus are not a very sensitive indicator of the type of interaction optimized. Mathematically, these phenomena result from the complex interplay between the single map energies and the interaction energies. In case of the low order gradient-type inter-map coupling energy all pinwheels are located at OD extrema, as expected from the used coupling energy. For other analyzed coupling energies, however, the remaining pinwheels are located either at OD saddle-points (low order product-type energy) or near OD borders (higher order gradient-type energy), in contrast to the expection that OD extrema should be energetically preferred. Remarkably, such correlations, which are expected from the gradient-type coupling energies, occur also in the case of the product-type energies. Remarkably, in case of product type energies pinwheels are located at OD saddle-points. which is not expected per se and presumably result from the periodic layout of OP and OD maps. Correlations between pinwheels and OD saddle-points have not yet been studied quantitatively in experiments and may thus provide valuable information on the principles shaping cortical functional architecture.
Our results demonstrate that, although distinct types of coupling energies can leave distinguishing signatures in the structure of maps shaped by interaction (as the OP map in our example), drawing precise conclusions about the coordinated optimization principle from observed map structures is not possible for the analyzed models. In the past numerous studies have attempted to identify signatures of coordinated optimization in the layout of visual cortical maps and to infer the validity of specific optimization models from aspects of their coordinated geometry [34], [37], [38], [40]–[42], [44], [49]–[51], [58], [59], [93]. It was, however, never clarified theoretically in which respect and to which degree map layout and geometrical factors of inter-map relations are informative with respect to an underlying optimization principle. Because our analysis provides complete information of the detailed relation between map geometry and optimization principle for the different models our results enable to critically assess whether different choices of energy functionals specifically impact on the predicted map structure and conversely what can be learned about the underlying optimization principle from observations of map structures.
We examined the impact of different interaction energies on the structure of local minima and ground states of models for the coordinated optimization of a complex and a real scalar feature map such as OP and OD maps. The models were constructed such that in the absence of interactions, the maps reorganized into simple stripe or blob pattern. In particular, the complex scalar map without interactions would form a periodic stripe pattern without any phase singularity. In all models, increasing the strength of interactions could eventually stabilize qualitatively different, more complex, and biologically more realistic patterns containing pinwheels that can become the energetic ground states for strong enough inter-map interactions. The way in which this happens provides fundamental insights into the relationships between map structure and energy functionals in optimization models for visual cortical functional architecture.
Our results demonstrate that the structure of maps shaped by inter-map interactions is in principle informative about the type of coupling energy. The organization of the complex scalar map that optimizes the joined energy functional was in general different for all different types of coupling terms examined. We identified a class of hPWC solutions which become stable for large inter-map coupling. This class depends on a single parameter which is specific to the used inter-map coupling energy. Furthermore, as shown in Text S1, pinwheel positions in rPWCs, tracked while increasing inter-map coupling strength, were different for different coupling terms examined and thus could in principle serve as a trace of the underlying optimization principle. This demonstrates that, although pinwheel stabilization is not restricted to a particular choice of the interaction term, each analyzed phase diagram is specific to the used coupling energy. In particular, in the strong coupling regime substantial information can be obtained from a detailed inspection of solutions.
In the case of the product-type coupling energies, the resulting phase diagrams are relatively complex as stationary solutions and stability borders depend on the magnitude of the OD bias. Here, even quantitative values of model parameters can in principle be constrained by analysis of the map layout. In contrast, for the gradient-type coupling energies, the bias dependence can be absorbed into the coupling strength and only selects the stationary OD pattern. This leads to relatively simple phase diagrams. For these models map layout is thus uninformative of quantitative model parameters. We identified several biologically very implausible OP patterns. In the case of the product-type energies, we found orientation scotoma solutions which are selective to only two preferred orientations. In the case of the low order gradient-type energy, we found OP patterns containing pinwheels with a topological charge of 1 which have not yet been observed in experiments. If the relevant terms in the coupling energy could be determined by other means, the parameter regions in which these patterns occur could be used to constrain model parameters by theoretical bounds.
The information provided by map structure overall appears qualitative rather than quantitative. In both low order inter-map coupling energies (and the gradient-type higher order coupling energy, see Text S1), hPWC patterns resulting from strong interactions were fixed, not exhibiting any substantial dependence on the precise choice of interaction coefficient. In principle, the spatial organization of stimulus preferences in a map is an infinite dimensional object that could sensitively depend in distinct ways to a large number of model parameters. It is thus not a trivial property that this structure often gives essentially no information about the value of coupling constants in our models. The situation, however, is reversed when considering the structure of rPWCs. These solutions exist and are stable although energetically not favored in the absence of inter-map interactions. Some of their pinwheel positions continuously depend on the strength of inter-map interactions. These solutions and their parameter dependence nevertheless are also largely uninformative about the nature of the interaction energy. This results from the fact that rPWCs are fundamentally uncoupled system solutions that are only modified by the inter-map interaction. As pointed out before, preferentially orthogonal intersection angles between iso-orientation lines and OD borders appear to be a general feature of coordinated optimization models in the strong coupling regime. Although the detailed form of the intersection angle histogram is solution and thus model specific, our analysis does not corroborate attempts to use this feature to support specific optimization principles, see also [52], [107], [108]. The stabilization of pinwheel crystals for strong inter-map coupling appears to be universal and provides per se no specific information about the underlying optimization principle. In fact, the general structure of the amplitude equations is universal and only the coupling coefficients change when changing the coupling energy. It is thus expected that also for other coupling energies, respecting the proposed set of symmetries, PWC solutions can become stable for large enough inter-map coupling.
Our analysis conclusively demonstrates that OD segregation can stabilize pinwheels and induce pinwheel-rich optima in models for the coordinated optimization of OP and OD maps when pinwheels are intrinsically unstable in the uncoupled dynamics of the OP map. This allows to systematically assess the possibility that inter-map coupling might be the mechanism of pinwheel stabilization in the visual cortex. The analytical approach developed here is independent of details of specific optimization principles and thus allowed to systematically analyze how different optimization principles impact on map layout. Moreover, our analysis clarifies to which extend the observation of the layout in physiological maps can provide information about optimization principles shaping visual cortical organization.
The common design observed in experimental OP maps [3] is, however, not reproduced by the optima of the analyzed optimization principles. Whether this is a consequence of the applied weakly nonlinear analysis or of the low number of optimized feature maps or should be considered a generic feature of coordinated optimization models will be examined in part (II) of this study [91]. In part (II) we complement our analytical studies by numerical simulations of the full field dynamics. Such simulations allow to study the rearrangement of maps during the optimization process, to study the timescales on which optimization is expected to take place, and to lift many of the mathematical assumptions employed by the above analysis. In particular, we concentrate on the higher order inter-map coupling energies for which the derived amplitude equations involved several simplifying conditions, see Text S1.
We studied the intersection angles between iso-orientation lines and OD borders. The intersection angle of an OD border with an iso-orientation contour is given by(63)where denotes the position of the OD zero-contour lines. A continuous expression for the OP gradient is given by . We calculated the frequency of intersection angles in the range . In this way those parts of the maps are emphasized from which the most significant information about the intersection angles can be obtained [19]. These are the regions where the OP gradient is high and thus every intersection angle receives a statistical weight according to . For an alternative method see [10].
We studied how the emerging OD map depends on the overall eye dominance. To this end we mapped the uncoupled OD dynamics to a Swift-Hohenberg equation containing a quadratic interaction term instead of a constant bias. This allowed for the use of weakly nonlinear analysis to derive amplitude equations as an approximate description of the shifted OD dynamics near the bifurcation point. We identified the stationary solutions and studied their stability properties. Finally, we derived expressions for the fraction of contralateral eye dominance for the stable solutions.
|
10.1371/journal.pgen.1003777 | Cell-Type Specific Features of Circular RNA Expression | Thousands of loci in the human and mouse genomes give rise to circular RNA transcripts; at many of these loci, the predominant RNA isoform is a circle. Using an improved computational approach for circular RNA identification, we found widespread circular RNA expression in Drosophila melanogaster and estimate that in humans, circular RNA may account for 1% as many molecules as poly(A) RNA. Analysis of data from the ENCODE consortium revealed that the repertoire of genes expressing circular RNA, the ratio of circular to linear transcripts for each gene, and even the pattern of splice isoforms of circular RNAs from each gene were cell-type specific. These results suggest that biogenesis of circular RNA is an integral, conserved, and regulated feature of the gene expression program.
| Last year, we reported that circular RNA isoforms, previously thought to be very rare, are actually a pervasive feature of eukaryotic gene expression programs; indeed, the major RNA isoform from hundreds of human genes is a circle. Previous novel RNA species that initially appeared to be special cases, of dubious biological significance, have subsequently proved to have critical, conserved biological roles. An almost universal characteristic of regulatory macromolecules is that they are themselves regulated during development and differentiation. Here, we show that the repertoire of genes expressing circular RNA, the relative levels of circular: linear transcripts from each gene, and even the pattern of splice isoforms of circular RNAs from each gene were cell-type specific, including examples of striking regulation. In humans, we estimate that circular RNA may account for about 1% as many molecules as poly(A) RNA. The ubiquity of circular RNA and its specific regulation could significantly alter our perspective on post-transcriptional regulation and the roles that RNA can play in the cell.
| Recently, we were surprised to find that the predominant RNA isoform from hundreds of human genes is a circle, and that circular RNAs were transcribed from thousands of genes in both human and mouse [1]. Circular RNA transcripts had been reported previously for a handful of genes. With the possible exceptions of the circular RNA isoforms of the Sry gene in mouse testis [2] and the muscleblind gene in Drosophila melanogaster [3] these were generally thought to be rare RNA species, perhaps the result of transcriptional noise. In humans, circular isoforms of the transcripts from the ETS and cytochrome P450 2C24 genes have also been reported; these were found to be inabundant compared to linear RNA isoforms from the same genes [3]–[5]. In recent years, two antisense circular RNAs were discovered and studied more intensely in humans: an antisense transcript from the INK4A-ARF locus, cANRIL, and an abundant antisense transcript to CDR1; the latter was recently reported to be a microRNA sink [6]–[9].
The ubiquitous expression of circular RNA in human and mouse cells has now been independently confirmed by high throughput sequencing of the RNase R treated, ribosomal-depleted fraction of RNA, combined with a previously published informatic algorithm to identify circular RNA [7] as well as by a second report characterizing RNA after ribosomal RNA depletion [8]. In the former report, a large majority of the circular isoforms we had described (1025 of 1319) were also identified by deep sequencing of RNase R-treated RNA. This overlap in circles identified in these two studies is noteworthy because the more recent report focused on fibroblasts, while we previously analyzed RNA isolated from leukocytes and pediatric leukemias.
Here, we describe a more systematic bioinformatic and statistical genome-wide study that significantly expands the catalogue of circular RNAs identified in human cells and reveals significant regulation of circular RNA expression. In many applications, computational challenges associated with mapping and with distinguishing between sequencing errors and sequence homology prevent reliable identification of structural variants, including circular RNAs. Indeed, although de novo splicing detection algorithms have been used in more than a thousand published studies, including studies aimed at identifying gene fusions and internal tandem duplications, most instances of scrambled exons in human RNAs, and thus the circular species that they represent, had eluded detection.
A major challenge in bioinformatic and statistical identification of novel RNA isoforms, particularly circular RNA, involves distinguishing bona fide evidence of scrambled exons in RNA from confounding factors such as sequence degeneracy at exon boundaries and sequencing errors. To address these challenges and to identify circular isoforms from public ENCODE RNA-Seq data, we developed a new bioinformatic approach. The main idea behind our computational method is that it refrains from qualitative hard thresholding of read alignment quality, and instead computes statistical averages of alignment quality scores. This approach allows us to distinguish putative novel splice junctions where the majority of reads align to the ‘novel’ junction with high quality alignment scores from those where reads with high alignment scores are rare. The method allows for systematic FDR-based thresholding, rather than qualitative cut-offs, to determine classification as a scrambled junction at a prescribed confidence level.
Our method was focused on identifying circular RNA transcribed from genes whose linear isoform exons are annotated: we first built a database of all scrambled junctions between annotated exon boundaries, essentially as previously described [1], extending the database to annotated hg19 UCSC ‘knowngene’ exon boundaries. Importantly, we did not impose a lower threshold on the length of annotated exons in our database, instead generating database entries of short annotated exons by an ‘in silico’ rolling circle. We required a minimum of 10 nt on both sides of a diagnostic read to span a scrambled exon-exon junction.
The improved sensitivity of this approach compared to previous methods allowed us to identify thousands of previously unreported circular isoforms and some very small circular RNAs, exemplified by a <150 nt circular RNA isoform ABTB1 resulting from the splicing of two short exons. Other small circular isoforms that we identified and confirmed include a 204 nt circular isoform from a single exon of LINC00340 - a long intergenic noncoding RNA, and a two exon circle of 151 nt from the RNA binding motif gene RBM5.
Experimental and bioinformatic noise can give rise to spurious evidence of circular transcripts, especially for highly expressed genes. To tackle this problem, we combined the bioinformatic approach above with a statistical strategy to distinguish reads supporting exon scrambling from reads likely to be homology and sequencing artifacts. Briefly, we did not impose any thresholds on alignment quality of either read 1 (aligning to a diagnostic scrambled exon-exon junction) or read 2 (aligning to a canonical isoform or a diagnostic scrambled junction). All read pairs with evidence of scrambled exon splicing at a particular pair of genomic coordinates were aggregated by averaging, measuring the overall quality of reads aligning to the putative circle.
We generated an empirical null distribution of average alignment qualities using “decoy” read pairs. These “decoy” reads had the property that read 1 mapped to a scrambled intragenic exon X – exon Y junction and read 2 mapped within the same gene but would be excluded from a circle composed of exons Y, Y+1, … X, see Figure 1A. These alignment qualities were averaged across all reads for each circular RNA and generated our null distribution of the alignment quality as depicted in Figure 1B. This approach allowed us to compute a per-isoform FDR (by referring the alignment score per isoform to the empirical null distribution) and reduce calls of false positive circular isoforms which riddled the data before this approach was applied.
Using this approach, we were able to enhance statistical discrimination between case 1: false positive evidence of circular RNA isoforms in highly expressed genes resulting from reads with sequencing errors observed due to high sampling of these genes, and case 2: bona fide low-level expression of circular isoforms from these highly expressed genes. For example, the vast majority of reads (99.995%) from putative circular isoforms of GAPDH had a FDR significantly surpassing our threshold of .025; these reads would be taken as evidence of circular isoform expression with a naïve approach.
We applied our method to a large publicly available data set from the ENCODE consortium (see Table 1), with the goal of identifying novel RNA circular isoforms and studying regulation of circular RNA. This dataset consisted of 76-nt paired-end reads from RNA isolated from 15 different cancer and non-cancer cell lines representing mesodermal, ectodermal, endodermal lineages and the pluripotent H1-hESC (see Table 1). Each RNA sample was depleted of poly(A) RNA, size selected to be above 200 nt, and subsequently subjected to ribosomal RNA depletion by RiboMinus (see ENCODE protocols). We have also made use of public analysis of the matched poly(A)-selected samples from these cell lines published by the consortium. Note that the statistics above are absolute counts not corrected for sequencing depth, which varied by sample.
Across the 15 cell types, at an FDR of .025, we found 46866 distinct intragenic splice junctions at annotated exon boundaries in 8466 genes. Across cell types, we detected the largest number of genes with evidence of circular RNA expression in the leukemia cell line K562 (16559 distinct circle-specific splice junctions); in the fetal lung fibroblast line AG04450, we detected 11590 distinct splicing circle-specific splice junctions and in the human foreskin fibroblast line BJ, we detected 7771 (this is not a typo: [7] reports exactly the same number of circular isoforms). Recently, 7771, 2229 and 485 splice junctions, of ‘low’, ‘medium’ and ‘high’ stringency, respectively, representing circular RNA were identified by another method in Hs68 cells, a human fetal foreskin fibroblast line [7].
We used the enzyme RNase R, a highly processive 3′ to 5′ exoribonuclease, to test our computational predictions of circular RNA species. This exonuclease is not expected to digest circular RNA because they lack the required free 3′ end but readily digests linear RNAs with a 3′ single stranded region of greater than 7 nucleotides [10]. We tested a panel of 8 putative circular RNAs varying in size, abundance and abundance of the corresponding linear isoforms: ABTB1, FAT1, HIPK3, CYP24A1, LINC00340, LPAR1, and PVT1. As positive controls, we included 3 genes with strong prior evidence of circularization: MAN1A2, RNF220 and CAMSAP1. We treated total RNA from HeLa cells with either RNase R or a mock enzyme treatment. For each sample, we performed an RT with random hexamer primers and used qPCR to quantify the change in abundance of species with scrambled exons compared to species with exons that we predicted found only in linear RNA isoforms, following treatment with RNase R. All the RNA species that we had predicted to be circular were resistant to RNase R whereas all predicted linear sequences were highly sensitive to RNase R (see Figure 2), providing strong evidence that our computational method specifically identifies circular RNA species.
Because the ENCODE libraries were constructed to preserve strand directionality, we could analyze the data for evidence that circular RNAs may serve as an RNA-dependent RNA polymerase as has been shown to occur in some siRNA pathways [11] and in viral or viroid replication [12]. Among paired end reads that support either sense or antisense circular RNA, with a diagnostic circular junction between exon boundaries annotated for linear RNA, we found a strong and significant bias in the directionality of reads (almost 100% of reads from >93% of putative circles). This bias supports the hypothesis that the significant majority of RNA circles formed using splice sites shared with annotated linear RNAs are transcribed from the same strand as the canonical linear RNA. By this analysis, the percentage of circular isoforms in the sense orientation (with respect to the linear isoform) was 96% for HMEC and >99% for the other 14 cell types. This evidence argues against a primary function of circular RNA serving as an RNA template for an RNA-directed polymerase. A small minority of reads had a polarity inconsistent with transcription in the same direction as linear RNA; we believe in most cases these represent artifacts of reverse transcription, perhaps induced by RNA secondary structure. Note that our intention was not to identify un-annotated antisense circular RNA, and we did not search for circular RNAs that might have been produced by splicing a primary transcript complementary to the annotated transcript.
We sought to determine the relative abundance of each circular RNA compared to its cognate linear RNA. This required estimating the relative abundance of each linear RNA, the relative abundance of each circular RNA, and an “equivalence factor” or normalization constant relating the number of RNA molecules represented by 1 measured unit of linear RNA to 1 measured unit of circular RNA. For linear RNA abundance, we used the estimates generated by the public ENCODE consortium analysis of polyadenylated fractions, represented in RPKM units (reads per kilobase of transcript per million mapped reads in the sample). Our estimate of each circular RNA isoform's abundance from sequencing was the number of read pairs, in poly(A)-minus fractions, in which one read spanned a circular junction (note that counting junctional reads in this manner inherently normalizes by gene length). To determine the equivalence factor (the number of junctional read counts expected for a circular RNA expressed at the same level as a transcript with an RPKM of 1), we measured the abundance of circular and linear isoforms of FAT1 and HIPK3 across three ENCODE cell lines (A549, AG04450 and HeLa) by qPCR. This allowed us to relate the abundance of the linear isoforms of FAT1 and HIPK3 as measured in units of RPKM to the abundance of the circular isoforms of these genes as measured in units of junctional read counts. Since the equivalence factor is the same for all genes in the genome, we were then able to compute the relative abundance of circular and linear isoforms for all genes detected in the sequencing data.
FAT1 and HIPK3 were chosen because they have large, abundant circular RNA isoforms and high linear RNA isoform expression, thus mitigating potential factors confounding this estimation such as rolling circle amplification of small circular RNA during the RT, and statistical uncertainty introduced by estimating the expression of low abundance circle or linear isoforms.
These estimates suggested that there was roughly 1 molecule of circular RNA for every 100 molecules of poly(A) RNA in the cell lines we evaluated: A549, AG04450 and HeLa. For roughly 50 genes in each cell line, circular transcript isoforms were estimated to be more abundant than linear isoforms (see Tables S1, S2, S3 for a complete list of the relative linear: circular isoforms per gene genome-wide). For most genes with circular RNA isoforms, the abundance of the circles was roughly 5–10% that of their linear counterparts. At least among this small sample of cell lines, the differences in growth rate and developmental origin do not appear to fundamentally alter the genome-wide rate of circular RNA expression.
As a spot-check of our sequencing based estimates of relative abundance of linear and circular isoforms, we performed a Northern blot for CAMSAP1 with total RNA from HeLa cells (see Figure 3). Sequencing based estimates suggested that the circular isoform of CAMSAP1 consisting of exons 2 and 3 was 20 times more abundant than the linear counterpart. The Northern blot shows that CAMSAP1 circular isoforms are more abundant than the linear isoform. Intriguingly, one of the major bands (at 1446 nt) is an unexpected circular isoform consisting of exon 2 - intron 2 - exon 3. RT-PCR bands consistent with both isoforms were amplified from RNase R treated HeLa RNA; Sanger sequencing of the gel-purified bands verified their structure (data not shown).
RNA-Seq also provided evidence for this retained-intron circle: in poly(A) depleted fractions of HeLa-S3, the highest read density was in exons 2 and 3 followed by intron 2, with other introns more than 10-fold lower. Of note, our estimate of the ratio of intron 2 to exon 3 expression (based on number of reads with zero mismatches to the genome) was somewhat higher in the nuclear fraction (38%) compared to in the cytosolic (10%) or RNA isolated from total cells (16%). There was also cell type variation of the ratio of intron 2 to exon 3 read density in “cell” fractions across the ENCODE data set, from 16% in HeLa-S3 down to 3.5% in NHEK, suggesting that intron-retention in CAMSAP1 circles may be under regulatory control.
We previously reported that genes with circular RNA transcripts tend to have larger introns than genome-wide averages. That analysis showed that even after controlling for the tendency for intron lengths to decrease from 5′ to 3′ along the canonical transcript [1], [13], the introns immediately flanking the exon boundaries that participated in the scrambled splice were significantly longer than average. Here we further investigated the relationship between intron length and circularized exons in this deeper survey of circular RNA expression. For each UCSC annotated gene, for each annotated splice site, we defined the flanking intron length for the 3′ and 5′ splice sites as the distance to the nearest upstream or downstream 5′ or 3′ splice site respectively. To control for systematic biases, for example, that genes expressing circular RNA isoforms have relatively large introns compared to genome-wide averages (as we have found previously), we performed the following analysis.
We ranked introns that flanked spliced exons generating circular RNA isoforms in two ways: 1) weighting the lengths of introns flanking each circular isoform by the abundance of the corresponding circular RNA (right panel in Figure 4); 2) counting each circular isoform once regardless of its expression level (left panel in Figure 4). For the analysis depicted in Figure 4, for each gene, we ranked the length of each intron according to its length. We then converted each rank value to a quantile: for example, the second largest intron in a gene with 5 introns would receive a quantile of 40 ( = 2/5 * 100%). For reference, under a null model where the rank of intron length had no relationship with propensity to flank a circular splice donor or acceptor, the heatmaps would have uniform intensity regardless of the quantile represented. We found that the relative length of the flanking intron did not reliably determine which exons were spliced to form an RNA circle, although within a gene, longer introns were more likely to flank circularized exons.
To test whether small variations in intron length might explain the dynamic range in intron length quantiles observed in Figure 4, we also examined the relative length of each intron flanking a diagnostic donor or acceptor site in the circle as a fraction of the largest intron length in the gene (Figure S1). Thus, if one of the introns flanking a splice site diagnostic of a circle were the longest intron in the gene, its ratio compared to the maximum intron length would be 1. The null distribution we considered was based on the relative length of the second vs. third largest intron in the set of genes evaluated. Unexpectedly, we found that, measured as a fraction of maximum intron length, introns flanking circular junctions were, on average, smaller than expected from the null distribution, perhaps explained by a single atypically long intron within genes expressing circular isoforms.
We explored regulation of circular RNA expression using the ENCODE RNA-Seq data for a number of cultured cell lines, then did an independent evaluation of some of the identified circle expression variation using qPCR (see Figure 5). Some of the genes we tested by qPCR (CYP24A1, PVT1 and LPAR1 and LINC00340) expressed circular RNA isoforms that were predicted from sequence data to vary across cell lines; others (FAT1, HIPK3) appeared from the RNA-Seq data to have constant levels of circular isoform expression in A549, AG04450 and HeLa cells (data not shown).
To assess variation in circular RNA expression genome-wide, we estimated the abundance of circular RNA from sequence data based on the diagnostic splice-junction counts described earlier; estimates of linear transcript abundance were from the ENCODE consortium's analysis of poly(A) gene expression. For a set of relatively highly expressed circular isoforms, we evaluated the fit of a Poisson model in which circular RNA expression increased with linear isoform abundance, controlling for effects of sequencing depth and incorporating experimental variation by treating experimental replicates as distinct. We also included cell-type effects to further account for circular RNA expression.
We observed the largest dynamic range in circular RNA production in the gene CYP24A1, a candidate oncogene encoding a component of the vitamin D3 metabolic pathway. Its linear mRNA product and the CYP24A1 protein have been reported to be expressed at elevated levels in many primary lung and other cancers, and in many lung cancer cell lines, including A549; no amplification of the CYP24A1 gene has been reported in A549 [14]–[19], although CYP24A1 is frequently amplified and mutated in primary human cancers [20].
Our statistical model also suggested that other highly expressed RNA circles had cell-type specific increases in expression that could not be accounted for by an increase in overall expression of the corresponding linear RNA. One example is the circular isoform of DOCK1, whose linear isoform encodes a “dedicator of cytokinesis”, a RacGEF, and was the most highly expressed circular isoform in MCF-7, a breast cancer cell line. DOCK1 also had the highest estimated ratio of circular: linear RNA expression in MCF-7 among all cell types in the ENCODE panel, including those where the linear isoform of DOCK1 was more highly expressed (see Figure 6).
Figure 6 depicts other examples of genes with cell-type-specific selective increases in the ratio of circular to linear RNA isoforms. One example was the much higher expression of a circular RNA isoform of RBM33 in K562 cells compared to the other cell types. We have previously detected RBM33 circles in human leukocyte and leukemia samples and mouse brain [1], suggesting the possibility of evolutionary conservation. RBM33 has not been extensively studied, but duplication of a locus including Sonic Hedgehog and RBM33 has been associated with congenital muscular hypertrophy [21]. Similarly, expression of a circular isoform of the long intergenic noncoding RNA LINC00340 was specifically elevated in H1-hESCs. In H1-hESCs, sequencing data suggested that the circular isoform of LINC00340 was the fourth most highly expressed circular RNA of all detected circular isoforms. Circular isoforms of other LINC RNAs, including LINC00263 and LINC00265, were also identified in our analysis. Because noncoding RNAs, including LINC RNAs are generally less well annotated than messenger RNAs, it is possible that our analysis may still have under-detected circular isoforms of these RNAs as we did not specifically attempt to improve their representation in the UCSC knowngene annotation.
A final highlighted example of cell-type-specific selective increases in the ratio of circular to linear RNA isoforms in Figure 6 is AMBRA1. Two different circular isoforms of AMBRA1 RNA were differentially regulated in MCF-7 and HepG2 cells. MCF-7 cells expressed higher levels of a longer isoform (362 nt) while a shorter isoform (182 nt) was more highly expressed in HepG2. AMBRA1 plays a key role in autophagy; deficient mice have excessive cell death by apoptosis [22]–[23].
In these specific examples, and in general, variation in the abundance of hundreds of circular RNA isoforms appeared to have little or no correlation with variation in the abundance of the cognate linear RNA between the cell lines we compared. In particular, we did not observe a correlation between circle-specific junctional counts and overall abundance of the corresponding RNA as measured by RPKM, even at the lowest levels of gene expression. Further evidence that RNA circles are not just an accidental aberration of normal splicing is provided by the fact that circular RNA isoforms are generated by splicing very specific pairs of exons (see discussion below).
When a gene encodes multiple alternatively spliced circular isoforms, what patterns characterize the use of splice acceptor and donor pairs to produce the circle junction? To study this question, we distinguished three broad classes of splice site pairings, which we term stereotyped, proximal and combinatorial pairing, respectively. Examples of each are depicted in Figure 7. For most genes that have circular RNA isoforms (the “stereotyped” class), a single splice site donor and acceptor pair were either used exclusively or strongly preferred in the splice that gave rise to the circular isoforms of the gene; this was the case, for example, for CYP24A1 and MCU. While CYP24A1 was the most highly expressed circular RNA among the examined cell lines and MCU was among the 20 most highly expressed circular RNAs in 9 different cell types, only one circular splice variant from each gene was overwhelmingly preferred (see Figure 7).
A variant of stereotyped splicing was exemplified by the circular RNA isoforms of MBOAT2. These isoforms were expressed at levels similar to MCU, but with a distinctly different pattern of splicing: one particular splice acceptor was highly preferred, but several alternative splice donors were used and each produced similar levels of the corresponding spliced circular RNA isoform. It is noteworthy that none of the exons that participate producing the MBOAT2 circles have been reported to participate in alternative splicing of linear RNA MBOAT2 isoforms.
For many transcripts in which multiple splice donor and multiple splice acceptor sites were used in circular splicing, proximal donor-acceptor pairs were strongly preferred. This “proximal” pattern of circular splicing is exemplified by the circular isoforms of ABCC1.
The “combinatorial” pattern of circular splicing is exemplified by CAMSAP1 and especially PICALM. Multiple splice donors and multiple splice acceptors participate in production of circular isoforms, with little preference for proximal donor and acceptor sites. In contrast to PICALM, across cell types, CAMSAP1 has a single dominant isoform.
Although detection of rare circular RNA isoforms increased with sampling depth of the RNA sequences (see Figure S3), within a gene, our data did not fit a simple model where overall expression of circular RNA isoforms predicted the diversity of circular RNA isoforms expressed (Figure S4). Figure S4 depicts other features of intragenic circular RNA splicing patterns across all genes: the majority of genes with detectable circular RNA expression had detectable levels of more than one circular isoform. Also in such genes, the number of splice donor and acceptor sites used in circular splicing was correlated: when more acceptor sites were used in circular RNA products from a particular gene, so were more donor sites.
Considering all genes with circular RNA isoforms, we found that cells generally expressed a small fraction of the number of circular RNA isoforms that could, in principle, be produced by splicing a downstream splice donor to an upstream splice acceptor (see Figure S4C). We quantified this fraction by defining a combinatorial index C, which compares the number of observed circular isoforms to the number of possible pairings of the detected acceptor and donor splice sites (see Methods). In general, regardless of the total expression level of circular RNA isoforms, half or less of the combinatorial space of circular isoforms—conditioned on acceptor donor and acceptor sites used in circular RNA splicing– had detectable expression, and many genes expressed the minimum number of potential circular RNA isoforms represented by the lowest value of C.
We used a statistical model to identify genes with regulated use of donor and acceptor sites characterizing the diagnostic non-canonical exon junction. For each gene and each cell type, we estimated the frequency with which each donor and acceptor splice site was used (see Tables S4, S5), and computed confidence intervals for the use of each site by cell type.
For hundreds of genes, we found statistical evidence of cell type-specific preferences in patterns of splice donor and acceptor usage in the biogenesis of circular RNA (Tables S4, S5). Three of these genes are shown in Figure 8. The RNF19B gene provides a simple and striking example. The only circular isoform of RNF19B RNA detected in NHLF was undetectable in any of the other cell lines examined. Conversely, the dominant circular RNF19B isoform in the other cells was undetectable in NHLF (see Figure 8). In a second example, a single splice acceptor was used in all circular LPAR1 RNAs identified in NHEK, NHLF and HSMM cells, whereas three different splice acceptors were represented in the circular LPAR1 RNAs found in two fetal fibroblast cell lines, AG04450 and BJ. The differences in diversity of circular isoforms were not explained by cell-type specific differences in overall LPAR1 expression.
ZFAND6 is a third example of a gene with regulated circular RNA expression. A549 cells expressed a single circular isoform, while the remaining cell types expressed two circular isoforms. These differences cannot be readily explained either by differences in sequencing depth, cell-type-specific differences in linear or circular RNA expression or any cell-type independent differences in the RNA, such as intron lengths or structure (see Figure 8). For example, among all the cells we examined, NHLF expressed the second highest levels of linear ZFAND6 RNA, but circular ZFAND6 RNAs were undetectable in these cells. Further, we do not observe any correlation between canonical alternative splicing and circular RNA splice site use or patterns in the three genes depicted in Figure 8 (see Figure S2).
To further assess evolutionary conservation of circular RNA expression across model organisms, we surveyed circular RNA expression using published RNA-Seq data from Drosophila brains [24]. This analysis revealed hundreds of genes encoding circular RNA isoforms in fly, including abundant expression of a previously described circular RNA isoform from the muscleblind locus [3]. Muscleblind was among the most highly expressed circular isoforms, but our analysis indicated that circular RNAs from 11 other genes in these samples were even more abundant: the most highly expressed putative circular RNAs were from Pka-C3, encoding a cAMP-dependent protein kinase, and scarecrow (scro), encoding an NK-2 homeobox protein. Other highly expressed RNA circles included Caps, ps, mGluRA, caps, snap25, jp, zfh2 and two genes of unknown function, CG40178 and CG17471. Overall, we found evidence for exon scrambling in more than 800 distinct Drosophila splice junctions supported by more than one sequencing read (Table S6).
Additional evidence supporting some evolutionary conservation of circular RNAs is found by considering mouse genes represented in brain RNA-Seq data [1]. Genes whose human orthologs expressed circular RNAs were statisteically more likely to have evidence for circular isoforms in the mouse RNA-Seq data. Roughly 4% of genes with expression in both mouse and human data and which encoded orthologous proteins also encoded circular RNA detected in both data sets compared to a null expected rate of 2.5%. This suggests modest conservation of circular RNA expression from loci with orthologous protein products, ignoring finer features that might influence conserved expression of circular RNA, such as developmental stage. In addition, several genes encoding exclusively non-coding RNA species, including IPW (Imprinted in Prader-Willi syndrome) and the oncogene PVT1 were expressed as circular isoforms in both mice and humans.
Characteristic changes during development and differentiation are a pervasive feature of global gene expression programs. We systematically searched for evidence of circular RNAs in a large corpus of RNA-Seq data generated by the ENCODE consortium as well as in RNA-Seq data from Drosophila brain. We found that circular RNA comprises a significant fraction of cellular RNA and that the relative abundance of circular isoforms and the specific patterns of splice site usage in RNA circularization are regulated in a gene-specific and cell-type specific manner. The results strongly suggest that circular RNAs are a common, abundant and potentially developmentally regulated component of the gene expression programs in diverse animal species.
To improve our sensitivity and specificity in detecting circular isoforms, we developed improved bioinformatic and statistical methods that enabled more reliable discrimination between bona fide evidence of exon scrambling and artifacts introduced by high throughput sequencing and sequence homology within a gene. This improved performance allowed us to detect a more extensive catalog of circular RNA than previously reported, including small RNA circles, RNA circles formed by non-canonical splicing of short exons and noncoding RNAs. Improved detection of circular RNA isoforms has also allowed us to characterize the extent of differential circular RNA splicing within a single gene, and to study variation in alternative splicing of circular RNA; indeed, this method may have wider applicability in the discovery of novel RNA splice junctions and detection of other variant sequences.
qPCR quantification and extensive analysis of RNA-Seq data has allowed us to estimate that the number of circular RNA molecules is roughly 1% of the number of poly(A) molecules in the cells we investigated. This estimate is remarkably similar to a report published more than 30 years ago, which found physical evidence of circular RNA by examining cellular RNA by electron microscopy [25].
We tested the hypothesis that circular RNAs might be the result of a background “noise” level of dysfunctional splicing. Under this model, we would expect a positive relationship between linear RNA isoform expression from a given gene and the probability of detecting a circular RNA isoform from that gene. We found no evidence of such a relationship, suggesting instead that even low rates of circular RNA production are regulated, or that highly expressed genes have evolved specific mechanisms to prevent splicing errors that could result in RNA circles.
Our initial report of the ubiquity of circular RNA, based on sequencing an ribosomal-RNA depleted RNA fraction, has since been confirmed in an independent study in which circular RNAs from human and mouse fibroblasts were enriched by treating RNA with RNase R, and in a second genome-wide search for evidence of circular RNA by sequencing ribosomal-RNA-depleted RNA samples [7], [8]. The analysis presented here significantly expands the catalog of circular RNAs expressed by humans and Drosophila. It is likely that human cells express even more circular RNAs than we report here: we did not attempt a ‘de novo’ identification of circular RNA, and instead focused on circular RNA produced by splicing at annotated exon boundaries.
For example, by analysis of a Northern blot for CAMSAP1 in HeLa cells, and a subsequent limited bioinformatic survey of 6 genes, we found evidence of cell-type specific variation in rates of intron retention as well as evidence that circular, intron-retained transcripts in HeLa-S3 cells may be sequestered in the nucleus and potentially exported to the cytoplasm. CAMSAP1 is a calmodulin regulated gene and has conserved circle expression in mouse and Drosophila (spp4), and it would be interesting to study intron retention in these organisms.
The most highly expressed circular RNA identified in our analysis was from the CYP24A1 gene, in a lung cancer cell line, A549. CYP24A1, which encodes 1,25-dihydroxyvitamin D 3 24-hydroxylase, has been suggested to play a role in the pathogenesis of many carcinomas [14], [18], [26]–[32]. We found that the circular isoform was expressed at levels comparable to the canonical linear form. Although the circular isoform of CYP24A1 (which includes all but the first and last exons of CYP24A1) could in principle encode a protein lacking the N terminal mitochondrial localization signal, we have not found evidence for such a protein by mass spectrometry on A549 cell lysates (unpublished data). This finding is consistent with other evidence that despite the formal possibility of translation of circular RNA by the eukaryotic ribosome [33]–[34], circular RNAs do not in general act by encoding a protein.
Recent reports have shown that an antisense circular transcript from the CDR1 locus is enriched for functional microRNA binding sites [7], [8]. However, in a preliminary analysis, we have not found evidence that enrichment of microRNA binding sites is a global feature under selection in the sequence of the thousands of circular RNAs profiled in this paper. For example, we see a roughly 5% enrichment of microRNA binding sites in a 66 nt window in sequence flanking diagnostic circular RNA junctions in circular RNAs which are highly expressed in at least one cell type compared to the number of binding sites in the junctional sequences flanking all detected circular RNA junctions (see Figure S5). The low enrichment is perhaps not surprising considering that the vast majority of the transcripts we surveyed in [1] and in this report are transcribed in the same sense with respect to the linear mRNA isoform. Therefore, except for ‘diagnostic’ junctional sequence and secondary and higher order structure, circular and linear isoforms would have the same potential to bind microRNA, albeit with a different degree of stability. Certainly, the potential genome-wide interplay between microRNAs and circular RNAs warrants further experimental and computational investigation.
Our findings that mouse orthologs of human genes with circular RNA products are themselves more likely to encode circular RNAs are consistent with a similar independent analysis of circular RNA conservation and support the hypothesis that circular RNAs have an evolutionarily conserved function [7]. Thus, although the abundance, ubiquity, and potential developmental regulation of circular RNAs all point to the possibility of important functional roles, their nature and mechanisms are still to be discovered.
Raw fastq files available on Sept. 3, 2012 were downloaded from the ENCODE project website and processed in batch using custom Perl scripts. At that time, 2 replicates from each of 15 cell types were available, with the exceptions that 1 HMEC and 3 NHEK were downloadable. We selected all long poly(A) minus reads banked at http://hgdownload.cse.UCSC.edu/goldenPath/hg19/encodeDCC/wgEncodeCshlLongRnaSeq. Read 1 and Read 2 reflect directionality of original RNA and were not processed symmetrically.
We constructed a custom database for sequence alignment as follows: all UCSC annotated exons in scrambled order were identified and for each pair, 66 nt from each side the 3′ and 5′ ends of flanking exon were concatenated. Sequence alignment of a 76 nt read hence required alignment with a minimum of 10 nt overhang. Cases where exons were <66 nt were treated separately, by concatenating exons upstream of the donor or downstream of the acceptor exon in the scrambled pair, or using ‘in silico’ rolling circle concatenation in cases where the total circle size was smaller than 132 nt.
In detail, read 1 and read 2 were not treated symmetrically as the input library was a directional RNA-Seq library. Read 1 was aligned to UCSC knowngenes and the human genome under bowtie2 default conditions [35]. Reads failing this alignment were aligned to a custom database of all scrambled exon-exon junctions for each UCSC knowngene isoform, again under bowtie2 default conditions. We culled the mate of each aligned read 1, and performed an alignment of this subset of reads to the above UCSC knowngenes and to the custom database of scrambled exon-exon junctions: (thus, in principle, we could have analyzed the data as described above focusing on read 2 and increased the number of detected junctional reads).
In conjunction with the alignment to the above database of exon-exon junctions, we modeled the null distribution for rates of mismatch of reads aligning to this database using a method that should be of general interest in discovery of structural variants using high throughput sequencing data. In overview, we considered all reads that aligned with qualities a) and b) below, without imposing hard thresholding on the quality of alignment of either read:
A read reflecting a circular RNA isoform transcribed from the same strand as the canonical isoform has the property that read 1 maps to the −orientation and read 2 to the +orientation.
Alignment scores were calculated using the bowtie2 default which for example, adds a ‘−6’ penalty for a mismatch between the reference and aligned read at a high quality base call. Summing penalties for mismatches produces an overall alignment score per read, one score for the read spanning junction (read 1) and one score for read 2. Three statistics measuring alignment were calculated for each pair of scrambled exons for each UCSC isoform supported by at least one read: the average alignment score of read 1, the average alignment score of read 2, and the average product alignment score of read 1, read 2 although this measure was not ultimately used to calculate the FDR reported.
In detail, to compute the FDR, we created a null distribution of the joint alignment statistics for read 1 and read 2 using reads where read 1 mapped to a junction between exon x and exon y (x> = y) and read 2 mapped upstream of exon y or downstream of exon x which is incompatible with it deriving from a circular RNA molecule. We used the pair of (read 1, read 2) alignment statistics from such reads to generate the FDR per isoform depicted in Figure 1.
Subsequently, all reads were filtered to an FDR level of .025 unless otherwise specified. See Table S7 for a complete list of scores.
HeLa total RNA was isolated by TRIZOL lysis followed by PureLink purification of the aqueous phase (Life Technologies). 2 micrograms of total RNA was treated in a 10 microliter reaction with 0 units (mock treatment) or 20 units of RNase R (Epicentre) in 1× RNase R buffer, 1 unit/microliter murine Ribonuclease Inhibitor (New England Biolabs), and incubated at 37DEGC for 1 hr. 1 microliter 1 mM EDTA, 1 microliter 10 mM each dNTP, and 1 microliter 100 microM random hexamer were added and the RNA denatured at 65DEGC for 5 min and placed on ice. 4 microliters 5× buffer (250 mM Tris-HCl pH 8, 125 mM KCl, 15 mM MgCl_2), 1 microliter murine Ribonuclease Inhibitor (40 units/microliter), and 1 microliter Superscript III (LIfe Technologies) were added; this cDNA reaction was incubated at 25 deg C 10 min, 50 deg C 50 min, 55 deg C 10 min, 85 deg C 5 min, 4 deg C hold. 0.5 microliter cDNA reaction was used as the template for qPCR and fraction resistant was computed as 2∧(RNase R C_t - Mock C_t).”
HeLa-S3, A549 and AG04450 cells were grown in standard media and conditions. RNA was harvested by lysing cells with the TRIZOL reagent and purified on Purelink columns under ethanol concentrations that retain small and large RNAs. Total RNA reaction was reverse transcribed using the SuperScript III First-Strand Synthesis System (Life Technologies, Carlsbad, CA) with random hexamers according to the manufacturer's instructions. 500 ng/ul of cDNA was then used for each qPCR validation; dilution series were performed at concentrations of .5, 5, 50 and 500 ng/ul. Each qPCR experiment was done in biological duplicate with 3 technical replicates each.
For each cell type, we downloaded gtf files with gene level RPKM (Reads per kilobase mapped) estimates from the Poly(A) fraction; eg. for SK-NS-H_RA, we downloaded the file: http://hgdownload.cse.UCSC.edu/goldenPath/hg19/encodeDCC/wgEncodeCshlLongRnaSeq/wgEncodeCshlLongRnaSeqSknshraCellPapGeneGencV7.gtf.gz
RPKMs were summed across all genes to estimate total annotated poly(A) transcript abundance. In parallel, we summed all reads mapping to a circular RNA junction. These two values provide total abundance estimates of poly(A) and circular RNA respectively, up to a normalizing constant. We determined that normalizing constant by performing qPCR with two calibrating genes: FAT1 and HIPK3. These genes were chosen for the reasons described in the main text. Standard curves were computed for each set of primers listed in Table S8 and used to compute relative expression of linear and circular RNA at the log scale. The difference between the log base 2 of calculated junctional circle counts and log base 2 RPKM and these differences was averaged for the 2 genes, and raised to the power 2 in order to normalize measurements. To test robustness of our estimates, we also performed the above analysis by imposing a filter on circles that could contribute to our estimate of total circle mass. That filter required a circular isoforms have greater than 5 counts in the cell type under consideration. This provided a conservative estimate of the moles of circular vs. poly(A) RNA. Using this filter, we obtained estimates of .6%, 2% and .6% for HeLa, A549 and AG04450 respectively, and is consistent with what we estimate without this filter.
For each circular isoform represented by at least 50 counts in one sample (and satisfying an FDR cut-off of .025), we fit a Poisson model per gene modeling circle counts by poly(A) gene expression (genexp), celltype and total circles (totcircles) using the glm poisson model in R with the formula:
cir∼log(genexp)+celltype+totcircles−1. Coefficients in this model were used to choose genes shown in Figure 6 and the table of values is listed in Table S9.
10 micrograms total RNA was denatured with glyoxal and run on a 2% agarose gel [36], followed by alkaline capillary transfer onto Brightstar-Plus nylon membrane (Ambion). 32P-labeled probe was generated by random-priming (Prime-It II kit, Stratagene) of a PCR product corresponding to exons 2 and 3 of CAMSAP1 (nt 161–423 of GenBank # NM_015447.3) and hybridized in modified Church buffer (0.5M sodium phosphate pH 7.2, 7% SDS, 10 mM EDTA) at 65DEGC for 16 hr. After washing, the blot was visualized by phosphorimaging (Typhoon, Molecular Devices).
The following procedure was used to access statistical significance of the use of donor and acceptor sites across cell types. We analyzed donor and acceptor sites separately. Splice sites represented by more than 50 counts (and satisfying an FDR cut-off of .025) in at least one cell type were analyzed using this approach. For each such donor and acceptor site that was supported by more than 5 reads and present in at least two cell types, we computed an exact .999 binomial confidence interval for its probability of use in that cell type. Sites with at least one pair of non-overlapping confidence intervals were identified and used to choose genes depicted in Figure 8. Cell types were collapsed over replicates. A table of all confidence intervals by site and cell type is listed in Table S4.
Poly(A) depleted RNA isolated from Drosophila brain, available at ftp://ftp-trace.ncbi.nlm.nih.gov/sra/sra-instant/reads/ByStudy/litesra/SRP/SRP007/SRP007416/ was aligned to a custom database of annotated Drosophila exon-exon junctions using Jan. 2012 flyBase exon annotation and previously described methods and filters. A complete list of detected events is listed in Table S6.
To access evolutionary conservation, orthology of protein products was defined by Inparanoid using the following databases: http://inparanoid.sbc.su.se/download/7.0_current/sqltables/
For statistical assessment of expression of circular isoforms between mouse and human, a list of orthologous genes expressed (measured by detected gene expression from RNA-Seq data sets used to measure circle abundance) was compiled (a total of 1402 genes). We then counted the number of genes in this table with more than 1 circle count in the mouse RNA-Seq data and with expression in the top 100 most expressed circular isoforms in one ENCODE (human) experimental replicate. 57 genes matched this criterion (4%). 147 genes on the list of 1402 were in the top 100 most expressed circular isoforms in one experimental replicate; 332 had more than 1 count in the mouse RNA-Seq data. Under an independence model, we expected 35 genes to match the joint criterion (2.5%). The residual from a chi-squared test for the independence model is (O-E)/sqrt(E) = 3.7, which corresponds to a one sided p value of .0001.
Read counts were summed across cell types and replicates for each isoform, defined here as a unique combination of gene, circularization splice donor coordinate, and circularization splice acceptor coordinate. Isoforms were filtered by requiring 20 or more read counts in total across ENCODE cell lines. Intron length was computed as described in the text. For each gene, intron lengths were considered either as fractions of the longest intron length within the gene, or as quantile ranks within the gene.
Read counts were summed across replicates for each unique combination of cell type, gene, circularization splice donor coordinate, and circularization splice acceptor coordinate.
Read counts were summed across cell types and replicates for each isoform, defined here as a unique combination of gene, circularization splice donor coordinate, and circularization splice acceptor coordinate. Genes where some annotated isoforms satisfied splice acceptor<donor and other isoforms satisfied donor<acceptor were removed from consideration; the remaining genes were then oriented such that all isoforms were acceptor upstream of donor.
For each gene, the combinatorial index C compares the number of observed circular isoforms to the number of possible pairings of the detected acceptor and donor sites; C = 1 means that all possible pairings were actually detected, whereas C = 0 means that the minimum possible number of pairings was detected (we adopted the convention that C is undefined when max. poss. isoforms = min. poss. isoforms). Precisely, C was defined for each gene as (# of distinct circular isoforms detected – min. poss. isoforms)/(max. poss. isoforms – min. poss. isoforms), where min. poss. isoforms = max(# of distinct acceptor sites detected, # of distinct donor sites detected), and max. poss. isoforms = # of combinations of 1 detected acceptor and 1 detected donor in which the acceptor is upstream of the donor. C is evaluated in Figure S4.
For each cell type and replicate, isoforms were defined here as unique combinations of gene, circularization splice donor coordinate, circularization splice acceptor coordinate, read 1 orientation, and read 2 orientation.
Instances in which the junction-defining read (“read 1”) and its mate-pair read (“read 2”) were on the same strand were removed from consideration. For each cell type, replicates were pooled, and isoforms were defined here as unique combinations of gene, circularization splice donor, circularization splice acceptor, and read 1 orientation. The percentage of circular isoforms in the sense orientation (with respect to the linear isoform) is 96% for HMEC and >99% for the other 14 cell types.
We downloaded a list of all high confidence microRNAs (mature.fa from http://mirbase.org/ftp.shtml annotated as ‘Homo sapiens’) and corresponding 6mer seed match (nt 2–7). For each possible non-canonically ordered exon X, exon Y pair within a transcript in the UCSC knowngene transcript database (enumeration beginning at 0) to 30, we generated a corresponding 132 nt sequence consisting of 66 nt upstream and 66 nt downstream of the exon-exon junction. For each group of exon X -exon Y sequences, we compared the number of microRNA seed matches per offset (from 0 to 126) divided by the total number of junctions evaluated. We compared these statistics for circular junctions expressed at rank <1000 in at least one cell type, ranked based on aligning paired end sequencing reads to a database of all UCSC knowngene exon-exon junctions and all other expressed circular junctions (see Figure S5). The rate of enrichment averaged 1.05 and was never more than 1.25 per offset. While this analysis does not strictly consider all microRNA binding sites within a circle, it samples a window including circular RNA sequence that, under most basic models where circular RNA were under selection to serve as a microRNA sink, would be enriched for microRNA seed matches.
|
10.1371/journal.ppat.1000944 | Two Novel Point Mutations in Clinical Staphylococcus aureus Reduce Linezolid Susceptibility and Switch on the Stringent Response to Promote Persistent Infection | Staphylococcus aureus frequently invades the human bloodstream, leading to life threatening bacteremia and often secondary foci of infection. Failure of antibiotic therapy to eradicate infection is frequently described; in some cases associated with altered S. aureus antimicrobial resistance or the small colony variant (SCV) phenotype. Newer antimicrobials, such as linezolid, remain the last available therapy for some patients with multi-resistant S. aureus infections. Using comparative and functional genomics we investigated the molecular determinants of resistance and SCV formation in sequential S. aureus isolates from a patient who had a persistent and recurrent S. aureus infection, after failed therapy with multiple antimicrobials, including linezolid. Two point mutations in key staphylococcal genes dramatically affected clinical behaviour of the bacterium, altering virulence and antimicrobial resistance. Most strikingly, a single nucleotide substitution in relA (SACOL1689) reduced RelA hydrolase activity and caused accumulation of the intracellular signalling molecule guanosine 3′, 5′-bis(diphosphate) (ppGpp) and permanent activation of the stringent response, which has not previously been reported in S. aureus. Using the clinical isolate and a defined mutant with an identical relA mutation, we demonstrate for the first time the impact of an active stringent response in S. aureus, which was associated with reduced growth, and attenuated virulence in the Galleria mellonella model. In addition, a mutation in rlmN (SACOL1230), encoding a ribosomal methyltransferase that methylates 23S rRNA at position A2503, caused a reduction in linezolid susceptibility. These results reinforce the exquisite adaptability of S. aureus and show how subtle molecular changes cause major alterations in bacterial behaviour, as well as highlighting potential weaknesses of current antibiotic treatment regimens.
| The treatment of serious infections caused by Staphylococcus aureus is complicated by the development of antibiotic resistance, and in some cases the appearance of more persistent bacteria that have a reduced growth rate resulting in small colony variants (SCV). Here we have shown using whole genome sequencing and gene replacement experiments on sequential S. aureus isolates obtained from a patient with a serious bloodstream infection, how S. aureus evolved into a multi-antibiotic resistant, persistent and almost untreatable SCV. Specifically we show that a minor DNA change in a S. aureus gene encoding an enzyme called RelA causes an accumulation of a small signalling molecule called (p)ppGpp, which in turn leads to persistent activation of the important bacterial stress response known as the stringent response. This is the first report of the involvement of the stringent response in S. aureus SCV formation and its association with persistent infection. Additionally, we have uncovered a novel mechanism of resistance to the new antimicrobial linezolid, caused by a mutation in a gene encoding a 23S rRNA methyltransferase. This study highlights the exquisite adaptability of this important pathogen in the face of antimicrobial treatment.
| The factors promoting persistence of bacterial infection in the face of apparently effective antimicrobial therapy have not been clearly defined. This particularly applies to Staphylococcus aureus, especially methicillin-resistant S. aureus (MRSA), which remains a major human pathogen that frequently causes invasive disease, often associated with a high mortality rate [1], [2], [3]. A number of bacterial factors have been associated with persistent bacteremia and failed antimicrobial therapy for serious MRSA infections, including reduced activity of the quorum sensing system agr, resistance to host antimicrobial peptides, and the evolution of reduced vancomycin susceptibility in patients treated with this antibiotic [4], [5]. Although traditionally considered an extracellular organism, recently it has been demonstrated that S. aureus can reside and persist in an intracellular state [6]. A staphylococcal phenotype that appears to be particularly associated with cellular invasion and clinical persistence is the small colony variant (SCV) phenotype [6], [7]. This is phenotypically characterised by reduced growth rate, small colony size and in some cases auxotrophism for hemin or menadione, related to mutations in genes encoding products involved in the electron transport system. Small colony variants of S. aureus have been associated with persistent and recurrent S. aureus infections, and with increased antimicrobial resistance [7]. The mechanisms of the SCV phenotype in S. aureus have been investigated in detail over a number of years. Defined hemB and menD mutants [6], [7] of laboratory S. aureus strains have defects in electron transport, and have demonstrated global transcriptional changes [8], increased cellular attachment, invasion and persistence [9], [10], [11], reduced antibiotic susceptibility [12], and reduced virulence [13]. However, despite this significant work the molecular correlates of persistence have not been definitively elucidated in clinical isolates of S. aureus.
One important bacterial response to stress and nutritional starvation, including antimicrobial challenge, is activation of the stringent response, mediated by intracellular accumulation of the alarmones ppGpp and pppGpp [(p)ppGpp], which are usually controlled by the activities of a synthetase (RelA) and a hydrolase (SpoT) [14]. In gram positive organisms, including S. aureus, a single rel gene encodes a protein with synthetase and hydrolase domains that controls the stringent response under stressful conditions, while other synthetases such as RelP and RelQ provide basal levels of (p)ppGpp during non stressful conditions [14], [15], [16], [17]. The stringent response has been associated with persistence of infection in Mycobacterium tuberculosis, where it is important for the long term survival of the organism [18], and recently has been linked with growth defects and vancomycin tolerance in E. faecalis [17]. However, different bacteria have developed different strategies to utilize the alarmones in intracellular signalling, with diverse regulatory changes found in different organisms [19]. The impact of an active stringent response has not been studied in S. aureus, mainly because a functional hydrolase domain of the RelA/SpoT homolgue in S. aureus (RelA) is essential for survival of the organism [19], [20], [21]. Although mupirocin is a strong inducer of the stringent response in S. aureus and has been used to investigate the transcriptional profile of an active stringent response in this organism, it also leads to RelA/SpoT independent transcriptional changes [19], [22], indicating that the mupirocin model alone is not an optimal strategy to study the stringent response in this organism. Although it could be anticipated that the bacterial stringent response would play a role in the adaptation of S. aureus to antimicrobial challenge during persistent infection, this has not been previously reported.
For many years vancomycin has been the mainstay of therapy for serious MRSA infections [5]. However, with increasing antimicrobial resistance in MRSA, including resistance to vancomycin, the newer, novel antimicrobials such as linezolid and daptomycin are the last available therapies in some patients [5]. While there are increasing reports of daptomycin non-susceptible S. aureus strains [23], reports of reduced linezolid susceptibility in S. aureus have, to date, been rare. Linezolid is the first in a new class of antimicrobials, the oxazolidinones, that bind to the A site of the peptidyl transferase centre (PTC) of the bacterial ribosome [24], inhibiting bacterial ribosomal protein synthesis. Resistance to linezolid in S. aureus has primarily been related to target site mutations in domain V of 23S rRNA, especially the G2576U mutation [25], [26]. Recently, however, a naturally occurring resistance gene cfr, which encodes Cfr methyltransferase and leads to modification of adenosine at position 2503 in 23S rRNA has been described in a single S. aureus isolate from Columbia, and in two staphylococcal clinical isolates from the USA [27], [28]. The cfr gene on the chromosome was associated with mobile genetic elements, suggesting the resistance mechanism may be transferable [27]. The S. aureus genome encodes a number of conserved RNA methyltransferases, including RlmN (encoded by SACOL1230), and although methylation of rRNA is a common mechanism of acquired antimicrobial resistance [29], mutations in chromosomally encoded RNA methyltransferases have not been linked to reduced linezolid susceptibility in S. aureus.
Recently, we treated a patient with persistent and recurrent methicillin-resistant S. aureus (MRSA) bacteremia despite extensive, appropriate antimicrobial therapy. The clinical isolates obtained following treatment demonstrated significant antimicrobial resistance, including reduced susceptibility to linezolid, and features characteristic for small colony variant strains (SCV) of S. aureus [6], [7]. However, phenotypic features suggested that mutations were not present in hemin or menadione biosynthesis genes. Therefore we investigated the mechanisms of persistence and antimicrobial resistance in these isolates using a combined comparative and functional genomics approach, and discovered a clinical isolate with a persistently activated stringent response, and a novel mechanism of reduced linezolid susceptibility.
A 73-year old man with end-stage renal failure was admitted with line related methicillin-resistant S. aureus (MRSA) bacteremia. The MRSA was susceptible to clindamycin, trimethoprim-sulfamethoxazole, ciprofloxacin, vancomycin, rifampicin and fusidic acid. He was commenced on intravenous vancomycin, and due to persistent S. aureus bacteremia, rifampin and ciprofloxacin were added. After 16 days of ongoing bacteremia and detection of heterogeneous vancomycin-intermediate S. aureus (hVISA), vancomycin was changed to oral linezolid and he completed 18 days of linezolid combined with rifampicin and ciprofloxacin. Multiple investigations including transesophageal echocardiogram, computed tomography of brain, chest, abdomen, pelvis and lumbar spine, and white cell/SPECT imaging did not reveal any definite focus. Eleven days later he developed fever, hypotension and back pain and blood cultures were again positive for MRSA, on this occasion a small colony variant (SCV). He was recommenced on oral linezolid and completed 6 weeks of therapy. Five days later he developed severe lumbar back pain and raised inflammatory markers. A single blood culture and a lumbar aspirate from the L3–4 region again cultured SCV-MRSA (Fig. 1). He was commenced on intravenous linezolid and completed 6 weeks of therapy, and was changed to trimethoprim-sulfamethoxazole for long-term suppressive treatment. Pulsed field gel electrophoresis demonstrated that the SCV strain JKD6229 emerged from the parental strain (JKD6210) [5] (Fig. 1), but JKD6229 did not demonstrate auxotrophism for hemin or menadione [7]. Both strains were multi-locus sequence type 5. During failed therapy, resistance to ciprofloxacin, rifampin and reduced susceptibility to linezolid developed (Table 1).
Initially, to understand the molecular determinants of clinical persistence in the SCV-MRSA isolate JKD6229 the transcriptional profile was analysed. Using microarray transcriptional analysis significant global gene expression changes were found in the SCV strain (JKD6229) compared to the parental strain (JKD6210) (349 genes up-regulated and 175 genes down-regulated ≥2-fold (see Fig. 2, Table 2 and Table S1). Changes included pronounced up-regulation (up to 80-fold) of genes encoding capsule biosynthesis in JKD6229 (cap5A to cap5P; SAV0149 to SAV0164). To confirm the biological impact of capsule gene transcriptional changes the capsule type of JKD6210 and JKD6229 was confirmed as type 5 by PCR [30], and a capsule immunoblot was then performed. This demonstrated significant enhancement of capsule production in JKD6229 compared to the parental strain JKD6210 and the capsule type 5 control strain Newman (Fig. 2B).
Intracellular persistence and the SCV phenotype of S. aureus has previously been associated with down regulation or complete loss of activity of the global quorum sensing accessory gene regulator (agr) [31], however all genes encoding the agr locus (SAV2036 to SAV2039) and the delta-hemolysin precursor (SAS1940a) were significantly up-regulated in the SCV strain JKD6229 (2 to 10-fold increased expression). Associated with this was up-regulation of two genes encoding exotoxins (alpha-hemolysin [SAV1163], 5.9-fold increase; enterotoxin P [SA1761], 2.8-fold increase), however the SAV1163 orthologue in JKD6210 and JKD6229 was found to be a pseudogene because of a point mutation introducing a premature stop codon.
Distinct differential regulation of genes involved in carbohydrate transport and metabolism, amino acid metabolism and oligopeptide transport was also detected. Genes involved in lactose utilization and galactose metabolism (SAV2189-SAV2194) were remarkably down-regulated (up to 100-fold), with similar changes found in genes encoding key glycolysis enzymes such as pgi (SAV0962, glucose-6-phophate isomerase), while genes with products potentially involved in metabolism of alternative carbon sources such as sucrose, fructose and galactitol (scrA, gatC, fruA, fruB) were up-regulated in SCV JKD6229. Amino acid metabolism was another distinct functional class that was up-regulated in SCV JKD6229. Genes encoding valine, leucine and isoleucine biosynthesis enzymes showed increased expression, as did genes such as rocD (SAV0957) and argJ (SAV0183) linked to ornithine and arginine production. Striking too was the up-regulation of the Opp3 oligopeptide transport system (SAV0986–SAV0994). Opp3 facilitates the acquisition 4–8 aa-long peptides from the extracellular environment and it is the only known functional oligopeptide transport system in S. aureus [32].
The prominent transcriptional changes detected in JKD6229 compared to JKD6210 suggested SCV JKD6229 had undergone important genetic changes. In addition, the transcriptional profile of SCV JKD6229 was significantly different to the transcriptional profile of the SCV hemB mutant, suggesting that mutations in other genes may be contributing to the SCV phenotype of this strain [8]. Therefore whole genome sequencing and comparison of the parental MRSA strain JKD6210 and the SCV JKD6229 was performed. Illumina short-read sequencing yielded 2.7 Mb of mappable data for each genome. After detailed reciprocal sequence comparisons and comparisons against the reference genomes S. aureus COL and S. aureus N315, the only changes detected in JKD6229 compared to JKD6210 were two nucleotide substitutions, two codon insertions, and the loss of a ∼15 kb plasmid (Table 3). The sequences for JKD6210 and JKD6229 were aligned to the genome sequence of N315 (also MLST 5–the same as JKD6210 and JKD6229) and this demonstrated that 97.2% of N315 was covered to a depth of ≥20 in both JKD6210 and JKD6229. The regions not covered in N315 by the JKD6210 sequence were the same as the regions not covered by the JKD6229 sequence. PCR and Sanger sequencing confirmed the presence of each mutation in SCV JKD6229, and PCR and plasmid analysis confirmed the loss of the plasmid in JKD6229. Annotation and BLAST analysis of the plasmid (denoted as pJKD6210) revealed a pUSA300-HOU-MS-like replicon (Fig. 3) with genes encoding beta-lactam resistance, but absence of the genes encoding cadmium resistance that are present on pUSA300-HOU-MS [33]. All four changes in nucleotide sequence were associated with a predicted amino acid change or addition, suggesting one or more of these mutations might be responsible for the phenotypic changes in JKD6229.
Two of the four mutations clearly corresponded with the acquired antibiotic resistance of SCV JKD6229. The change in rpoB that led to a H481Y substitution is a mutation commonly linked with rifampin resistance in S. aureus [34] and the amino acid insertion in parC (encoding Topoisomerase IV) likely contributed to reduced ciprofloxacin susceptibility in this strain. Single mutations in topoisomerase IV without additional mutations in DNA gyrase are often associated with low-level quinolone resistance [35], as demonstrated in JKD6229 (Table 1).
The ‘CAA’ insertion in SACOL1230, encoding RlmN, a ribosomal RNA large subunit methyltransferase, was associated with the linezolid exposure of SCV JKD6229. RlmN methylates 23S ribosomal RNA at adenosine 2503, and deletion of the gene renders S. aureus more susceptible to linezolid [36]. Linezolid resistance in clinical isolates of staphylococci is often linked to G2576T mutations in domain V of the 23S rRNA genes [37] or acquisition of a plasmid-encoded cfr (methyltransferase), which also methylates ribosomal RNA at position 2503 [28]. Interestingly, a 23S rRNA T2500A mutation has also been previously linked to linezolid resistance in a clinical isolate of S. aureus [38], but mutations in SACOL1230 have never been reported. The ‘CAA’ insertion in JKD6229 is predicted to incorporate an additional glutamate to the motif (DIDACCGQ’Q’) at the extreme C-terminus of the enzyme, a motif that is absolutely conserved among diverse Gram positive and negative bacteria [36]. An allelic replacement experiment was performed, where the normal SACOL1230 sequence from JKD6210 was replaced with the mutated SACOL1230 allele from JKD6229 using pKOR1 [39]. The SCV clinical strain (JKD6229) and the mutant JKD6300 (JKD6210 with SACOL1230 ‘CAA’ insertion) both demonstrated an increase in linezolid MIC within the susceptible range when an Etest using a 2 McFarland inoculum was used (Table 1).
The fourth mutation occurred in SACOL1689. Based on high amino acid sequence similarity (71% amino acid similarity) to an ortholog in Streptococcus mutans, SACOL1689 (relA) is predicted to encode a bifunctional enzyme that modulates the amount of the intracellular signalling molecules guanosine 3′-diphosphate 5′-triphosphate and guanosine 3′, 5′-bis(diphosphate), abbreviated to (p)ppGpp [40]. Accumulation of (p)ppGpp, activates the bacterial stringent response leading to a switch to “survival mode” [17], [18]. The mutation in relA (SACOL1689) was of particular interest with respect to SCV formation because of the involvement of this gene in the bacterial stringent response and the potential impact on growth characteristics of an enhanced stringent response. Additionally, the microarray transcriptional profile of JKD6229 suggested the stringent response was active in this strain, with upregulation of amino acid catabolism pathways, significant over expression of genes encoding oligopeptide transport proteins, up-regulation of genes associated with isoleucyl tRNA limitation (including ilvB and ilvD, 4 to 6-fold increase; and leuABCD, up to 7-fold increase) [41], over expression of genes encoding extracellular proteases (hrtA, splB, SAV1612, SAV1613), and up-regulation of the quorum sensing system agr. The transcriptional profile was very similar to the profile of S. aureus after in vitro induction of the stringent response by exposure to mupirocin, an agent that inhibits isoleucyl tRNA synthetase [22].
Structural and functional studies of RelA in Streptococcus mutans (RelASm) have shown that the enzyme can modulate intracellular levels of (p)ppGpp through a N-terminal hydrolase domain and C-terminal synthetase domain that act antagonistically in a ligand-dependant manner, either by degrading (p)ppGpp within the hydrolytic domain or converting GDP or GTP to (p)ppGpp within the synthetic domain [40]. Scanning mutagenesis of RelASm has defined regions of the enzyme that are critical for its hydrolytic function [40]. An alignment of the N-terminus of RelA from SCV JKD6229 (RelASCV) with RelASm shows that the F128Y mutation occurred in a region known to be critical for hydrolase function in S. mutans (Fig. 4A). Thus, the alignment data and microarray results suggested that RelASCV might have impaired (p)ppGpp hydrolase function leading to an accumulation of (p)ppGpp and the persistent activation of the stringent response. To confirm that the RelA F128Y mutation was causing accumulation of (p)ppGpp, the relA allele from SCV JKD6229 was introduced into the parental strain JKD6210, and ppGpp levels were measured using the fluorescent chemosensor PyDPA [42]. As predicted, a significant increase in ppGpp levels was demonstrated in JKD6229 (clinical SCV) and the mutant RelA F128Y mutant JKD6301, compared to the parental strain JKD6210 (Fig. 4B and C), suggesting that the F128Y mutation reduced the hydrolase activity of RelA. This is the first time an activated stringent response has been implicated as a mechanism of SCV formation in clinical S. aureus.
The phenotypic features and impact on virulence of a persistently activated stringent response have not been previously investigated in S. aureus, because of the inability to generate a mutant strain without RelA hydrolase activity [19]. Therefore, the discovery of the clinical strain JKD6229 with the active stringent response, and creation of the relA mutant JKD6301 provided a unique opportunity to investigate the active stringent response in this organism. A number of phenotypic characteristics were investigated (Fig. 5 and 6). JKD6301 demonstrated a reduced growth rate in MH broth, and reduced colony size on HBA after 24 hours incubation indicating that the relA mutation contributed significantly to the growth defect of the clinical SCV strain JKD6229 (Fig. 5A and B). An analysis of vancomycin susceptibility in JKD6301 using macromethod Etest [5], and population analysis profile (data not shown) demonstrated no increase in vancomycin resistance in the mutant compared to JKD6210 (Table 1), indicating that although the stringent response has been linked to vancomycin tolerance in E. faecalis [17], the relA mutation alone was not responsible for the reduced vancomycin susceptibility in JKD6229. Susceptibility to other antimicrobials was also unchanged in JKD6301 compared to JKD6210 (Table 1).
Previous studies have demonstrated enhanced invasion and persistence of some S. aureus SCV strains [6]. Therefore, the attachment, invasion, and persistence potential of the clinical isolate pair, and the relA mutant strain were tested (Fig. 5C, D). Bacterial attachment to HeLa cells was decreased in JKD6229 and JKD6310 compared to the parental strain JKD6210, while invasion was increased only in the SCV strain JKD6229, indicating that the stringent response promotes factors that facilitate bacterial attachment but these changes alone are not sufficient to enhance invasion. In contrast to reported studies of electron transport deficient SCV strains, after 72 hours incubation there was no difference in intracellular persistence of SCV JKD6229 compared to the other strains. This observation might reflect the activated agr expression and increased toxin gene expression in JKD6229 (Table 2) which is unusual for SCV S. aureus where reduced agr expression and alpha-toxin expression is thought to promote intracellular persistence without lethal effects on the host cell [7], [31], [43].
The larval stage of the Greater Wax Moth (Galleria mellonella) is an invertebrate model used to assess S. aureus virulence [44]. A comparison in this model of the virulence of parental strain JKD6210 with SCV JKD6229 and the relA mutant JKD6301 demonstrated a marked reduction in virulence in the SCV strain JKD6229 and also in the relA mutant (Fig. 6). The attenuation of SCV JKD6229 and JKD6310 was not due to their reduced growth rate compared to JKD6210 because the infected larvae had equivalent bacterial burden after 48 hours incubation. These experiments indicate that the increased persistence of SCV JKD6229 is associated with a reduced ‘virulence’ phenotype caused by the relA mutation.
In this study comparative and functional genomics has demonstrated the remarkable adaptive response of S. aureus to antimicrobial challenge during chronic infection, where four point mutations were sufficient to permit the strain to persist and resist multiple antibiotic therapies. This confirms the role of sequential point mutations in S. aureus adaptation during persistent infection, initially described by Mwangi et al [45]. We have uncovered a novel mechanism of growth inhibition contributing to SCV formation by S. aureus through mutation of relA and activation of the stringent response, and have described for the first time phenotypic features of an active stringent response in S. aureus, associated with profound global transcriptional changes. Analysis of the stringent response in S. aureus has been previously hampered by the inability to generate relA knock-out strain in this organism [21], confirming the unique nature of the naturally occurring clinical isolate JKD6229. Here, using the clinical strain JKD6229 and a mutant with a single base swap in relA (JKD6301), we demonstrate that an active stringent response in S. aureus leads to a reduced growth rate and features characteristic of SCV strains, as well as attenuated virulence in the G. mellonella invertebrate infection model. These data contrast with a recent report describing attenuated virulence of a S. aureus Rsh synthetase mutant in a murine infection model [20], suggesting that both persistent activation or inactivation of the stringent response is associated with attenuated virulence in S. aureus. The specific impact of the relA mutation was clearly demonstrated by replicating the same single nucleotide change in relA from the SCV strain JKD6229 into the parental strain JKD6210, and measuring the cellular levels of ppGpp using the fluorescent chemosensor PyDPA (Fig. 4). It is likely that the mutation detected in relA of JKD6229 partially impairs hydrolase function of the enzyme, leading to accumulation of the alarmones (p)ppGpp, but not cell death as has been described following complete loss of hydrolase function [19]. Despite the attenuated virulence of the strain in the invertebrate model, it was associated clinically with a persistent infection, suggesting that the mutation leading to permanent activation of the stringent response in this strain may have provided a survival advantage during chronic infection. Further analysis of the clinical impact of an active stringent response in S. aureus is now needed, with particular focus on the impact of this response on bacterial immune evasion, persistence and response to antimicrobial treatment.
In addition to an activated stringent response the clinical strain JKD6229 harboured a number of mutations leading to the reduced antimicrobial susceptibility that also promoted persistent infection. Most intriguingly, we have described for the first time a codon insertion in the methyltransferase gene SACOL1230 (RlmN) that reduces linezolid susceptibility in clinical S. aureus. Early reports of linezolid resistance in S. aureus, and other Gram positive organisms, suggested that mutations in domain V of 23S rRNA are primarily responsible for resistance [25], [26], in particular the G2576T mutation, which continues to be detected in resistant strains from multiple S. aureus lineages [46], [47], [48]. Recently, mutations in ribosomal proteins L3 and L4 have also been associated with linezolid resistance in staphylococci [48], [49], and it has also become apparent that changes in ribosomal methylation can affect susceptibility to linezolid and other antimicrobials in S. aureus and other organisms [50], [51], [52]. The conserved methyltransefrase RlmN methylates 23S rRNA at position A2503 and a S. aureus strain with a knock-out of the gene encoding RlmN demonstrated a 2-fold increase in linezolid susceptibility [36]. Additionally, an acquired mechanism of linezolid resistance due to acquisition of cfr has recently been described [53], [54]. The product of cfr hypermethylates 23S rRNA at position A2503 leading to the presence of not one, but two methyl groups which affects drug binding [50]. The impact of the codon insertion in SACOL1230 in our strain was confirmed by an allelic exchange experiment where the identical insertion was created in the linezolid susceptible parent strain JKD6210. Although the change in linezolid MIC was not large, this is consistent with previous reports of changes in linezolid resistance in S. aureus due to acquisition of the methyltransferase cfr, where prolonged incubation was required to detect an increase in MIC using Etest [53]. We therefore propose that the CAA insertion in SACOL1230 enhanced ribosomal methylation in the clinical isolate JKD6229 leading to a reduction in linezolid susceptibility. The clinical impact of subtle changes in linezolid susceptibility of S. aureus have not been defined. However, similar to recent findings with reduced vancomycin susceptibility in this organism [5], subtle reductions in susceptibility to an antibiotic may significantly impact the outcome of therapy, especially in patients with deep-seated infection as occurred in this case.
Two additional mutations were detected in JKD6229, as well as the loss of a ∼15 kb plasmid. The mutation in rpoB was clearly linked to the acquired rifampin resistance in JKD6229, and has been previously described [34]. Likewise, the codon insertion in parC contributed to an increase in quinolone MIC of the organism [35]. The plasmid which was present in JKD6210, but absent in JKD6229 (pJKD6210), shared high sequence homology to pUSA300-HOU-MS and encodes beta-lactam resistance, but did not contain the genes encoding cadmium resistance which are present on pUSA300-HOU-MS [33].
Over recent years there has been significant interest in the role of small colony variants of S. aureus in persisting and relapsing infections, and intracellular invasion and persistence is a frequently described feature of these strains [7]. While an understanding of the genetic determinants of SCV S. aureus has focussed on mutations in genes encoding hemin, menadione or thymidine biosynthesis [55], [56], [57], our data clearly demonstrates the heterogeneic nature of this phenotype, with permanent activation of the bacterial stringent response also leading to a growth defect in S. aureus. Not surprisingly, the global transcriptional profile and phenotypic features of stringent response S. aureus demonstrate significant differences to those of the defined hemB and menD mutants, while the transcriptional profile of the stringent response SCV JKD6229 shared significant similarity to the profile of S. aureus after exposure to mupirocin [22]. For example, auxotrophs for hemin, menadione or thymidine have been shown to have reduced tricarboxylic acid cycle (TCA) metabolism leading to reduced electron transport [8], [58], [59]. The SCV strain JKD6229 was not an auxotroph for hemin or menadione and the genes for the TCA cycle were up-regulated in JKD6229 compared to JKD6210. The SCV strain JKD6229 demonstrated increased intracellular invasion, however there was no increase in persistence compared to the parental strain (Fig. 5). Increased fibronectin-binding protein gene expression was found in JKD6229, possibly contributing to increased cellular invasion, as has been previously described for the hemB mutant [11]. However, the absence of increased persistence is interesting. It has previously been demonstrated that reduced agr expression and alpha-toxin expression occurs in clinical SCV S. aureus, and in the hemB and medD mutants [31], and it has been suggested that these changes favour intracellular persistence by avoiding lysis of the invaded cells [7], [43]. In the SCV strain JKD6229, increased expression of the agr locus was demonstrated; an unusual finding for an SCV strain, but this is also associated with the mupirocin induced stringent response in S. aureus [22], and could potentially explain the failure to demonstrate increased intracellular persistence. An interesting finding of this study was the profound increase in expression of capsule biosynthesis genes, associated with a significant increase in capsule production in JKD6229, demonstrated by capsule immunoblot (Fig. 2B). A previous microarray transcriptional comparison of the hemB mutant to its parental strain also demonstrated an increase in capsule biosynthesis genes, however not to the same degree found in JKD6229 [8]. Given the association of staphylococcal capsule production with innate immunity evasion mechanisms and virulence in animal models [60], this phenotypic change in the SCV strain JKD6229 also likely contributed to the clinical behaviour of the organism.
The growth defect of SCV JKD6229 was incompletely replicated in the relA mutant strain JKD6301, suggesting that additional factors contributed to the growth defect of the clinical strain. Although it appears unlikely that the other mutations detected in JKD6229 would lead to an additional growth defect, step-wise generation of each mutation in JKD6301 would be required to confirm this. Another unanswered question from this study is the mechanism of reduced vancomycin susceptibility in the strain JKD6229, which demonstrated a heterogenous-vancomycin intermediate S. aureus (hVISA) phenotype based on the macromethod Etest result (Table 1) [5]. Although mutations of relA in the Gram positive pathogen E. faecalis have been linked to vancomycin tolerance in that organism, there was no change in vancomycin susceptibility of the relA mutant JKD6301 compared to the parent strain JKD6210 demonstrating that an activated stringent response did not alter vancomycin susceptibility in this strain. Interestingly, the transcriptional profile of JKD6229 which was a hVISA, demonstrated some similarities to the transcriptional profiles of other hVISA strains, including enhanced capsule expression and reduced expression of the gene encoding protein A [30]. Finally, it is unlikely that other genomic differences were missed during our comparative genomics analysis. We performed a de novo assembly of the JKD6210 and JKD6229 sequences which revealed similar genome size (approx 2.7 Mb), and 97.2% coverage of the N315 genome at a depth of ≥20 for both strains. Regions of N315 not covered in the JKD6210 and JKD6229 sequences were identical, indicating that these regions were unique to N315. To reduce false positive SNP detection during comparative genomics analysis we set a threshold of a minimum depth of coverage at a SNP of ≥20, and that the reads covering that position are all uniquely and unambiguously aligned to the reference genome. Although a small possibility exists that SNPs with very low read coverage or SNPs within repeat regions might be missed in out comparative genomics analysis, this is unlikely.
In summary, using comparative and functional genomics to investigate the mechanisms of staphylococcal persistence in a patient with a very difficult-to-treat infection, we have detected a new mechanism of SCV S. aureus, and we have described for the first time the features of an activated stringent response in this organism. Also, a novel mechanism of reduced linezolid susceptibility has been described. Further work to determine the relationship between the stringent response and outcome of staphylococcal infections is required, as well as an exploration of the frequency of mutations in the staphylococcal gene encoding RlmN in patients treated with linezolid. This study highlights the limitations of current antimicrobial treatment strategies in patients with serious S. aureus infections.
This study was performed in accordance with Austin Health Human Research Ethics Committee guidelines. The de-identified clinical details described in this manuscript constitute a medical case report that did not require formal Human Ethics Committee approval or Informed Patient Consent.
Bacterial strains and plasmids used in the study are listed in Table 1. Staphylococcal strains were stored in glycerol broth at −80°C and subcultured twice onto Horse Blood Agar (Oxoid) for 48 h before being used for any experiment. Unless otherwise indicated all S. aureus isolates were grown in BHIB (Oxoid), and E. coli grown in LB broth (Oxoid). When required media was supplemented with the following antibiotics at the indicated concentrations: for E. coli, ampicillin 100 µg/mL; for S. aureus RN4220, chloramphenicol 10 µg/mL; for S. aureus clinical isolates, chloramphenicol 25 µg/mL. For all DNA and RNA extractions, or for experimental inoculum preparations when the SCV strain JKD6229 was used, a subculture onto solid media was performed to confirm that the strain retained the SCV phenotype. For all phenotypic experiments growth conditions were carefully controlled, and all strains were grown to the same OD600 prior to analysis.
Vancomycin MICs were determined by microbroth MIC according to CLSI criteria [61]. The detection of vancomycin hetero-resistance was performed by macromethod Etest for vancomycin and teicoplanin as well as vancomycin population analysis, as previously described [62], [63]. A positive macromethod Etest result for hVISA was defined as vancomycin plus teicoplanin MIC≥8 µg/mL, or teicoplanin MIC≥12 µg/mL [5]. The MICs for daptomycin, gentamicin, linezolid, rifampicin and ciprofloxacin were performed by Etest (AB Biodisk), according to manufacturer's instructions. For linezolid MIC testing a 2 McFarland saline suspension was used, because of previous problems in detecting linezolid resistant strains of S. aureus using standard Etest [53]. Other antibiotic susceptibilities were performed by agar dilution according to CLSI criteria [61]. Pulsed-field gel electrophoresis (PFGE) and multilocus sequence typing (MLST) were also performed as previously described [62], [64].
Analysis of S. aureus growth rate was performed using 50 mL Muller Hinton II broth by inoculating 500 µL of an overnight broth culture. The optical density of the broth was read at 600 nm using a spectrometer. Assessment of colony size on solid media was performed by a blinded operator by measuring the size of 100 single colonies on Horse Blood Agar using callipers after 24 hours incubation. An analysis for hemin and menadione auxotrophism for the SCV strain JKD6229 was performed using chemically defined medium (CDM) [65] as previously described, and assessed after overnight incubation [55].
Microarray transcriptional analysis was performed with TIGR version 6 S. aureus arrays, as previously described [30]. For preparation of total RNA shaking flasks (50 mL BHI broth in 250 mL flasks) were inoculated with 500 µL overnight BHI broth culture and incubated on a 225 rpm shaker at 37°C. Optical density was closely monitored, and one millilitre of sample was collected at exponential growth phase (optical density at 600 nm of 0.5) and 0.5 mL RNA stabilization reagent (RNA later, Qiagen) was added and mixed immediately. The mixture was allowed to stand in room temperature for 10 minutes before total RNA was extracted using the RNeasy micro kit (Qiagen). RNA extractions and hybridisations were performed on four different occasions, and the dye swapped with each biological replicate. The images were combined and quantified using ImaGene™ ver 5.1 (Biodiscovery), and then imported into BASE and analyzed using Bioconductor and Limma [66], [67]. The fold ratio of gene expression for the SCV strain (JKD6229) relative to the parental MRSA (JKD6210) was calculated. Using a modified t-test P-values were calculated and adjusted for multiple testing using false discovery rate (FDR) correction. A≥2-fold change with a P value less than 0.05 was considered significant and included in an analysis of differentially expressed genes. Microarray data has been submitted to GEO with accession number GSE20957.
The capsule typing (CP5 and CP8) by multiplex PCR and quantification by immunoblot was performed as previously described [30]. Briefly, crude CP extracts were prepared using 10 mL of an overnight BHI broth culture adjusted to an OD600 of ∼0.5. Serial two-fold dilutions of CP extracts were loaded onto a nitrocellulose membrane using a dot-blot apparatus. After blocking with 5% skim milk, the membrane was hybridised with CP5-specified rabbit antiserum, hybridised with sheep anti-rabbit IgG peroxidase conjugate (Chemicon, Australia), and the image acquired and analysed using the LAS-3000 Luminescent Image Analysis System (Fujifilm, Tokyo, Japan).
Genome sequences for the parental strain JKD6210 and the clinical SCV strain JKD6229 were obtained from an Illumina Genome Analyzer II using 36 cycle paired-end chemistry. Reads were mapped to the reference strains S. aureus COL (Genbank NC_002951.2) and S. aureus N315 (Genbank NC_002745.2) using SHRiMP. SNP/DIPs were detected using Nesoni 0.14, a software tool for analysing high-throughput DNA sequence data (http://bioinformatics.net.au/software). Nesoni tallied the raw base counts at each mapped position in each of the reference strains, and then compared them using Fisher's Exact Test to find variable nucleotide positions in JKD6229 relative to JKD6210. To exclude the possibility that mutations in JKD6229 may have occurred in regions not present in S. aureus COL or N315, de novo assembly of JKD6210 and JKD6229 was performed using Velvet 0.7.55 [68] and the above read mapping and SNP/DIP detection was performed, using the resulting contigs as reciprocal reference sequences. For SNP detection a depth of coverage of ≥20 was required at the allele. The read data for JKD6210 and JKD6229 have been deposited in the NCBI Sequence Read Archive as part of Study accession number SRP001289.
Standard procedures were used for DNA manipulation, molecular techniques, PCR and sequencing [63], [69]. The loci containing the relA nucleotide substitution and the ‘CAA’ insertion in rlmN (from JKD6229 were amplified (Table S2), cloned with the vector pKOR1 and then generated in the parental strain JKD6210 as previously described [63]. The generation of the allele swap in JKD6210 using pJKD6318 was performed as previously described [63], with some modifications. For the final selection step, 100 µL of a 48 hour BHI broth culture (incubated at 30°C) was inoculated into 10 mL BHI broth with 5% horse blood and 400 µg/mL anhydrotetracycline. The broth was incubated for 24 hours at 37°C on a shaker at 225 rpm. The culture was then diluted to 10−5 and 10 µL of a range of dilutions plated on several HBA and BHI agar plates. After 24 hours incubation at 37°C, single colonies were patched on BHI agar plates with and without chloramphenicol, and screened for the correct allele swap. The correct allele swaps were confirmed, and introduction of unwanted mutations excluded, by PCR amplification and Sanger sequencing of the whole relA and SACOL1230 locus from the mutants strains JKD6301 and JKD6300, respectively.
The presence of ppGpp was detected as previously published [42], with some modifications. Briefly, S. aureus strains (JKD6210, JKD6229 and JKD6301) were grown in 25 mL of BHI broth at 37°C with vigorous shaking. Serine hydroxymate (concentration 0.5 mg/mL) was added to one flask of JKD6210 for 10 minutes to induce the stringent response and provide a positive control for the assay. Cells were harvested by centrifugation at OD600 of 0.5. Following addition of 100% methanol, vigorous vortexing and centrifugation to pellet cellular debris, the supernatant containing ppGpp was collected and concentrated by freeze drying overnight. The dried extracts were then resuspended in 1 mM HEPES buffer, pH 7.4 containing 16% DMSO (v/v) and two-fold serial dilutions were performed in the same buffer. To each dilution, PyDPA to a final concentration of 25 µM was added. Fluorescence was observed using a hand held Wood's UV lamp (365 nm) and a FLUOstar Omega microplate reader (Ex 344 nm/Em 470 nm) (BMG Labtech, Offenburg, Germany).
A HeLa cell line was used to test the invasive and intracellular persistence abilities of the clinical and mutant S. aureus strains. HeLa cells were seeded and grown in DMEM cell culture with 5% fetal bovin serum (FBS) in 24 well plates, and infected by the addition of approx 5×106 CFU of an overnight broth culture. The correct starting inoculum was confirmed by colony counts. After 1 hour incubation at 35°C in an incubator with 5% CO2, the infected cells were washed with pre-warmed PBS 6 times to wash away the unattached bacteria and fresh DMEM with 5% FBS and supplemented with 400 µg/mL gentamicin and 40 µL/mL lysostaphin was added into each well and incubated for a further 72 hours. The infected HeLa cells were sampled before adding antibiotics to assess bacterial attachment/invasion, and at 1 hour post addition of antibiotics (to assess invasion), and at 24 hours and 72 hours after adding antibiotics (to assess intracellular persistence). The cell cultures were lysed by PBS supplemented with 0.05% saponin and plated on BHI agar plates. CFUs on plates were counted after 48 hr incubation at 37°C.
The previously described invertebrate S. aureus infection model Galleria mellonella [44] was used to study the pathogenesis of clinical and mutant strains. G. mellonella in the final instrar larval stage were used in groups of 16, and weighed to confirm no difference in size between groups. A HPLC syringe was used to inject 10 µL of bacterial suspension (approx 0.5–1.0×106 CFU) into each caterpillar via the last left proleg. Bacterial colony counts were performed to confirm consistency of inoculum and caterpillars injected with PBS and caterpillars that were not injected were included as controls. Each experiment was repeated on at least 4 different occasions. To determine the bacterial burden in infected caterpillars 48 hours after inoculation an assessment of the S. aureus CFU per caterpillar was performed on a subset of caterpillars.
Non-parametric tests were used to analyse the results of colony size, bacterial attachment, invasion and persistence assays. Statistical analyses were performed using the two-tailed Mann-Whitney U test, with a p<0.05 set for statistical significance. Growth curves and stringent response activity were analysed using a one way analysis of variance (ANOVA) at each time point, and Kaplan Meier plots of G. mellonella killing results were analysed using the log rank test. All analyses were performed using Prism 4 for Macintosh ver 4.0 (GraphPad Software Inc., CA, USA).
|
10.1371/journal.pbio.1000045 | Analysis of the Chloroplast Protein Kinase Stt7 during State Transitions | State transitions allow for the balancing of the light excitation energy between photosystem I and photosystem II and for optimal photosynthetic activity when photosynthetic organisms are subjected to changing light conditions. This process is regulated by the redox state of the plastoquinone pool through the Stt7/STN7 protein kinase required for phosphorylation of the light-harvesting complex LHCII and for the reversible displacement of the mobile LHCII between the photosystems. We show that Stt7 is associated with photosynthetic complexes including LHCII, photosystem I, and the cytochrome b6f complex. Our data reveal that Stt7 acts in catalytic amounts. We also provide evidence that Stt7 contains a transmembrane region that separates its catalytic kinase domain on the stromal side from its N-terminal end in the thylakoid lumen with two conserved Cys that are critical for its activity and state transitions. On the basis of these data, we propose that the activity of Stt7 is regulated through its transmembrane domain and that a disulfide bond between the two lumen Cys is essential for its activity. The high-light–induced reduction of this bond may occur through a transthylakoid thiol–reducing pathway driven by the ferredoxin-thioredoxin system which is also required for cytochrome b6f assembly and heme biogenesis.
| To grow optimally, photosynthetic organisms need to constantly adjust to changing light conditions. One of these adjustments, called state transitions, allows light energy to be redistributed between the two photosynthetic reaction center complexes in a cell's chloroplasts. These complexes act in concert with other components of the photosynthetic machinery to turn light energy into cellular energy. A key component in the regulation of state transitions is the chloroplast protein Stt7 (also known as STN7), which can modify other proteins by adding a phosphate group. When light levels change, the oxidation level of a pool of another chloroplast component, plastoquinone, changes, which in turn activates Stt7, inducing it to phosphorylate specific proteins of the light-harvesting complex of one reaction center. As a result, a portion of this light-harvesting complex is transferred from one photosynthetic reaction center to the other, thereby optimizing photosynthetic efficiency. Here, we have addressed the configuration of Stt7 within the thylakoid membrane of the chloroplast and the molecular mechanisms underlying its activation. Our data reveal that the level of Stt7 protein changes drastically under specific environmental conditions, that the protein does not need to be present in a one-to-one ratio with its targets for activity, and that it associates directly with a number of components of the photosynthetic machinery. The protein-modifying domain of Stt7 is exposed to the outer side of the thylakoid membrane, whereas the domain critical for regulation of its activity lies on the inner side of the thylakoid membrane. These results shed light on the molecular mechanisms that allow photosynthetic organisms to adjust to fluctuations in light levels.
| Photosynthetic organisms are constantly subjected to changes in light conditions. These organisms have developed different mechanisms to rapidly acclimate to this changing environment. At one extreme, when the absorbed light excitation energy vastly exceeds the assimilation capacity of the photosynthetic apparatus, these organisms need to protect themselves. Excess light energy is dissipated as heat through nonphotochemical quenching, which involves conformational changes in the light-harvesting system of photosystem II [1]. In contrast, under low light, photosynthetic organisms optimize the absorption capacity of their antenna systems. This is especially true when changes in light quality occur that lead to the preferential stimulation of either photosystem II (PSII) or photosystem I (PSI), which are linked through the photosynthetic electron transport chain. Under these conditions, balancing of the light excitation energy between the antenna systems of PSII and PSI occurs through a process called state transitions [2–4]. Upon preferential excitation of PSII, the plastoquinone pool is reduced, a process that favors binding of plastoquinol to the Qo site of the cytochrome b6f complex and leads to the activation of a thylakoid protein kinase required for the phosphorylation of the light-harvesting system of PSII (LHCII) [5,6]. In the green alga Chlamydomonas reinhardtii, the LHCII protein set consists of Type I (Lhcbm3, Lhcbm4, Lhcbm6, Lhcbm8, and Lhcbm9), Type II (Lhcbm5), Type III (Lhcbm2 and Lhcbm7), and Type IV (Lhcbm1) proteins, and of CP26 and CP29 [7]. Because of their nearly identical sequences and size, several of these Lhcbm proteins cannot be distinguished by SDS-PAGE. Most of them fractionate into two major bands called P11/P13 (Type I) and P17 (Type III). CP29, Lhcbm5, P11, P13, and P17 are phosphorylated during a state 1 to state 2 transition [7–9]. Although CP29 and Lhcbm5 are mobile during state transitions, it is not yet clear which among the other LHCII proteins of C. reinhardtii are mobile [10,11]. The phosphorylation of LHCII is followed by a displacement of LHCII from PSII to PSI, thus increasing the size of the PSI antenna at the expense of the PSII antenna and rebalancing the excitation energy between both photosystems. Binding of the mobile LHCII to PSI requires the PsaH subunit [12]. This state corresponds to state 2. The process is reversible as preferential excitation of PSI leads to the dephosphorylation of LHCII by unknown phosphatases and its return to PSII (state 1).
State transitions can be induced, not only by changes in light conditions, but also through changes in cellular metabolism. Thus, in C. reinhardtii, the process can be triggered when the level of ATP is low or when the cells are grown in the dark under anaerobiosis. These conditions lead to the influx of reducing equivalents into the plastoquinone pool and to the activation of the LHCII kinase [13]. Moreover, transition from state 1 to state 2 in C. reinhardtii is associated with a switch from linear to cyclic electron transfer [14,15]. In this alga, the major role of state transitions appears to be ATP homeostasis. In land plants, the LHCII kinase can also be activated in the dark by the addition of sugar compounds [16]. Recent studies further confirm that state transitions are not limited to the balancing of excitation energy between the photosystems but that they also play a major role in the adjustment of the light reactions with carbon metabolism [17].
In C. reinhardtii, transition from state 1 to state 2 causes the displacement of 80% of LHCII from PSII to PSI, as deduced from the measurements of the quantum yield of PSI and PSII charge separation [18]. In contrast in land plants, only 15% of LHCII is mobile during state transitions, although this process is associated with considerable structural rearrangements of the thylakoid membranes [19]. The large size of the mobile LHCII antenna leads to significant changes in fluorescence yield during state transitions in C. reinhardtii, a feature that has been exploited for the screening of mutants deficient in this process [9,20]. Such a screen has revealed the existence of the thylakoid Ser-Thr protein kinase Stt7 (AA063768) [21]. Mutants deficient in this kinase are deficient in LHCII phosphorylation and fail to undergo a transition from state 1 to state 2. The Stt7 protein kinase is associated with the thylakoid membrane and contains a potential transmembrane domain upstream of the catalytic kinase domain. In Arabidopsis thaliana, the ortholog STN7 (NP_564946) is also specifically involved in LHCII phosphorylation and state transitions [21,22]. At this time, it is not yet clear whether Stt7 and STN7 act directly on LHCII or whether they act further upstream as part of a kinase cascade. The cytochrome b6f complex plays a key role in the activation of the kinase [4]. It is thus very likely that Stt7/STN7 interacts directly with this complex. If LHCII is the substrate of Stt7/STN7, an interaction between the two is expected. Here, we have used coimmunoprecipitations and pull-down experiments to show that interactions of this type do indeed occur. Our data reveal that Stt7 acts in catalytic amounts. We also show that Stt7 contains a transmembrane region with the catalytic domain on the stromal side of the thylakoid membrane and the N-terminal region in the lumen. This domain appears to play a key role in the regulation of the kinase activity.
To understand how the Stt7 kinase functions, we first estimated its abundance, in particular its molar ratio compared with the cytochrome b6f complex under state 2 conditions. An antibody was raised against Stt7 and the amount of Stt7 was estimated using recombinant Stt7 protein for calibration. A similar calibration was performed with the cytochrome b6f complex (see Figure S1). This analysis revealed that the molar ratio between Stt7 and the cytochrome b6f complex is 1:20, clearly indicating that Stt7 is present at substoichiometric levels compared to the photosynthetic complexes.
The Stt7/STN7 protein kinase has been shown to be associated with the thylakoid membrane [21]. However, it is not known whether Stt7/STN7 acts singly or whether it is associated with other proteins in a larger complex whose composition might change during state transitions. To test this possibility, thylakoid membranes from the stt7 mutant complemented with Stt7 containing a haemagglutinin (HA)-tag at its C-terminal end (Stt7-HA) were isolated. Membranes were prepared from cells in state 1 obtained by illumination in low light (6 μmol m−2 s−1) in the presence of DCMU (3-(3,4-dichlorophenyl)-1,1-dimethylurea) for 30 min and in state 2 by incubating the cells under anaerobic conditions for 30 min in the dark. The occurrence of state transitions was verified by measuring the change in maximum fluorescence (Fmax). The thylakoid membranes were solubilized with n-dodecyl-β-maltoside and fractionated by sucrose density gradient centrifugation. Individual fractions of the two gradients from Stt7-HA thylakoid membranes were separated by PAGE and then tested by immunoblot analysis using antibodies directed against HA, Cytf, PsaA, D1, CP26, CP29, and Lhcbm5 (Figure 1A). Under both state 1 and state 2 conditions, the Stt7 protein kinase was associated with a large complex that partly overlaps with the high molecular weight fractions of PSI and the cytochrome b6f complex but not with PSII (Figure 1A). No major changes in the distribution of the thylakoid complexes were observed under the state 1 and state 2 conditions used. A similar distribution of the complexes was found in the stt7 mutant, confirming that the high molecular weight LHC complexes are not only formed under state 2, but also under state 1 conditions (Figure 1B).
Although the level of Stt7-HA was between 25%–50% compared to Stt7 in wild-type cells (Figure S2), state transitions proceeded to the same extent as in the wild type (unpublished data). Moreover, the amount of photosynthetic complexes was the same in Stt7-HA, stt7, and wild-type cells (Figure S2). As expected, immunoblots with an anti–P-Thr antiserum revealed increased phosphorylation of several proteins in state 2, notably the major LHCII proteins P11, P13, and P17 in the wild-type strain (Figure 1A). Moreover, a weak phosphorylation signal corresponding to Lhcbm5 was detected under state 2 conditions in the same high molecular weight fractions containing PSI. Although an increase of phosphorylation was also observed for the PSII core proteins CP43 and D2 under state 2 conditions, in other experiments, no significant increase of phosphorylation of these proteins was detected between state 1 and state 2.
The immunoblots in Figure 1A indicate that the levels of Stt7 are significantly higher in state 2 than in state 1. To examine this further, cells containing Stt7-HA were grown for 2 h under state 2 conditions, and growth was continued for 4 h either under state 2 or state 1 conditions (Figure 2). At different time points, aliquots of cells were processed for immunoblot analysis with HA and Cytf antibodies. Whereas the level of Stt7 remained the same under continuous state 2 conditions, its level decreased gradually 1 h after the shift to state 1 conditions. It decreased 4-fold after 2 h and more than 20-fold after 4 h (Figure 2C). After shifting the cells to state 2 conditions, the level of Stt7 increased 2-fold after 2 h but did not reach its initial value (unpublished data). Addition of cycloheximide did not affect the decline of Stt7, indicating that no newly synthesized protease is involved in this process (Figure 2D). However, addition of a protease inhibitor mixture to the cells abolished the degradation of Stt7 under state 1 conditions (Figure 2E).
The degradation of Stt7-HA under prolonged state 2 conditions monitored by immunoblotting with the HA antiserum could be due to the removal of the HA tag from Stt7. To test this possibility, we repeated this experiment with the Stt7-HA strain by using both Stt7 and HA antiserum. In both cases, a decline of the Stt7 kinase was confirmed under state 2 conditions (Figure S3A). We checked that this decrease also occurs with untagged Stt7 (Figure S3B). To identify the type of proteases involved, different protease inhibitors were tested under the same conditions as above: ACA (ε-aminocaproic acid) (Sigma) (50 mM), AEBSF (4-(2-aminoethyl)-benzenesulfonyl fluoride hydrochloride) (Roche) (5 mM), NEM (N-ethylmaleimide) (Sigma) (10 mM), phenylmethanesulfonyl fluoride (Sigma) (5mM), and EDTA (50 mM). Leupeptin and NEM were also used at different concentrations (see Figure S4). Samples were taken and analyzed at different time points. Whereas NEM completely prevented the breakdown of Stt7, the serine protease inhibitors ACA and AEBSF had no effect, and EDTA enhanced this process (Figure S4A and S4B). Other inhibitors of cysteine proteases besides NEM, such as E64 and leupeptin, prevented Stt7 degradation under prolonged state 1 conditions (Figure S4C, S4D, and S4E). Although no convincing proof of chloroplast Cys proteases has been reported, their existence cannot be excluded. A bioinformatic search for Cys proteases in plastids of A. thaliana did not reveal any convincing candidate (see Text S1).
As high-light treatment is known to lead to the inactivation of the LHCII protein kinase [23,24], we tested whether Stt7 was stable under these conditions. Cells adapted to state 2 were subjected to high-light treatment (900 μmol m−2 s−1), and the Stt7 levels were determined by immunoblotting at different times. Under these conditions, a steady decrease of Stt7 was observed (Figure 2F), which could be fully prevented by addition of leupeptin (Figure 2G). Under the same light regime, the level of PSII and of PSI were nearly unaffected (Figure 2F). However, measurements of the ratio of variable (Fv) over maximal florescence (Fv/Fmax) revealed that this ratio decreased from 0.7 to 0.2, indicating photodamage to PSII without apparent decrease of D1 protein level.
We took advantage of the decrease in Stt7 under prolonged state 1 conditions to test whether the cells are still able to switch from state 1 to state 2 under these conditions. Cells were first grown under state 1 conditions as shown in Figure 2. After different time periods, cells were collected and assayed for state transitions using the uncoupler FCCP (carbonyl cyanide p-fluoromethoxyphenylhydrazone), a known inducer of transition to state 2 [13]. The fluorescence emission spectra at low temperature and the level of Stt7 were measured under state 1 and state 2 conditions. Figure 2H shows that transition to state 2 occurred readily in these cells collected after 4 h under state 1 conditions, although the amount of Stt7 protein kinase was decreased 20-fold (Figure 2I). This indicates that the Stt7 kinase acts in catalytic amounts.
Activation of the LHCII kinase depends critically on the cytochrome b6f complex [4]. Moreover, this kinase is likely to interact with its putative substrate LHCII. To test whether the Stt7 kinase interacts with the cytochrome b6f complex and/or LHCII, thylakoid membranes from the Stt7-HA strain in state 1 or state 2 were solubilized with dodecyl maltoside and immunoprecipitated with HA antiserum. The immunoprecipitates were fractionated by SDS-PAGE and immunoblotted with antibodies against several known thylakoid proteins. Figure 3A shows that the LHCII proteins P13, P11, P17, CP29, CP26, and Lhcbm5 were coimmunoprecipitated with the Stt7 kinase. The signal obtained with CP29 was only detectable under state 2 conditions. Signals were also observed with Cytf and Rieske protein as well as with the PSI subunit PsaA. In contrast, no interaction was observed between Stt7 and D1 from PSII and with subunit α of ATP synthase (Figure 3A), indicating that the interactions detected between Stt7-HA and the other photosynthetic proteins are specific. Furthermore, no signal was observed with the untagged wild-type strain (Figure 3A). Reciprocal immunoprecipitations with PsaA, Cytf, and P11 antibodies confirmed the interaction of these proteins with Stt7 (Figure 3B).
To further test the interaction of Stt7 with the cytochrome b6f complex, the Stt7-HA strain was transformed with petA containing a His tag at its 3′-end [25]. After testing the homoplasmic state of the transformed strain for petA-His, thylakoid membranes were isolated, solubilized, and the cytochrome b6f complex was purified on a Ni-NTA column. After washing the column, immunoblotting of the eluted fraction revealed that a small portion of Stt7 kinase was associated with the tagged cytochrome b6f complex, but not with the untagged strain (Figure 4A). To determine which of the subunits of the cytochrome b6f complex interacts with Stt7, a pull-down assay with GST-Stt7 and solubilized purified cytochrome b6f complex was performed. The results (Figure 4B) show that the Rieske protein, but not Cytf, could be eluted from the GST-Stt7 column, indicating that Stt7 interacts with the Rieske protein.
Although a putative transmembrane domain within Stt7 is predicted by several algorithms [21], the presence of four Pro residues within this domain raises some questions, and two models need to be considered. In the first, the N-terminal end of Stt7 would be separated from the large catalytic domain of Stt7 by a transmembrane domain. In the second model, Stt7 could be localized entirely on one side of the thylakoid membrane. To distinguish between these possibilities, the Stt7 protein was tagged with FLAG and HA at its N-and C-terminal ends, respectively. It should be noted that FLAG-Stt7-HA is not functional but that it stably accumulates in the thylakoid membranes (see below). Because the presumed substrates of the Stt7 kinase are localized on the stromal side of the thylakoid membrane, it is expected that the kinase domain of Stt7 is also located on the stromal side. This was tested by isolating intact thylakoid membranes from the strains containing Stt7-HA or FLAG-Stt7-HA and by subjecting the thylakoid membranes to mild digestion with protease V8. The resulting protein extracts were then examined by PAGE and immunoblotting using antibodies directed against HA, FLAG, PsaD, and OEE2. PsaD is known to be partially exposed to the stromal side, whereas OEE2 is entirely located on the lumenal side of the thylakoid membrane. OEE2 and the main body of PsaD are thus expected to be protected from any external protease. Under conditions (75 μg/ml V8) in which proteolysis mildly affected the OEE2 protein and led to partial digestion of PsaD, the level of Stt7 was significantly decreased as measured with HA antibodies, confirming that the kinase domain is located on the stromal side of the membrane (Figure 5A). Similar results were obtained with Stt7-HA (unpublished data). In contrast when antibodies against FLAG were used, products of smaller size were detected, indicating that the N-terminal end of Stt7 is localized on the lumenal side of the thylakoid membrane (Figure 5A). Sonication of the thylakoid membranes followed by V8 protease treatment revealed that the levels of protected fragments detected with the FLAG antibodies were significantly reduced as also observed with the lumenal protein OEE2 (Figure 5A). Sonication also led to enhanced degradation of the HA-tag of Stt7, presumably because the domain of the kinase exposed to the stroma was more accessible to the protease under these conditions and/or the thylakoid membrane was damaged. In contrast, no difference was observed for PsaD with and without sonication.
The orientation of Stt7 was further tested with the yeast split-ubiquitin system [26]. The C-terminal fragment of ubiquitin (Cub) was fused to either the N- or C-terminal end of Stt7 and expressed in yeast together with the N-terminal end of ubiquitin (Nub) fused to the C-terminal end of the endoplasmic reticulum (ER) protein Alg5, which places Nub on the cytoplasmic side of the membrane. Ubiquitin was reconstituted only with the construct in which Cub was fused to the C-terminal end of Stt7 but not when it was fused to the N-terminal end (Figure 5B). Because membrane proteins usually insert with their lumen domain in the periplasmic space of yeast, the two-hybrid results are fully compatible with the topology of Stt7 derived from the protease protection studies.
The transmembrane domain of Stt7 near its N-terminal end is preceded by a region that contains two conserved Cys separated by four residues that are also conserved in the orthologous STN7 kinase of Arabidopsis [21]. In land plants, it has been shown that under high light, the LHCII kinase is inactivated through the ferredoxin-thioredoxin system [23,24]. One possibility is that these conserved Cys are targets of this redox system. To test the role of these residues, the two Cys were changed individually to Ser or Ala by transforming the stt7 mutant with the Stt7-HA-C68S/A and Stt7-HA-C73S/A constructs. In all four cases, the mutant kinase accumulated as in Stt7-HA cells (Figure 6A). However, the low-temperature fluorescence emission spectra measured under conditions inducing state 1 or state 2 were nearly identical in these mutants, indicating that they are deficient in state transitions (Figure 6A). As expected, an increase of PSI fluorescence at 715 nm, which is characteristic for state 2, was detected in the rescued Stt7-HA strain. Moreover, the Cys mutants failed to phosphorylate LHCII under state 2 conditions (Figure 6B). Thus, the Cys residues are critical for kinase activity, although they are separated from the catalytic kinase domain by the transmembrane region.
The insertion of a FLAG-tag near the N-terminal end of mature Stt7 abolished state transitions and kinase activity but did not affect the stable accumulation of the protein (Figure 6A and 6B). Moreover, the presence of the FLAG tag specifically prevented the coimmunoprecipitation of Stt7 with the Rieske protein, but not with Cytf (Figure 6C). At first view, this appears to contradict the results of the pull-down experiment, which indicate that the Rieske protein, but not Cytf, interacts with Stt7. In the case of the pull-down experiment, recombinant Stt7 protein was used in which the transmembrane may not be correctly folded. This domain could be responsible for the binding to Cytf. In the case of the immunoprecipitation, solubilized thylakoid membranes were used in which the interaction between Stt7 and Cytf is preserved. The addition of the FLAG epitope appears to prevent proper interaction of Stt7 with the Rieske protein. In contrast, coimmunoprecipitations of Stt7-C68S occurred both with Cytf and the Rieske protein (Figure 6C). Taken together, these results indicate that the N-terminal end of Stt7 plays a crucial role in the activation of its kinase activity and that this domain may be involved in the interaction with the Rieske protein.
State transitions lead to a considerable reorganization of the antenna systems of PSII and PSI in C. reinhardtii. Moreover, they are accompanied by large changes in the PSI-LHCI supercomplex [11,27]. Fractionation of solubilized thylakoid membranes by sucrose density gradient centrifugation revealed no major changes in the distribution of Stt7 and the LHCII proteins between PSII and PSI under state 1 and state 2 conditions. Interestingly, Stt7 is associated with a large molecular weight complex which cofractionates with the high molecular weight fractions of the cytochrome b6f complex and PSI, but clearly not with PSII (Figure 1). These partial cofractionations of Stt7 with the cytochrome b6f complex and PSI are compatible with the coimmunoprecipitation experiments (Figure 3). The exact composition of these complexes remains to be determined.
Our results on the fractionation of the complexes differ slightly from those of Takahashi et al. [11]. Whereas these authors found a clear shift of CP26, CP29, and Lhcbm5 towards the PSI region upon a transition from state 1 to state 2, which they attribute to the formation of a large PSI-LHCII supercomplex, we only detected a small increase in the ratio between high and low molecular weight fractions of the sucrose gradient in state 2 for CP26 and Lhcbm5 (Figure 1A). In the case of CP29, there was no significant difference in its distribution in state 1 and state 2. These differences may be due to the fact that state 2 and state 1 were induced through different means in our study and that of Takahashi et al. They used FCCP plus NaF and DCMU plus staurosporine for inducing state 2 and state 1, respectively, whereas we used anaerobiosis and vigorous aeration in the presence of DCMU. We further confirmed the presence of the LHCII proteins detected in the high molecular weight fractions under state 1 conditions in the stt7 mutant.
Although we could not detect major changes in the distribution of the LHC complexes in the sucrose density gradient under state 1 and state 2 conditions, there were changes in the phosphorylation patterns. A phosphorylated form of Lhcbm5 was detectable in the high molecular weight fractions under state 2, but not state 1 conditions (Figure 1A). Takahashi et al. [11] observed phosphorylated forms of CP29 and Lhcbm5 in this fraction. A phosphorylated form of CP29 in the high molecular weight fraction was also reported by Kargul et al. [10]. The differences between these studies might be partly accounted for by the different specificities of the anti–P-Thr antibodies used.
A striking feature is that the level of Stt7 is significantly lower in state 1 than in state 2. This was further investigated by a state 2–state 1 time course experiment (Figure 2). Within 2 h after shifting from state 2 to state 1 conditions, the level of Stt7 decreased to one fourth of state 2 levels. The level of Stt7 decreased further to 2%–5% of state 2 levels after 4 h in state 1, indicating that Stt7 is unstable under prolonged state 1 conditions. Thus, Stt7 is more abundant and stable under conditions where it is active. The decrease of Stt7 under state 1 conditions could be prevented by addition of inhibitors of Cys proteases to intact cells. It is therefore possible that under state 2 conditions, Stt7 is more protected from proteases either because of a posttranslational modification, e.g., phosphorylation, or of its association with other proteins. Although experimental evidence for the presence of cysteine proteases in chloroplasts is weak, their existence and role in Stt7 turnover cannot be excluded. It is possible that other proteases are involved in this process, such as the FtsH and Deg proteases, which are known to degrade thylakoid membrane proteins [28]. Some Deg proteases are indeed sensitive to NEM [29].
We note that the decrease in Stt7 occurs only after a prolonged period in state 1, whereas state transition is a short-term acclimation response that occurs within minutes. It is therefore possible that the control of the Stt7 level is part of a long-term acclimation response. There is indeed evidence for a role of the ortholog STN7 of A. thaliana in such a process [30].
The level of Stt7 decreases under high light. Under the same conditions, PSII levels remained constant. In land plants, the LHCII kinase is known to be inactivated through the ferredoxin-thioredoxin system under high light [31]. It is thus conceivable that the kinase is less stable when it is maintained for a prolonged period in its inactive state, induced either by state 1 conditions or high light. Although the redox state of the plastoquinone pool is critical for activation of the kinase, it does not solely determine the level of Stt7. Under state 1 conditions, the plastoquinone pool is oxidized, whereas it is expected to be reduced under high light.
A close interaction between Stt7 and the cytochrome b6f complex is apparent from the coimmunoprecipitation results. This complex is known to play a critical role for the activation of the kinase during a state 1 to state 2 transition [32]. Mutants deficient in cytochrome b6f complex fail to phosphorylate LHCII and are blocked in state 1 [32].
The original state transition model postulates that upon activation of the Stt7 kinase through the cytochrome b6f complex, the kinase is released from the complex to phosphorylate LHCII. However, we find that the association of the Stt7 kinase with the cytochrome b6f complex does not markedly change between state 1 and state 2. One possibility is that another downstream kinase is phosphorylated by Stt7, which in turn phosphorylates LHCII. Alternatively, the Stt7 kinase may be part of a large supercomplex that includes the cytochrome b6f complex and the PSII-LHCII complex. We have not been able to detect complexes of this kind, although it is possible that they are only formed transiently. The PetO subunit of the C. reinhardtii b6f complex is known to be phosphorylated during state transitions [33]. Despite several attempts, we were unable to detect any interaction between Stt7 and PetO based on coimmunoprecipitations or yeast two-hybrid screens. It remains to be seen whether the phosphorylation of PetO depends on Stt7.
To identify which subunit of the cytochrome b6f complex interacts with Stt7, pull-down experiments were performed. The Rieske protein was identified as an interactant (Figure 4B). Based on structural studies of the mitochondrial bc1 and of the chloroplast b6f complexes, electron transfer between plastoquinol at the Qo site and cytochrome f (Cytf) is mediated by the Rieske protein which moves from a proximal position when the Qo site is occupied by plastoquinol to a distal position when the Qo site is unoccupied [25,34,35]. It has been suggested that this dynamic behavior of the Rieske protein could be coupled to the activation of the Stt7 kinase [36–38]. Such a dynamic model is compatible with the low abundance of the Stt7 kinase with a molar ratio of 1:20 relative to the cytochrome b6f complex found in this study. Assuming one cytochrome b6f complex for one PSII core complex and an average of ten LHCII proteins per PSII reaction center [39], the molar ratio of Stt7 kinase to LHCII protein can be estimated at 1:200. At first sight, the kinase would have to phosphorylate several LHCII substrates and may undergo multiple rounds of activation. It is possible that the Stt7 kinase acts first on the LHCII located in the edges of the grana and that this phosphorylation induces extensive remodeling of the thylakoid membrane as reported recently during state transitions in A. thaliana [19]. These changes may facilitate the access of Stt7 to LHCII in the grana core. Alternatively, the observed strong LHCII phosphorylation under state 2 conditions could be due to very flexible movements of PSII-LHCII supercomplexes in the grana core and grana margins.
An intriguing component of the cytochrome b6f complex is its single chlorophyll a molecule whose chlorine ring lies between helices F and G of the PetD subunit, whereas the phytyl chain protrudes near the Qo site [25]. Interestingly, mutants affected in the binding site of chlorophyll a, besides having reduced cytochrome b6f turnover, also display a decreased rate of transition from state 1 to state 2 [36]. This region may thus either be an interaction site for the N-terminal region of Stt7 or it could act as a sensor for the presence of plastoquinol at the Qo site and initiate a signaling pathway through the chlorophyll a molecule towards the catalytic domain of Stt7 on the stromal side of the thylakoid membrane.
The coimmunoprecipitation experiments indicate that Stt7 is associated with LHCII in most cases under both state 1 and state 2 conditions. LHCII could be the direct substrate of Stt7 or, alternatively, Stt7 could be part of a multikinase complex, which ultimately phosphorylates LHCII. The only marked difference in coimmunoprecipitation between state 1 and state 2 was observed for CP29 (Figure 3). CP29 is particularly interesting. First, this monomeric LHCII together with CP26 and Lhcbm5 has been proposed to act as linker between the LHCII trimers and the dimeric PSII reaction center for the transfer of excitation energy [11,39–41]. Second, during transition from state 1 to state 2, CP29 undergoes hyperphosphorylation: in addition to Thr6 and Thr32, which are phosphorylated in state 1, Thr16 and Ser102 are phosphorylated in state 2 [10,42]. Third, electron microscopy (EM) analysis of PSII-LHCII complexes lacking CP29 could not distinguish between C2S2 and PSII monomeric complexes [39]. Fourth, in maize bundle sheath cells, which carry out mostly cyclic electron flow, the few remaining PSII complexes are monomeric [43]. Taken together, these results raise the possibility that the phosphorylation of CP29 in state 2 may act as a switch for cyclic electron flow with the detachment of CP29 from PSII and its monomerization.
The coimmunoprecipitation experiments also reveal an interaction of Stt7 with the PSI complex. This finding is surprising, as one would expect that the kinase acts on LHCII bound to PSII and that it would not interact with PSI. However, movement of PSI-LHCI complexes from the stromal lamellae to the grana margins following LHCII phosphorylation occurs in land plants, and it was proposed that the PSI absorption cross section is increased in this region through interaction of PSI-LHCI with the phosphorylated LHCII originating from the grana [44]. It is possible that the grana margins constitute a platform where the observed interactions of Stt7 with PSI could occur. It is not known whether the kinase is active or inactive when it is bound to PSI, and the role of this association remains to be determined.
Analysis of the Stt7 amino acid sequence by bioinformatic means predicts the presence of a single transmembrane domain. However, this putative transmembrane domain contains four Pro residues that may prevent the formation of an α helix. It was therefore important to test the topology of Stt7 by experimental means using Stt7 tagged at its N-terminal end with FLAG and at its C-terminal end with HA. Using intact thylakoid membranes, we showed that whereas the C-terminal end of Stt7 is susceptible to protease digestion, the N-terminal end is protected, indicating that Stt7 indeed contains a transmembrane domain with the kinase domain on the stromal side and the N-terminal end in the lumen. The N-terminal region contains Cys68 and Cys73, which are conserved in the ortholog of Stt7 in land plants [21]. Among the seven Cys residues in Stt7, these are the only conserved Cys residues between Stt7 and STN7. In land plants, the LHCII kinase is inactivated by high-light treatment through the ferredoxin-thioredoxin system [31]. The two conserved Cys could therefore be the targets of this redox system and/or play a major role in the activation of the kinase. By changing either of the two Cys to Ala or Ser, the kinase was inactivated. There was no phosphorylation of LHCII under state 2 conditions, and the mutants were blocked in state 1. Possibly a disulfide bridge between these two Cys is required for kinase activity, or redox changes of this disulfide bridge are critical for its activity (Figure 7). The question arises how the redox state of these two Cys in the lumen is regulated through the redox state of the stromal compartment. Recently, at least two components of a transthylakoid thiol–reducing pathway have been identified in chloroplasts. The CcdA thiol disulfide transporter is a polytopic thylakoid protein with two highly conserved Cys in membrane domains, which is able to convey reducing power from the stroma to the lumen [45]. The second component is the Hcf164 thioredoxin-like protein, which acts as a thiol disulfide oxidoreductase [46,47]. This system, which is also conserved in bacteria, is thought to be involved in cytochrome c and cytochrome b6f assembly but could also have additional roles. In this respect, given the tight physical association of Stt7 with the cytochrome b6f complex and the requirement of this active complex for the activation of the kinase, it is tempting to propose that the redox state of the Cys68 and Cys73 couple is controlled through the same thioreduction system that operates in cytochrome b6f assembly. These changes in redox state of Stt7 could in turn induce conformational changes of the kinase and affect both its activity and stability.
Chlamydomonas reinhardtii wild-type and mutant cells were grown as described [48]. The stt7 mutant and stt7 complemented with Stt7-HA were used [21]. The HA-tag consists of six copies of the HA peptide YPYDVPDYA inserted at the C-terminal end of Stt7. In some experiments, the double-tagged FLAG-Stt7-HA was used with FLAG inserted after the Stt7 transit peptide and the HA-tag at the C-terminal end of Stt7. Strains were maintained on Tris-acetate-phosphate (TAP) medium at 25 °C in dim light (10 μmol m−2 s−1). The stt7 mutant strain complemented with Stt7-HA was also transformed with the ph6FA1 plasmid [25] in order to obtain a strain expressing Stt7-HA and cytochrome f tagged with a His-tag. Homoplasmicity of this strain was checked by PCR.
Stt7-HA consists of six copies of the HA epitope YPYDVPDYA inserted at the C-terminal end of Stt7. The molecular mass of this HA tag is 8.4 kDa and that of Stt7-HA is 85 kDa. Compared with Stt7-HA, FLAG-Stt7-HA contains in addition the FLAG tag RDYKDHDGDYKDHDIDYKDDDDKS with a molecular mass of 3 kDa inserted after the 41 amino acid transit peptide of Stt7.
Cultures were grown in TAP medium to a density of 2 × 106 cells/ml. Cells were subsequently concentrated 10-fold in HSM medium. State 1 was induced by incubating cells in 10−5 M DCMU (3-(3,4-dichlorophenyl)-1,1-dimethylurea) in dim light (10 μmol m−2 s−1) under strong aeration, and state 2 was obtained by incubating cells under anaerobic conditions in the dark. Cells were harvested at 6,000g for 10 min and resuspended in buffer at a density of 108 cells/ml and broken in a French press at 1,200 psi. Thylakoid membranes (0.8 mg/ml) were prepared as described [11] and solubilized with 0.9% n-dodecyl-β-maltoside for 30 min in ice, then 0.5 ml were layered on a sucrose density gradient (0.1–1.3 M sucrose in 5 mM Tricine-NaOH [pH 8.0], 0.05% n-dodecyl-β-maltoside) and centrifuged at 280,000g for 16 h in a SW40 Beckman rotor. After centrifugation, the gradient was divided into 18 fractions that were analyzed by SDS/PAGE and immunoblotting. The antibodies used were against HA, FLAG, Cytf, PsaA, D1, CP26, CP29, Lhcbm5, and P-Thr.
Cultures were grown in TAP medium to a density of 2 × 106 cells/ml. After a 10-fold concentration, cells were incubated in HSM medium under anaerobic conditions in the dark for 2 h. Cells were then maintained under state 2 conditions or cultured under dim light under strong aeration (state 1 conditions). In some cases, cycloheximide (10 μg/ml) or protease inhibitor cocktail (Roche) (2× concentration recommended by the manufacturer) were added prior to the onset of state 1 conditions.
Chlorophyll fluorescence emission spectra were recorded with a Jasco FP-750 spectrofluorimeter using intact cells at a concentration of 106 cells/ml frozen in liquid nitrogen. The excitation light had a wavelength of 435 nm, and emission was detected from 650 to 800 nm. Fv and Fmax measurements at room temperature were performed with a Hansatech PAM fluorimeter.
Proteins from thylakoid membranes isolated from cells in either state 1 or state 2 were separated on a 15% SDS-polyacrylamide 6 M urea gel and transferred to nitrocellulose membranes. The membranes were blocked with bovine serum albumin and incubated with rabbit anti-phosphothreonine antibody (Cell Signaling Technology).
Thylakoid membranes (0.8 mg/ml) were solubilized with n-dodecyl-β-maltoside for 30 min in ice, and nonsolubilized material was removed by centrifugation at 12,000g for 10 min at 4 °C. Fifty microliters of anti–HA-affinity matrix was added to 0.3 mg of chlorophyll of solubilized membranes and incubated overnight at 4 °C. The beads were washed five times in TBS-BSA (100 mM Tris/HCl [pH 7.5], 150 mM NaCl, 0.05% BSA), and the bound proteins were eluted in 40 μl of 2× SDS loading buffer (100 mM Tris/HCl [pH 6.8], 4% SDS, 0.2% bromophenol blue, 20% glycerol) for 30 min at room temperature. The immunoprecipitated proteins were analyzed by immunoblotting with antibodies against subunits of the photosynthetic complexes.
The full-length Stt7 cDNA was cloned in the pGEX-4T-1 expression vector (Amersham Pharmacia Biotech). Proteins were expressed in Escherichia coli as GST fusion and purified with the GST fusion system kit (Amersham Pharmacia Biotech). Two micrograms of GST fusion protein were immobilized on glutathione Sepharose 4B beads (incubation for 2 h at 4 °C) and mixed with 50 μg of solubilized thylakoid extract in 0.5 ml of PBS. After overnight incubation at 4 °C on a rotary shaker, beads were washed four times with the same buffer, resuspended in 30 μl of SDS gel-loading buffer (100 mM DTT, 2% SDS), and the eluate was fractionated by SDS-PAGE.
Thylakoids were washed three time in 50 mM Hepes (pH 7.4), 0.3 M sucrose [49] without protease inhibitors and resuspended in NH4HCO3 25 mM. Samples were sonicated in a waterbath sonicator with 1-min sonication and 30-s cooling five times. Endoproteinase GLU-C (Sigma) was added at indicated concentrations to thylakoid membranes at a chlorophyll concentration of 0.3 mg/ml at room temperature for 15 min. The digestion was arrested by TCA precipitation and samples were resuspended in 1× loading buffer containing 8 M Urea. Further analyses of the samples were performed by SDS-PAGE and immunoblotting.
All steps were performed at 4 °C with 1 mM AEBSF (Roche) as protease inhibitor in all buffers. After solubilization with 0.9% n-dodecyl-β-maltoside, 640 μg of thylakoid membranes (chlorophyll equivalent) were incubated 2 h with Ni-NTA matrix (Qiagen) in the presence of 200 mM NaCl and 10 mM imidazole. The matrix was washed five times in washing buffer (TBS, 200 mM NaCl, 20 mM imidazole), and the bound proteins were eluted in 2× 1 ml of elution buffer (TBS, 250 mM NaCl, 300 mM imidazole). The eluted proteins were analyzed by immunoblotting with antibodies against subunits of the photosynthetic complexes.
The split-ubiquitin experiments were performed using the DUALmembrane kit 3 from Dualsystems Biotech (http://www.dualsystems.com). Cub fused to the artificial transcription factor LexA-VP16 was fused to either the N-terminal end of Stt7 (minus the chloroplast signal peptide) or to its C-terminal end (including STE leader sequence) according to the manufacturer's protocol using SfiI restriction sites. These constructs were then tested for their ability to release LexA-VP16 when coexpressed with an integral membrane protein (Alg5) fused at its C-terminal to the N-terminal half of ubiquitin (Nub). If both Cub and Nub are located in the cytoplasm, spontaneous reassociation will occur, and ubiquitin-specific proteases will be recruited, releasing LexA-VP16, which will in turn activate the reporter genes (Ade and His).
|
10.1371/journal.pcbi.1000174 | Identifying Cognate Binding Pairs among a Large Set of Paralogs: The Case of PE/PPE Proteins of Mycobacterium tuberculosis | We consider the problem of how to detect cognate pairs of proteins that bind when each belongs to a large family of paralogs. To illustrate the problem, we have undertaken a genomewide analysis of interactions of members of the PE and PPE protein families of Mycobacterium tuberculosis. Our computational method uses structural information, operon organization, and protein coevolution to infer the interaction of PE and PPE proteins. Some 289 PE/PPE complexes were predicted out of a possible 5,590 PE/PPE pairs genomewide. Thirty-five of these predicted complexes were also found to have correlated mRNA expression, providing additional evidence for these interactions. We show that our method is applicable to other protein families, by analyzing interactions of the Esx family of proteins. Our resulting set of predictions is a starting point for genomewide experimental interaction screens of the PE and PPE families, and our method may be generally useful for detecting interactions of proteins within families having many paralogs.
| We consider the problem of detecting protein interactions from genome sequences when the potential interacting partners belong to large families of similar (homologous) proteins. Many computational methods for predicting protein interactions rely on similarity to a pair of known interacting proteins. When the proteins in question are members of large groups of similar proteins within the same organism (paralogs), the problem of inferring the correct interactions becomes difficult. To illustrate the problem, we undertook prediction of interactions of some highly expanded protein families of Mycobacterium tuberculosis (Mtb), which are believed to contribute to the bacterium's ability to infect human beings. To generate predictions, we analyzed patterns of coevolution in a small subset of likely interacting proteins, and extended these patterns to predict additional interactions throughout the genome. Our results provide a map for experimental probes of the Mtb interaction network, for the benefit of drug and vaccine discovery. More generally, our procedure is applicable to detecting interactions of proteins that belong to large families of paralogs in any organism with a sequenced genome.
| Tuberculosis remains a health problem of global importance [1]. Despite the availability of the genome sequence of Mycobacterium tuberculosis (Mtb) for nearly a decade [2], the biology of the pathogen, particularly the molecular mechanisms by which it achieves virulence, remains poorly understood. Probing the molecular interaction network of Mtb therefore is an important step in the fight against tuberculosis disease.
The PE and PPE gene families in Mtb make up nearly 10% of the bacterium's coding DNA [2]. The two families combined have about 150 members, amounting to 4% of the open reading frames (ORFs) in Mtb. The PE and PPE gene families account for much of the genomic difference between Mtb and other (nonpathogenic) mycobacterial genomes [3],[4]. Therefore it is thought that they may have a role in Mtb's virulence and host-specificity. A subset of PE proteins is displayed on the bacterium's cell surface [5], can elicit an immune response [6], and may be a source of antigenic diversity for Mtb [7]. PPE proteins have also been found on the cell surface [8],[9], may be secreted [10], and can confer virulence [11]. These studies indicate the likely importance of the PE and PPE gene families in pathogenesis. More extensive characterization of their function, interactions, and roles in infection are therefore important areas for investigation.
Genome analysis suggests that the PE and PPE families are functionally linked [12]–[14]. Pairs of PE and PPE genes are frequently found adjacent on the Mtb genome, and the structure of a complex of one such PE/PPE protein pair was recently characterized [13]. These results indicate that there may be many other instances of interactions between PE and PPE proteins. However, with only one complex characterized so far, it remains unclear which specific members of the two families interact. The 87 PE and 65 PPE proteins (depending on similarity threshold) in the Mtb H37Rv genome generate ∼6,000 possible pairwise combinations. It may be that dozens of biologically relevant PE/PPE complexes remain to be characterized. Because the PE and PPE families can interact with the host immune system [5],[6],[11], combinatorial formation of complexes might enable immune evasion during tuberculosis infection. Mapping the PE/PPE interaction network is therefore of critical importance for accelerating drug discovery. Because PE and PPE proteins are difficult to express and purify experimentally [13], new computational methods are needed to detect likely PE/PPE complexes and efficiently prioritize experiments.
Perhaps the most straightforward bioinformatic approach for detecting PE/PPE complexes is to simply predict interaction of the PE/PPE pairs found in the same operon [15]–[18]. Some 14 pairs of PE and PPE genes, including the one complex that has been structurally characterized to date [13], are found adjacent on the genome, in the same orientation, with minimal intergenic distance, and with the PE 5′ to (upstream of) the PPE (the PE proteins in such pairs do not include any of the repeat-containing PE_PGRS proteins). Because of this recurring genome organization motif, such pairs are likely expressed in the same operon [19]. However, these same-operon PE/PPE pairs comprise less than 10% of the total number of PE and PPE genes in Mtb. The majority of PE and PPE proteins are found unpaired in the genome, and it is possible that some of these interact despite not having genomic proximity. Computationally detecting PE/PPE complexes not found by the operon method is therefore an important challenge.
The tendency for proteins to coevolve with their interaction partners has been described [20]–[23], and bioinformatic methods to detect protein coevolution have therefore been proposed for predicting protein interactions [24]. The idea is to exploit the correlation of the phylogenetic distance matrices of two protein families whose members are known from experiments to interact. Known interacting proteins tend to be found at analogous regions of their respective phylogenetic trees [20],[24],[25] (which can also be represented as distance matrices). Such methods can accurately pair ligands with receptors [24], and could potentially be used to infer interactions between the PE and PPE families. However, a difficulty of applying these methods in our case is that benchmarking predictions requires a set of experimentally determined interactions, and currently only a single known example of a PE/PPE complex exists [13]. Our challenge, therefore, for the computational prediction of PE/PPE interactions is the evaluation of predictions given the currently limited number of known PE/PPE interactions.
We combined the operon method, coevolution analysis, and structural knowledge of interacting domains to develop a coevolution-based strategy to predict PE/PPE complexes in the Mtb H37Rv strain. Some 289 predicted complexes resulted from the application of our method. To validate the predictions, we used several published mRNA expression datasets from Mtb to assess PE/PPE coexpression in vivo. A significant overlap was seen between coevolved and coexpressed PE/PPE gene pairs, supporting the coevolution-based predictions, and resulting in a high-confidence list of possible complexes. To demonstrate the extensibility of our method to other protein families, we performed a similar analysis of interactions of the ESAT-6/CFP-10 (Esx) family of proteins. Our results are a starting point for experimental genomewide screens of PE/PPE and Esx complexes, and our method may be applicable to other functionally linked protein families in Mtb and other microbial pathogens.
We assumed that each interacting pair of PE/PPE proteins must have complementary interfaces, and that the residues in these interfaces may coevolve due to positive selective pressure on the interaction. Although we currently do not have sufficient data from PE/PPE complexes to accurately predict residue-residue interactions from sequence using correlated mutations analysis [26]–[29], we can delineate the likely interacting regions by their similarity to the structurally characterized PE/PPE interacting domains [13].
We assumed that PE/PPE gene pairs adjacent on the genome, and in the same orientation, are in expression operons, as has been shown for Rv2431c/Rv2430c [13]. The components of protein complexes and metabolic pathways in prokaryotes are often located together on the genome in operons [19]. These operons are transcribed as a single, polycistronic mRNA. Genes located on an operon usually function together, and often form protein complexes. We predict thirteen other PE/PPE gene pairs lie in operons (Figure 1A) based on their short intergenic distance (<100 bp) and same transcription direction. These pairs have a high degree of coexpression (average mRNA correlation 0.59 for operon-paired, 0.05 for genomewide PE/PPE gene pairs, see Materials and methods), suggesting that these PE/PPE pairs are indeed in operons.
Finally, we assumed that PE/PPE pairs in operons are likely to interact in a manner similar to the structurally characterized, operon-coded, PE/PPE complex of Rv2431c/Rv2430c [13]. To support our assumption that bacterial operons tend to code protein complexes, we analyzed the tendency for annotated E. coli protein complexes to reside in operons in the EcoCyc database [30]. We extracted 280 complexes, involving 692 proteins, from EcoCyc. We asked what fraction of protein pairs found in complexes also had their genes in the same operon, and found this to be 49% (942 protein pairs in operons out of 1918 protein pairs in complexes). To assess the significance of this result, we shuffled the identity of the genes in complexes by replacing each with a random E. coli gene, and re-assessing the overlap. One thousand shufflings were performed and an overlap of 49% was never achieved; in fact, the highest overlap obtained was 2%. We conclude there is a significant tendency for bacterial protein complexes to be coded in the same operon. While this does not guarantee that proteins coded in operons interact, given a known example of an operon-coded PE/PPE complex, we might expect PE/PPE pairs similarly organized in operons to interact.
Figure 1 illustrates our method for detecting pairs of coevolved PE and PPE genes (and thus, possible interacting proteins). Figure 1A shows all PE and PPE gene pairs that lie in the same orientation of 5′ PE → PPE 3′ with no more than 100 bp separation between the PE and PPE genes. These PE/PPE pairs are likely within the same operon [15]–[18], and are summarized in Table 1. We refer to these as the ‘operon pairs’; they form the training data for our method. PE and PPE protein sequences coded by the operon pairs are aligned to the sequence of the appropriate subunit of the PE/PPE complex of Rv2431c/Rv2430c (Figure 1B). Next, the structure-based multiple alignment is used to generate phylogenetic distance matrices, which contain pairwise protein similarity relationships (Figure 1C). Notice that each equivalent row in the matrix is an operon-paired (and hence assumed interacting) PE/PPE pair. These are called the ‘reference matrices’. For all (operon-paired or otherwise) PE and PPE protein sequences in the Mtb genome, distance vectors to the reference matrices are generated (Figure 1D). The correlation between these vectors, Cij, is a measure of the PE/PPE pair's possible coevolution. Next, Cij scores are further processed (Figure 1E) to yield Sij, the paralog matching score for the predicted complex of PEi and PPEj.
The probable interacting regions of all PE and PPE proteins in the Mtb genome were delineated. This was done using the ClustalW program [31] to perform a multiple sequence alignment of each protein family to a secondary structure profile derived using the DSSP program [32] on the appropriate subunit of the known PE/PPE complex structure. The secondary structural alignment was motivated by the observation that the PE and PPE proteins of known structure are composed of long α-helices interspersed with turns and loops, and our intuition that insertions and deletions would preferentially occur in regions outside the helices. The alignment was visually inspected to remove outlying or poorly-aligned sequences. All remaining PE and PPE sequences in the alignment were truncated to eliminate regions not aligned to the structure. In many cases both the PE and PPE proteins contained additional domains in their C-terminal regions, including the PGRS repeats in PE_PGRS proteins. All subsequent sequence analysis in this work was performed with these truncated sequences. We reasoned that limiting our analysis to homologous interacting domains would facilitate detection of coevolution relevant to protein interaction, and would therefore not be confused by spurious coevolution signals from regions not involved in PE/PPE interface. This additional truncation step was performed because we observed that most PPE, and some PE, domains have additional domains, low complexity regions, or membrane helices C-terminal to the conserved interacting domains. The resulting alignments are provided in the Supporting Information ( Datasets S1 and S2).
Phylogenetic distance matrices for the subset of PE and PPE proteins linked in the 14 operons (Table 1) were constructed using the ClustalW program [31]. For each of the PE and PPE families, the 14 sequences in operon pairs were manually extracted from the full-family alignment. The 14-sequence subalignments were then loaded into ClustalW to generate 14×14 distance matrices. Phylogenetic distance matrices represent the pairwise distance between protein sequences. In ClustalW, pairwise distances between sequences are measured by the fraction of mismatches in ungapped positions of an alignment of two sequences. If our assumption that operon-paired PE/PPE genes code complexes were correct, we reasoned that there would be a correlation of the two distance matrices when the genes of both matrices are respectively ordered by the genomic position of the operon in which they occur (Figure 1C). Such a correlation between matrices would be consistent with previous analyses demonstrating correlation of distance matrices for known interacting proteins [20]. We indeed found that the PE and PPE matrices were correlated, with a Pearson correlation coefficient of 0.84.
To assess the significance of the correlation of the PE and PPE matrices, we performed random shuffling of the matrices' gene order, thus removing any mapping of paired genomic position between the matrices. One million shuffling steps were performed, and the frequency with which the shuffled correlation exceeded the correlation from the operon-ordered matrices was recorded. The correlation of 0.84 was never exceeded in 106 matrix shuffling steps (the maximum correlation in any shuffling was 0.20). These results suggest that the PE and PPE matrices, ordered by operon position, may represent an optimal pairing of PE/PPE proteins, and, in light of previous findings of correlated distance matrices of interacting proteins [20], support the hypothesis that PE/PPE operons code complexes.
The correlation of the operon-ordered distance matrices can be visualized using phylogenetic trees to provide an intuitive feeling for the results. To illustrate this, we generated trees from distance matrices in the ClustalW program (Figure 2). In the two trees, operon-paired PE and PPE proteins are in the same-colored shaded region, illustrating similar topologies of the trees. This qualitative tree similarity illustrates the notion of coevolution of the PE and PPE families.
We next used the correlated 14×14 PE and PPE distance matrices as reference matrices to evaluate pairwise correlations between the 86 PE proteins and 65 PPE proteins in the Mtb genome, excluding those present in operon pairs. This was done by generating a distance vector of length 14 for each protein in a PE/PPE pair. The vector contained the distance between the protein being tested and the 14 members of the appropriate reference matrix (PE or PPE). The Pearson correlation for the two vectors was calculated to obtain a measure of the coevolution of the test PE/PPE pair (Figure 1E). Notice that here we are taking the correlation of two vectors of length 14, as opposed to our earlier calculation of the correlation of the two 14×14 matrices. The coevolution of 5590 PE/PPE pairs was evaluated using this approach. We define the coefficient Cij, as the correlation which measures the coevolution of PEi with PPEj.
Pairwise correlations between all PEi and PPEj (Cij) were further processed using a reciprocal ranking procedure, to produce a predictive paralog matching score, Sij. This was done because we noticed that many PE/PPE pairs had high Cij values (average Cij = 0.54). The distribution of Cij is shown in Figure 3A (blue bars). From the histogram, it is clear that a great number of PE/PPE pairs have a high Cij. This may reflect the overall coevolution of the two families, but is not of use in a prediction scheme, as nearly all pairs have a high score. Such a result is inconsistent with our intuition that in a large collection of proteins, only a relatively small number of the possible pairs should interact. Further, we found that the Cij distributions for operon pairs and all PE/PPE pairs do not differ significantly (Kolmogorov–Smirnoff (KS) test, α = 0.05, k = 0.29, P = 0.16). In the reciprocal ranking procedure, the predicted complex of PEi and PPEj was assigned a high Sij only if PEi and PPEj were mutually at the top of each protein's list of interaction partners ranked by Cij. In other words, PEi and PPEj were required to be reciprocally the most coevolved partners in order to get a high Sij (see Materials and methods). Figure 3A, shows the distribution of Sij scores (red bars). The distribution of Sij suggests it is a more useful measure than Cij for complex prediction, as the bulk of the Sij scores are low (reflecting that in a large dataset, most protein pairs do not form complexes). The operon pairs have a significantly higher Sij than PE/PPE pairs overall (KS test, α = 0.05, k = 0.97, P–value<.0001), a result illustrated in Figure 3B. We conclude that the reciprocally ranked coevolution score, Sij, performs better than Cij for predicting protein interactions.
To evaluate the predictive value of the two pairwise PE/PPE scores, Cij and Sij, we assessed recovery of the 14 operon linked PE/PPE pairs used to generate the reference matrices. Of the 14 pairs, all were given Sij scores in the top 5% implying the method could be used to detect complexes with reasonably high accuracy. In contrast, only 3/14 (20%) of the operon pairs were given Cij scores in the top 5%. A mere 8/14 pairs (60%) had Cij scores above the median, implying poor recovery by the raw distance vector correlations. The relationship between Cij and Sij is illustrated in Figure 3B. The distribution of scores shows that Sij as a predictor of PE/PPE complexes gives much better recovery of operon-paired PE/PPE proteins than Cij, and therefore is likely a better indicator of interaction.
To further evaluate prediction accuracy, and to determine a prediction score threshold, we compared the sensitivity (also called true positive rate or TPR) and 1-specificity (also called the false positive rate or FPR) for Sij and Cij. Sets of positive and negative interactions were defined, and an Sij threshold of 0.75 was found to capture the best balance between sensitivity and specificity (Materials and methods). Applying a prediction criterion of Sij≥0.75 gave 289 PE/PPE pairs or roughly 5% of the possible 5,590 PE/PPE pairs genomewide. We therefore proceeded with our analysis by taking the predictions with Sij scores in the top 5%.
To see if the coevolution-based predictions were biologically sensible, we analyzed correlations in mRNA expression (which we call ‘coexpression’) of possible interacting PE/PPE pairs. We reasoned that interacting proteins would tend to be expressed at similar times to perform their biological functions together [33]–[37], and that if some of our predicted interactions were correct, we should see non-random enrichment in coexpression of the genes encoding the predicted complexes. Gene microarray data from Mtb was compiled from nine published datasets covering a broad range of experimental conditions (Table S1) in the Gene Expression Omnibus (GEO) database [38]. Vectors of expression values for each PE or PPE gene were generated, and used to derive a correlation, Rij, for the expression of each PE/PPE gene pair (see Materials and methods).
To confirm our intuition that predictions of PE/PPE coexpression and coevolution should overlap significantly, we analyzed the distribution of the two coevolution scores (Cij and Sij) combined with the coexpression score, Rij. The resulting distributions of the two score combinations are shown in Figure 4. In the figure, coevolution (Cij or Sij) is shown on the x axes; coexpression (Rij) is shown on the y axes. Red dots represent the 14 operon pairs; green dots represent the 182 inter-operon pairs (in which the PE and PPE are from different operon pairs); blue dots represent the other 5,394 genomewide PE/PPE pairs. Dashed lines are drawn to represent the upper 5% threshold for each score. Figure 4A illustrates our earlier assessment that Cij is not a useful prediction score due to the operon pairs not having a significantly different distribution from all PE/PPE pairs. Also notable in Figure 4A, only a minority of the operon pairs is found in the top 5% by both methods (3/14 operon pairs or 21%). Figure 4B shows better recovery of operon pairs in the top 5% by both methods (12/14 operon pairs or 86%). In light of these results, we concluded that the paralog matching score, Sij, is superior to Cij for predicting PE/PPE complexes. We therefore chose to combine Sij and Rij for subsequent predictions, and Cij was not used further in this study. The Sij and Rij scores for all PE/PPE pairs are provided in the Supporting Information (Dataset S3).
To assess the statistical significance of the overlap between predictions from coevolution and coexpression, we again employed a KS test. We asked whether the coexpression values of the top 5% coevolved PE/PPE genes (excluding the operon pairs) were higher than coexpression values of PE/PPE gene pairs overall. We found this to be the case (KS test, P = 0.02, α = 0.05, k = 0.09). From this we conclude that the PE and PPE proteins we predict to interact tend to be coexpressed, which we take as additional evidence for their possible interaction.
To assess the specificity of interaction in PE/PPE operon pairs, we analyzed coevolution and coexpression of inter-operon pairs, the PE/PPE pairs in which both proteins are from different operon pairs. A KS test showed that inter-operon pairs (Figure 4, green crosses) had significantly lower score distributions than all other pairs by both coevolution scores, Cij and Sij, and the coexpression score, Rij (P≪0.0001 in all tests). We noticed a bimodal distribution of Cij (Figure 4A). Using a k-means algorithm we identified two clusters: a larger (4,208 protein pairs), positive-valued cluster with mean Cij = 0.78 and a smaller (1,382 pairs), negative-valued cluster with mean Cij = −0.21. The negative cluster contained no operon pairs, and was more than twice as likely to contain inter-operon pairs than the higher group (7% of the negative cluster; 3% of the positive one). We interpret these results as evidence of negative selection, both at the amino acid and gene expression levels, against cross-reactivity of operon paired PE and PPE proteins, and conclude that PE/PPE operon pairs, in general, interact specifically.
To demonstrate that our method is extensible to protein families other than PE and PPE we studied the ESAT-6/CFP-10 (Esx) family of proteins, which include some secreted antigens [39]. We chose the Esx family because they, like PE and PPE, tend to be found in operon pairs, some of which are known to code interacting proteins [40]–[42]. We applied our method to the 22 Esx proteins in Mtb H37Rv, in an identical manner to our analysis of the PE/PPE pairs, and found that known Esx interacting pairs and operon pairs, were given a high Sij (coevolution) by our method, and that many of these were supported by a high Rij (coexpression) (Figure S1). The Sij and Rij scores for the Esx analysis are given in the Supporting Information (Dataset 4). We conclude that our method has the potential to predict interactions in protein families beyond PE and PPE.
To generate a high-confidence set of predicted PE/PPE complexes, we took the overlap of the top 5% by both coevolution (Sij) and coexpression (Rij), yielding the 35 pairs shown in Table 2. The same predicted interactions are shown in a network representation in Figure 5. Panel A shows that 6 of the 12 operon pairs in the top 35 predicted complexes are predicted to interact specifically. That is, the PEs in this group do not appear to interact with PPEs other than their operon partner, and vice versa. The specificity of interaction in operon pairs is also suggested by the tendency for inter-operon pairs to have low scores (Figure 4, green crosses). Figure 5B shows predicted cross-reaction of inter-operon pairs Rv3872/Rv1387, Rv1195/Rv3478, Rv3477/Rv1196, and Rv2769c/Rv1039c. Notice that the inter-operon interactions in Figure 5 are between pairs of proteins in the same colored regions in the phylogenetic trees in Figure 2, suggesting that pairs of paralogs with sufficiently similar sequences (nearby on a tree) could also cross-react. A high degree of mRNA coexpression (Table 2) provides additional evidence that there could be some cross reactivity between these PE/PPE operons. Cross-reactivity between genome-paired Esx proteins has been noted previously in Mtb [40], and it may be that the subunits of closely related PE/PPE complexes can similarly cross-react to confer functional flexibility as with the Esx family. However, our finding of negative coevolution and coexpression of the majority of the 196 possible inter-operon pairs (Figure 4) suggest that the four interactions we predict are exceptions rather than the rule.
Figure 5C shows possible cell surface-associated PE/PPE complexes. Four of the 6 PE proteins are PE_PGRS proteins (Rv0109, Rv0754, Rv1803c, and Rv2487c,), which are thought to be variable surface antigens displayed on the exterior of Mtb cells [5]. The other two PE proteins (Rv0151c and Rv0160c) were predicted in a previous study to contain membrane beta barrels [43], and thus are also likely localized to the cell surface. The PE and PPE proteins here appear to have multiple interaction specificities, particularly PE proteins Rv0160c and Rv0754, each of which is linked to several PPEs. Rv0160c and Rv0754 also have overlapping patterns of interaction as they share three predicted PPE partners: Rv0355c, Rv0442c, and Rv1801. However, in contrast to the example of Rv1195 and Rv3477 (Figure 5B), Rv0160c and Rv0754 do not have remarkably high sequence similarity (58%; both proteins are more similar to many other PEs) or close distance in the PE phylogenetic distance matrix. We would therefore conclude that the patterns of cross-reactivity shown in Figure 5C cannot be explained simply by sequence similarity. Future structural studies could reveal the detailed residue interactions responsible for complex formation among these proteins. Because of the possible cell surface localization of the proteins in Figure 5C, it may be that they are part of multiprotein cell-surface complexes involving varying combinations of PPE proteins interacting with surface-localized PE proteins. We truncated all proteins to include only the PE or PPE domains; therefore our method predicts that PE_PGRS and membrane-associated PE proteins interact with PPEs through their PE domains. All of these interpretations await experimental confirmation.
Two of the 14 Mtb H37Rv operon pairs (Rv1040c/Rv1039c and Rv3622c/Rv3621c) were not among the 35 putatively interacting PE/PPE pairs identified by our procedure. Both of these operon pairs exceeded our coevolution (Sij) threshold, but were slightly below our coexpression (Rij) threshold of 0.34 (Rij = 0.26 and 0.25, respectively; see also Figure 4). Employing a lower Rij threshold, for example the 85th percentile (Rij = 0.24), would result in the predicted interaction of all 14 operon pairs.
To predict new PE/PPE interactions we analyzed sequences homologous to the interacting domains in the known PE/PPE complex [13], without explicitly limiting our analysis to defined interface residues between the subunits as in [44]. We reasoned that, because both PE and PPE are helical, that in different paralogous complexes the registry of the helices could change, bringing different residues into the interface. This may be especially true for PE and PPE, which are highly elongated in shape (the characterized PE/PPE complex is about 108 Å long by 26 Å wide [13]), and, as a result, most residues in either subunit are near (within 10 Å) a residue in the other subunit. Computational and experimental studies [29],[45] have found evidence for energetic coupling of distant residues (>10 Å) in a number of protein families. While these studies focused on protein monomers, it is possible that the finding of long-range residue-residue interactions could also apply to complexes. With this in mind, we were cautious about excluding regions homologous to either subunit in which a mutation might not be ‘felt’ in the other subunit.
To test whether excluding residues distant from the complex interface would influence our results, we inspected the PE/PPE structure [13] and found a contiguous region spanning residues 101–148 of the PPE with C-α distances greater than 10 Å from the nearest residue of the PE. We constructed a modified PPE alignment omitting these residues (and homologous regions of aligned PPE proteins) and repeated our analysis, and found no significant change in the distance matrix made from the unmodified PPE alignment (correlation of PE and modified PPE distance matrices = 0.84; no improved correlations in 106 matrix shufflings). We conclude that including some non-interface residues, at least in the case of PE/PPE, did not significantly bias our results.
The large expansion of the PE and PPE genes in Mtb [12] allowed us to obtain results using the genome sequence of Mtb strain H37Rv alone, without bringing in data from other genomes. Our method could therefore in principle be applied to large families of interacting paralogs in microbial genomes without necessarily having any closely related genome sequences from other microbes.
To explore whether adding genome data from other mycobacteria would improve our results, we searched an additional 14 mycobacterial genomes for PE/PPE operon pairs. Orthologs to H37Rv PE/PPE operon pairs are summarized in Figure S2. Ninety additional operon pairs were found, adding to the 14 of Mtb H37Rv to give 104 operon pairs in total. However, only 24 of these operon pairs had PE or PPE domains with amino acid sequences different (usually just by an amino acid or two) from the 14 H37Rv operon pairs. The 38-protein reference matrices derived from multiple genomes showed a reasonable increase in correlation over the 14-protein reference matrices from H37Rv (0.91 up from 0.84; no higher correlations yielded by shuffling the matrices' gene order 106 times). No clear improvement in the results was evident from using the multi-genome set of operon pairs (data not shown), perhaps because each of the added 24 sequences was highly similar to one already in the 14 H37Rv operon pairs. However, it is likely that as still more mycobacterial genomes are sequenced, new sequence variants of PE and PPE domains will be discovered, and this may improve our results.
In the five Mtb strains analyzed, 65/70 or 93% of the operon pairs were conserved (Figure S2). Five operon pairs were missing either the PE or PPE protein. In particular, the Mtb C strain has single-gene deletions in three operon pairs (Rv1040c/Rv1039c, Rv1806/Rv1807, and Rv2107/2108. It is possible that in Mtb strains with ‘broken’ operon pairs, another interacting partner is able to interact with the orphaned gene, possibly restoring the PE/PPE complex's function, or introducing new complexes that help these strains survive in their environmental niches.
It is possible that a single Mtb strain (in this case H37Rv), in which the PE and PPE families are more expanded relative to other mycobacteria [12], provides a broad sampling of the tolerated residue variations in these families, as proteins with many paralogs are thought to be under negative selective pressure on their interactions with paralogs other than their cognate partner [46]. Thus, the more specifically interacting PE/PPE protein pairs there are in a genome, the more residue variation we might see, due to positive selective pressure on interaction specificity. The extent of interaction promiscuity between the PE and PPE families is unknown, but our observations are consistent with negative selection on promiscuous interaction. This suggests that there may be some advantage to Mtb in maintaining interaction specificity of PE and PPE proteins, at least in those that are operon-paired (and which we also predicted to interact with some degree of specificity, Figure 5A). We conclude that our analysis was not significantly affected by the inclusion of only a single genome, and that this could be a useful approach for mining the interactions of newly sequenced genomes for which there are initially no (or just a few) related genomes to compare to.
Related prediction methods [20],[21],[24] that compared phylogenetic distance matrices relied on a training set of experimentally determined protein-protein interactions to build the reference matrices, or large sets of orthologs of a few known interacting proteins [25]. At the time of this study, there is only a single experimentally characterized PE/PPE interaction [13]. We resolved this impasse by employing the operon method [15]–[18] to define a high-confidence set of predicted interactions (including the known complex) to build phylogenetic distance matrices capable of capturing some of the residue covariation patterns in PE/PPE complexes. The validity of our results depends on future verification of this assumption. Even so, the high degree of coevolution and coexpression seen in operon-paired PE/PPE genes, combined with definitive experimental characterization of at least one PE/PPE complex [13], implies that our assumption that operon-paired PE/PPE genes code complexes is fair.
We envision testing the PE/PPE complexes predicted by our approach with a scalable high-throughput strategy. Rapid cloning methods such as ligation-independent cloning (LIC) [47]–[49] could be employed to rapidly build up PE/PPE co-expression plasmids like that described for the Rv2431c/Rv2430c complex [13]. These strategies allow the experimentor to avoid time-consuming restriction and ligation steps during cloning. Using LIC, putative PE/PPE complexes could rapidly be screened to assess whether further study, including structural characterization, would be worthwhile. Using our set of predicted interactions to prioritize experiments would likely reduce the required number of assays to successfully characterize complexes. As all PE/PPE pairs in our set of operon pairs were found in the top 5% of predictions it is possible that the success rate of such prioritized assays, relative to all-vs-all screening, could be increased by an order of magnitude or more.
A method for predicting the interactions of the PE and PPE families of proteins in M. tuberculosis, beyond those simply linked by the operon method, is proposed. The method combines known interacting domain structure, genomic operon organization, and protein coevolution, and predicts that 35 pairs of PE and PPE proteins interact. Our method can be applied to a single genome if sufficient numbers of paralogs are present, or could be used in a multi-genome framework. A subset of the predictions from our coevolution-based method is confirmed by high mRNA coexpression, suggesting their biological relevance, and likely weeding out false positives. Our results may be a useful starting point for experimentally probing the interactions of PE, PPE, and other microbial protein families.
Annotations for the Mtb H37Rv genome were downloaded from the NCBI FTP site (ftp://ftp.ncbi.nih.gov). PE and PPE genes were identified from these annotations. The gene coordinates and orientation information provided in the annotations were used to compile a list of adjacent PE/PPE pairs in the same orientation, with the PE protein located 5′ (upstream) to the PPE protein, and with no more than 100 base pairs intergenic separation. Increasing the intergenic distance cutoff to 500 base pairs did not result in any additional PE/PPE pairs.
280 multiprotein complexes involving 692 proteins, and 2,909 unique operons involving 4,510 genes, were extracted from the EcoCyc database [30]. A total of 1918 pairwise protein interactions were found in the complexes. Of these pairs, 942 or 49% of the pairs were also found together in an operon. To assess the significance of the overlap, the identities of the proteins in the complexes were randomized (each protein was replaced with a unique, random E. coli protein), and the co-occurrence in operons was reassessed. One thousand shuffling trials were performed and the overlap of 49% was not met or exceeded in any of the trials. The maximum overlap achieved in any trial was 2%.
For each of the PE and PPE families, family member sequences were selected from the SwissProt database [50] by two criteria: 1. the sequence was annotated as belonging to either the PE or PPE protein families in Pfam, or 2. the protein was otherwise annotated as belonging to one of these families. Multiple alignment of the protein sequences was performed using the ClustalW program using default parameters, and a secondary structure profile generated by the DSSP program [32] and the known structure [13]. Alignments were visually inspected and hand-edited to omit sequences with obvious low homology. Rv3893c (PE36), though classified by SwissProt as a PE protein, appeared to be an outlier and was therefore removed from the multiple alignment. Rv3892c (PPE69), its genome-paired neighbor, was kept in the PPE multiple alignment. Because of the omission of Rv3893c, the PE/PPE pair of Rv3893c/Rv3892c was not included in the genome-paired reference set. Rv3892c was included in subsequent predictions.
The resulting structure-based alignments had 87 proteins in the PE alignment, and 65 proteins in the PPE alignment. Multiple alignments were truncated to include only those positions homologous to proteins in the known PE/PPE complex of Rv2431c/Rv2430c. Rv2431c and Rv2430c are among the shortest members of the PE and PPE families, respectively, and inspection of the complex structure suggests that these sequences may represent the minimal interface regions necessary to form a complex. Many PE and PPE proteins have other domains, low-complexity regions, and transmembrane domains C-terminal to their PE or PPE domain. These regions are unlikely to participate directly in the PE/PPE interaction; this was our reasoning for removing these regions from the alignment.
From the full-family multiple alignments, Phylip distance matrices were generated in the ClustalW program [31]. This resulted in an 87×87 matrix for the PE family and a 65×65 matrix for the PPE family. These matrices would be used subsequently in predictions to give us the distance between any pair of PE sequences and any pair of PPE sequences.
Next, phylogenetic distance matrices were made for just the 14 pairs of operon-paired PE and PPE proteins. This was done by extracting the 14 operon-paired sequences from each of the full-family multiple alignments. The resulting subalignments were used to generate a 14×14 PE distance matrix and a 14×14 PPE matrix. Importantly, the two 14×14 matrices were ordered so that the ith protein in the PE matrix was the operon partner of the ith protein in the PPE matrix. We would later use these matrices as ‘reference’ matrices for prediction of non-operon-paired PE/PPE complexes.
Phylogenetic trees were generated from the PE and PPE 14-sequence subalignments using the ClustalW program [31]. The correction for multiple substitutions was not used to generate the trees. Bootstrapping of the trees was done within the ClustalW program.
Let X be the 14×14 distance matrix of PE proteins, and Y be the 14×14 distance matrix of PPE proteins. Xij is the percent divergence of PEi and PEj; likewise Yij is the percent divergence of PPEi and PPEj. Xi is the vector of length 14 for the distances of PEi from all PEj (including for i = j, in which case Yij = 0.0); likewise Yi.
To determine the correlation between the ordered 14×14 distance matrices, X and Y, the Pearson correlation is taken. To avoid counting protein distances twice, only the unique elements in the matrices are taken (that is, we count i,j pairs but not j,i).L = 14, the number of operons with paired PE and PPE genes. rXY is a measure of the coevolution of the operon-paired subsets of the PE and PPE protein families.
Next we derive Cij, a measure of the coevolution of PEi with PPEj, for all PE and PPE proteins in the Mtb genome, including but not limited to proteins in the operon-paired set. For PEi, we generate Ai, a distance vector of length 14, containing the distances from PEi to each of the 14 PEk in the 14×14 reference matrix. Bj is equivalently generated for PPEj. To get Cij, a measure of the coevolution of PEi with PPEj, the Pearson correlation of Ai and Bj is calculated.Note that here we are taking the correlation of two vectors with length 14.
To generate the paralog matching score, Sij, a reciprocal ranking procedure was used. For each PE, a ranked list of the most coevolved PPEs was produced. The same was done for PPEs. Then, for each PEi, we recorded the position of each PPEj protein in the PE's list of PPEs ranked by Cij to give ri→j, the number of PPEs ranking below PPEj. The reciprocal procedure was performed to yield rj→i.N is the total number of PE/PPE pairs. Using the above formula, a high score is assigned only if both pairs were high on each other's list of most coevolved potential interacting partners.
Positive examples of PE/PPE complexes were defined as all of the 14 operon-paired proteins. A dataset of negative PE/PPE interactions are not currently available, so we made the assumption that operon-paired PE/PPE interactions were specific, and therefore a PE in one operon would not interact with a PPE in a different operon. 182 inter-operon PE/PPE pairs resulted, and were used as a negative set.
Next, for a range of prediction thresholds from 0.0 to 1.0, true positive (TP), true negative (TN), false positive (FP), and false negative (FN) rates were determined. The sensitivity, or true positive rate (TPR), was calculated as1-specificity, or the false positive rate (FPR), was calculated as
In a receiver operator characteristic curve (not shown), the prediction threshold corresponding to the upper left-most portion of the curve represents the optimum compromise between TPR and FPR. For prediction with Sij, this threshold was roughly 0.75. Taking all PE/PPE pairs with Sij≥0.75 gives roughly 5% of the total 5,590 possible PE/PPE pairs, in which all 14 of the operon pairs were included.
Nine Mtb gene expression datasets (Table S1) in the Gene Expression Omnibus (GEO) [38] were downloaded. All available Mtb datasets in GEO were used excluding for consistency those that studied deletion mutants or attenuated strains. Also for consistency, only datasets that reported gene expression changes as a ratio of a sample and a reference were used. Gene expression data from the nine studies were represented as a matrix where the rows were genes and the columns were experiments. To construct an expression vector for a gene, the data from each of the nine studies were concatenated. Combined expression vectors were made up from the field labeled ‘VALUE’ in the data files. In all data sets, this value represents the measured expression level of a gene under experimental conditions versus that gene in a reference sample. Various normalization schemes were applied by the authors of the individual datasets to correct for scale differences due to differing intensities among genes in response to different experiments. Because of the difficulty in combining these schemes to make a self-consistent combined dataset, we chose not to further normalize the expression data.
Correlation coefficients of gene expression vectors were calculated for all possible pairs of genes. To obtain a correlation coefficient for genes i and j over N experiments, the Pearson correlation coefficient, Rij, was used.where gix and gjx are the expression values reported in the GEO data file for genes i and j, respectively, in experiment x. For each pair of genes analyzed, combined expression vectors were truncated by deleting experiments in which either or both genes had missing values. Thus N varied for each pair of genes assessed. In all, 734 experiments were used for the inference of pairwise coexpression relationships between pairs of PE and PPE genes.
The Kolmogorov–Smirnoff (KS) test asks whether two collections of random samples come from the same distribution. We want to know if the coexpression scores for a group of PE/PPE gene pairs predicted to code interacting proteins has a different distribution (with a higher mean) than PE/PPE gene pairs overall. Because we expect the linked proteins to have a higher-valued mean than the unlinked proteins, we used the one-tailed version of the KS test. An α significance criteria of 0.05 was used for hypothesis acceptance/rejection.
The PE/PPE complex described in [13] was analyzed using the RasMol program [51]. The structure was visually inspected to identify a contiguous region spanning residues 101–148 of the PPE with C-α distances greater than 10 Å from the nearest residue of the PE. The PPE multiple alignment was modified by deleting all columns that aligned to this contiguous region.
The 22 Esx family genes were identified in Mtb H37Rv from NCBI annotations. The genes were divided into two groups: ESAT-6 paralogs (12 proteins) and CFP-10 paralogs (10 proteins). We based this categorization on the observation that 20 of the 22 the Esx genes are organized into 10 adjacent (operon) pairs on the genome, with a gene similar to CFP-10 lying upstream from a gene similar to ESAT-6. We therefore categorized upstream genes as CFP-10 paralogs and downstream genes as ESAT-6 paralogs. The two annotated Esx genes not in operon pairs, Rv1793 and Rv3017c, were both judged to be ESAT-6 paralogs from visual inspection of a multiple alignment. The 10 Esx operon pairs were used to build reference matrices as with PE and PPE. Coevolution (Sij) and coexpression (Rij) scores were derived using the same procedure as for PE and PPE.
For each PE or PPE gene in an operon pair, orthologs were manually extracted from the TB Database (http://www.tbdb.org). The results are summarized in a tab-delimited file in the Supporting Information (Dataset 5).
The ORF identifiers, gene names, and SwissProt accession codes of the PE and PPE proteins analyzed in this study are listed in Text S1.
|
10.1371/journal.pcbi.1006785 | Assessing the performance of real-time epidemic forecasts: A case study of Ebola in the Western Area region of Sierra Leone, 2014-15 | Real-time forecasts based on mathematical models can inform critical decision-making during infectious disease outbreaks. Yet, epidemic forecasts are rarely evaluated during or after the event, and there is little guidance on the best metrics for assessment. Here, we propose an evaluation approach that disentangles different components of forecasting ability using metrics that separately assess the calibration, sharpness and bias of forecasts. This makes it possible to assess not just how close a forecast was to reality but also how well uncertainty has been quantified. We used this approach to analyse the performance of weekly forecasts we generated in real time for Western Area, Sierra Leone, during the 2013–16 Ebola epidemic in West Africa. We investigated a range of forecast model variants based on the model fits generated at the time with a semi-mechanistic model, and found that good probabilistic calibration was achievable at short time horizons of one or two weeks ahead but model predictions were increasingly unreliable at longer forecasting horizons. This suggests that forecasts may have been of good enough quality to inform decision making based on predictions a few weeks ahead of time but not longer, reflecting the high level of uncertainty in the processes driving the trajectory of the epidemic. Comparing forecasts based on the semi-mechanistic model to simpler null models showed that the best semi-mechanistic model variant performed better than the null models with respect to probabilistic calibration, and that this would have been identified from the earliest stages of the outbreak. As forecasts become a routine part of the toolkit in public health, standards for evaluation of performance will be important for assessing quality and improving credibility of mathematical models, and for elucidating difficulties and trade-offs when aiming to make the most useful and reliable forecasts.
| During epidemics, reliable forecasts can help allocate resources effectively to combat the disease. Various types of mathematical models can be used to make such forecasts. In order to assess how good the forecasts are, they need to be compared to what really happened. Here, we describe different approaches to assessing how good forecasts were that we made with mathematical models during the 2013–16 West African Ebola epidemic, focusing on one particularly affected area of Sierra Leone. We found that, using the type of models we used, it was possible to reliably predict the epidemic for a maximum of one or two weeks ahead, but no longer. Comparing different versions of our model to simpler models, we further found that it would have been possible to determine the model that was most reliable at making forecasts from early on in the epidemic. This suggests that there is value in assessing forecasts, and that it should be possible to improve forecasts by checking how good they are during an ongoing epidemic.
| Forecasting the future trajectory of cases during an infectious disease outbreak can make an important contribution to public health and intervention planning. Infectious disease modellers are now routinely asked for predictions in real time during emerging outbreaks [1]. Forecasting targets can revolve around expected epidemic duration, size, or peak timing and incidence [2–5], geographical distribution of risk [6], or short-term trends in incidence [7, 8]. However, forecasts made during an outbreak are rarely investigated during or after the event for their accuracy, and only recently have forecasters begun to make results, code, models and data available for retrospective analysis.
The growing importance of infectious disease forecasts is epitomised by the growing number of so-called forecasting challenges. In these, researchers compete in making predictions for a given disease and a given time horizon. Such initiatives are difficult to set up during unexpected outbreaks, and are therefore usually conducted on diseases known to occur seasonally, such as dengue [7, 9, 10] and influenza [11]. The Ebola Forecasting Challenge was a notable exception, triggered by the 2013–16 West African Ebola epidemic and set up in June 2015. Since the epidemic had ended in most places at that time, the challenge was based on simulated data designed to mimic the behaviour of the true epidemic instead of real outbreak data. The main lessons learned were that 1) ensemble estimates outperformed all individual models, 2) more accurate data improved the accuracy of forecasts and 3) considering contextual information such as individual-level data and situation reports improved predictions [12].
In theory, infectious disease dynamics should be predictable within the timescale of a single outbreak [13]. In practice, however, providing accurate forecasts during emerging epidemics comes with particular challenges such as data quality issues and limited knowledge about the processes driving growth and decline in cases. In particular, uncertainty about human behavioural changes and public health interventions can preclude reliable long-term predictions [14, 15]. Yet, short-term forecasts with an horizon of a few generations of transmission (e.g., a few weeks in the case of Ebola), can yield important information on current and anticipated outbreak behaviour and, consequently, guide immediate decision making.
The most recent example of large-scale outbreak forecasting efforts was during the 2013–16 Ebola epidemic, which vastly exceeded the burden of all previous outbreaks with almost 30,000 reported cases resulting in over 10,000 deaths in the three most affected countries: Guinea, Liberia and Sierra Leone. During the epidemic, several research groups provided forecasts or projections at different time points, either by generating scenarios believed plausible, or by fitting models to the available time series and projecting them forward to predict the future trajectory of the outbreak [16–26]. One forecast that gained particular attention during the epidemic was published in the summer of 2014, projecting that by early 2015 there might be 1.4 million cases [27]. This number was based on unmitigated growth in the absence of further intervention and proved a gross overestimate, yet it was later highlighted as a “call to arms” that served to trigger the international response that helped avoid the worst-case scenario [28]. While that was a particularly drastic prediction, most forecasts made during the epidemic were later found to have overestimated the expected number of cases, which provided a case for models that can generate sub-exponential growth trajectories [29, 30].
Traditionally, epidemic forecasts are assessed using aggregate metrics such as the mean absolute error (MAE) [12, 31, 32]. This, however, only assesses how close the most likely or average predicted outcome is to the true outcome. The ability to correctly forecast uncertainty, and to quantify confidence in a predicted event, is not assessed by such metrics. Appropriate quantification of uncertainty, especially of the likelihood and magnitude of worst case scenarios, is crucial in assessing potential control measures. Methods to assess probabilistic forecasts are now being used in other fields, but are not commonly applied in infectious disease epidemiology [33, 34].
We produced weekly sub-national real-time forecasts during the Ebola epidemic, starting on 28 November 2014. Plots of the forecasts were published on a dedicated web site and updated every time a new set of data were available [35]. They were generated using a model that has, in variations, been used to forecast bed demand during the epidemic in Sierra Leone [21] and the feasibility of vaccine trials later in the epidemic [36, 37]. During the epidemic, we provided sub-national forecasts for the three most affected countries (at the level of counties in Liberia, districts in Sierra Leone and prefectures in Guinea).
Here, we apply assessment metrics that elucidate different properties of forecasts, in particular their probabilistic calibration, sharpness and bias. Using these methods, we retrospectively assess the forecasts we generated for Western Area in Sierra Leone, an area that saw one of the greatest number of cases in the region and where our model informed bed capacity planning.
This study has been approved by the London School of Hygiene & Tropical Medicine Research Ethics Committee (reference number 8627).
Numbers of suspected, probable and confirmed Ebola cases at sub-national levels were initially compiled from daily Situation Reports (or SitReps) provided in PDF format by Ministries of Health of the three affected countries during the epidemic [21]. Data were automatically extracted from tables included in the reports wherever possible and otherwise manually converted by hand to machine-readable format and aggregated into weeks. From 20 November 2014, the World Health Organization (WHO) provided tabulated data on the weekly number of confirmed and probable cases. These were compiled from the patient database, which was continuously cleaned and took into account reclassification of cases avoiding potential double-counting. However, the patient database was updated with substantial delay so that the number of reported cases would typically be underestimated in the weeks leading up to the date at which the forecast was made. Because of this, we used the SitRep data for the most recent weeks until the latest week in which the WHO case counts either equalled or exceeded the SitRep counts. For all earlier times, the WHO data were used.
We used a semi-mechanistic stochastic model of Ebola transmission described previously [21, 38]. Briefly, the model was based on a Susceptible–Exposed–Infectious–Recovered (SEIR) model with fixed incubation period of 9.4 days [39], following an Erlang distribution with shape 2. The country-specific infectious period was determined by adding the average delay to hospitalisation to the average time from hospitalisation to death or discharge, weighted by the case-fatality rate. Cases were assumed to be reported with a stochastic time-varying delay. On any given day, this was given by a gamma distribution with mean equal to the country-specific average delay from onset to hospitalisation and standard deviation of 0.1 day. We allowed transmission to vary over time in order to capture behavioural changes in the community, public health interventions or other factors affecting transmission for which information was not available at the time. The time-varying transmission rate was modelled using a daily Gaussian random walk with fixed volatility (or standard deviation of the step size) which was estimated as part of the inference procedure (see below). We log-transformed the transmission rate to ensure it remained positive. The behaviour in time can be written as
d log β t = σ d W t (1)
where βt is the time-varying transmission rate, Wt is the Wiener process and σ the volatility of the transmission rate. The basic reproduction number R0,t at any time was obtained by multiplying βt with the average infectious period. In fitting the model to the time series of cases we extracted posterior predictive samples of trajectories, which we used to generate forecasts.
Each week, we fitted the model to the available case data leading up to the date of the forecast. Observations were assumed to follow a negative binomial distribution. Since the ssm software used to fit the model only implemented a discretised normal observation model, we used a normal approximation of the negative binomial for observations, potentially introducing a bias at small counts. Four parameters were estimated in the process: the initial basic reproduction number R0 (uniform prior within (1, 5)), initial number of infectious people (uniform prior within (1, 400)), overdispersion of the (negative binomial) observation process (uniform prior within (0, 0.5)) and volatility of the time-varying transmission rate (uniform prior within (0, 0.5)). We confirmed from the posterior distributions of the parameters that these priors did not set any problematic bounds. Samples of the posterior distribution of parameters and state trajectories were extracted using particle Markov chain Monte Carlo [40] as implemented in the ssm library [41]. For each forecast, 50,000 samples were extracted and thinned to 5000.
We used the samples of the posterior distribution generated using the Monte Carlo sampler to produce predictive trajectories, using the final values of estimated state trajectories as initial values for the forecasts and simulating the model forward for up to 10 weeks. While all model fits were generated using the same model described above, we tested a range of different predictive model variants to assess the quality of ensuing predictions. We tested variants where trajectories were stochastic (with demographic stochasticity and a noisy reporting process), as well as ones where these sources of noise were removed for predictions. We further tested predictive model variants where the transmission rate continued to follow a random walk (unbounded, on a log-scale), as well as ones where the transmission rate stayed fixed during the forecasting period. When the transmission rate remained fixed for prediction, we tested variants where we used the final value of the transmission rate and ones where this value was averaged over a number of weeks leading up to the final fitted point, to reduce the potential influence of the last time point, at which the transmission rate may not have been well identified. We tested variants where the predictive trajectory was based on the final values and start at the last time point, and ones where it started at the penultimate time point, which could, again, be expected to be better informed by the data. For each model and forecast horizon, we generated point-wise medians and credible intervals from the sample trajectories.
To assess the performance of the semi-mechanistic transmission model we compared it to three simpler null models: two representing the constituent parts of the semi-mechanistic model, and a non-mechanistic time series model. For the first null model, we used a deterministic model that only contained the mechanistic core of the semi-mechanistic model, that is a deterministic SEIR model with fixed transmission rate and parameters otherwise the same as in the model described before [21]:
d S d t= − R 0 Δ I c + I h N S (2) d E 1 d t= − R 0 Δ I c + I h N S − 2 ν E 1 (3) d E 2 d t= 2 ν E 1 − 2 ν E 2 (4) d I c d t= 2 ν E 2 − τ I c (5) d I h d t= τ I c − γ I h (6) d R d t= γ I h (7) d A d t= τ I c (8) Y t∼ NB ( A t − A t − 1 , ϕ ) (9)
where Yt are observations at times t, S is the number susceptible, E the number infected but not yet infectious (split into two compartments for Erlang-distributed permanence times with shape 2), Ic is the number infectious and not yet notified, Ih is the number infectious and notified, R is the number recovered or dead, A is an accumulator for incidence, R0 is the basic reproduction number, Δ = 1/τ + 1/ν is the mean time from onset to outcome, 1/ν is the mean incubation period, 1/τ + 1/γ is the mean duration of infectiousness, 1/τ is the mean time from onset to hospitalisation 1/γ the mean duration from notification to outcome and NB(μ, ϕ) is a negative binomial distribution with mean μ and overdispersion ϕ. All these parameters were informed by individual patient observations [39] except the overdispersion in reporting ϕ, and the basic reproduction number R0, which were inferred using Markov-chain Monte Carlo with the same priors as in the semi-mechanistic model.
For the second null model, we used an unfocused model where the weekly incidence Z itself was modelled using a stochastic volatility model (without drift), that is a daily Gaussian random walk, and forecasts generated assuming the weekly number of new cases was not going to change:
d log Z= σ dW (10) Y t∼ NB ( Z t , ϕ ) (11)
where Y are observations, σ is the intensity of the random walk and ϕ the overdispersion of reporting (both estimated using Markov-chain Monte Carlo) and dW is the Wiener process.
Lastly, we used a null model based on a non-mechanistic Bayesian autoregressive AR(1) time series model:
α t + 1∼ N ( ϕ α t , σ α ) (12) Y t *∼ N ( α t , σ Y * ) (13) Y t= max ( 0 , [ Y t * ] ) (14)
where ϕ, σα and σY* were estimated using Markov-chain Monte Carlo, and […] indicates rounding to the nearest integer. An alternative model with Poisson distributed observations was discarded as it yielded poorer predictive performance.
The deterministic and unfocused models were implemented in libbi [42] via the RBi [43] and RBi.helpers [44] R packages [45]. The Bayesian autoregressive time series model was implemented using the bsts package [46].
The paradigm for assessing probabilistic forecasts is that they should maximise the sharpness of predictive distributions subject to calibration [47]. We therefore first assessed model calibration at a given forecasting horizon, before assessing their sharpness and other properties.
Calibration or reliability [48] of forecasts is the ability of a model to correctly identify its own uncertainty in making predictions. In a model with perfect calibration, the observed data at each time point look as if they came from the predictive probability distribution at that time. Equivalently, one can inspect the probability integral transform of the predictive distribution at time t [49],
u t = F t ( x t ) (15)
where xt is the observed data point at time t ∈ t1, …, tn, n being the number of forecasts, and Ft is the (continuous) predictive cumulative probability distribution at time t. If the true probability distribution of outcomes at time t is Gt then the forecasts Ft are said to be ideal if Ft = Gt at all times t. In that case, the probabilities ut are distributed uniformly.
In the case of discrete outcomes such as the incidence counts that were forecast here, the PIT is no longer uniform even when forecasts are ideal. In that case a randomised PIT can be used instead:
u t = P t ( k t ) + v ( P t ( k t ) − P t ( k t − 1 ) ) (16)
where kt is the observed count, Pt(x) is the predictive cumulative probability of observing incidence k at time t, Pt(−1) = 0 by definition and v is standard uniform and independent of k. If Pt is the true cumulative probability distribution, then ut is standard uniform [50]. To assess calibration, we applied the Anderson-Darling test of uniformity to the probabilities ut. The resulting p-value was a reflection of how compatible the forecasts were with the null hypothesis of uniformity of the PIT, or of the data coming from the predictive probability distribution. We calculated the mean p-value of 10 samples from the randomised PIT and found the corresponding Monte-Carlo error to be negligible (maximum standard deviation: sp = 0.003). We considered that there was no evidence to suggest a forecasting model was miscalibrated if the p-value found was greater than a threshold of p ≥ 0.1, some evidence that it was miscalibrated if 0.01 < p < 0.1, and good evidence that it was miscalibrated if p ≤ 0.01. In this context it should be noted, though, that uniformity of the (randomised) PIT is a necessary but not sufficient condition of calibration [47]. The p-values calculated here merely quantify our ability to reject a hypothesis of good calibration, but cannot guarantee that a forecast is calibrated. Because of this, other indicators of forecast quality must be considered when choosing a model for forecasts.
All of the following metrics are evaluated at every single data point. In order to compare the forecast quality of models, they were averaged across the time series.
Sharpness is the ability of the model to generate predictions within a narrow range of possible outcomes. It is a data-independent measure, that is, it is purely a feature of the forecasts themselves. To evaluate sharpness at time t, we used the normalised median absolute deviation about the median (MADN) of y S t ( P t ) = 1 0 . 675 median ( | y − median ( y ) | ) (17)
where y is a variable with CDF Pt, and division by 0.675 ensures that if the predictive distribution is normal this yields a value equivalent to the standard deviation. The MAD (i.e., the MADN without the normalising factor) is related to the interquartile range (and in the limit of infinite sample size takes twice its value), a common measure of sharpness [33], but is more robust to outliers [51]. The sharpest model would focus all forecasts on one point and have S = 0, whereas a completely blurred forecast would have S → ∞. Again, we used Monte-Carlo samples from Pt to estimate sharpness.
We further assessed the bias of forecasts to test whether a model systematically over- or underpredicted. We defined the forecast bias at time t as
B t ( P t , x t ) = 1 − ( P t ( x t ) + P t ( x t − 1 ) ) (18)
The least biased model would have exactly half of predictive probability mass not concentrated on the data itself below the data at time t and Bt = 0, whereas a completely biased model would yield either all predictive probability mass above (Bt = 1) or below (Bt = −1) the data.
We further evaluated forecasts using two proper scoring rules, that is scores which are minimised if the predictive distribution is the same as the one generating the data. These scores combine the assessment of calibration and sharpness for comparison of overall forecasting skill. The Ranked Probability Score (RPS) [52, 53] for count data is defined as [50]
RPS ( P t , x t ) = ∑ k = 0 ∞ ( P t ( k ) − 1 ( k ≥ x t )) 2 . (19)
It reduces to the mean absolute error (MAE) if the forecast is deterministic and can therefore be seen as its probabilistic generalisation for discrete forecasts. A convenient equivalent formulation for predictions generated from Monte-Carlo samples is [47, 50]
RPS ( P t , x t ) = E P t | X − x t | − 1 2 E P t | X − X ′ | , (20)
where X and X′ are independent realisations of a random variable with cumulative distribution Pt.
The Dawid-Sebastiani score (DSS) only considers the first two moments of the predictive distribution and is defined as [50]
DSS ( P t , x t ) = ( x t − μ P t σ P t) 2 + 2 log σ P t (21)
where μ P t and σ P t are the mean and standard deviation of the predictive probability distribution, respectively, estimated here using Monte-Carlo samples.
For comparison, we also evaluated forecasts using the absolute error (AE) of the median forecast, that is
AE ( P t , x t ) = | median P t ( X ) − x t | (22)
where X is a random variable with cumulative distribution Pt.
All scoring metrics used are implemented in the R package accompanying the paper. The goftest package was used for the Anderson-Darling test [54] and the scoringRules package for the RPS and DSS [55].
The semi-mechanistic model used to generate real-time forecasts during the epidemic was able to reproduce the trajectories up to the date of each forecast, following the data closely by means of the smoothly varying transmission rate (Fig 1). The overall behaviour of the reproduction number (ignoring depletion of susceptibles which did not play a role at the population level given the relatively small proportion of the population infected) was one of a near-monotonic decline, from a median estimate of 2.9 (interquartile range (IQR) 2.1–4, 90% credible interval (CI) 1.2–6.9) in the first fitted week (beginning 10 August, 2014) to a median estimate of 1.3 (IQR 0.9–1.9, 90% CI 0.4–3.7) in early November, 0.9 (IQR 0.6–1.3, 90% CI 0.2–2.2) in early December, 0.6 in early January (IQR 0.3–0.8, 90% CI 0.1–1.5) and 0.3 at the end of the epidemic in early February (IQR 0.2–0.4, 90% CI 0.1–0.9).
The epidemic lasted for a total of 27 weeks, with forecasts generated starting from week 3. For m-week ahead forecasts this yielded a sample size of 25 − m forecasts to assess calibration. Calibration of forecasts from the semi-mechanistic model were good for a maximum of one or two weeks, but deteriorated rapidly at longer forecasting horizons (Fig 2). The two semi-mechanistic forecast model variants with best calibration performance used deterministic dynamics starting at the last fitted data point (Table 1). Of these two, the forecast model that kept the transmission rate constant from the value at the last data point performed slightly better across forecast horizons than one that continued to change the transmission rate following a random walk with volatility estimated from the time series. There was no evidence of miscalibration in both of the models with best calibration performance for two-week ahead forecasts, but increasing evidence of miscalibration for forecast horizons of three weeks or more. Calibration of all model variants was poor four weeks or more ahead, and all the stochastic model variants were miscalibrated for any forecast horizon, including the one we used to publish forecasts during the Ebola epidemic (stochastic, starting at the last data point, no averaging of the transmission rate, no projected volatility).
The calibration of the best semi-mechanistic forecast model variant (deterministic dynamics, transmission rate fixed and starting at the last data point) was better than that of any of the null models (Fig 3A and Table 2) for up to three weeks. While there was no evidence for miscalibration of the autoregressive null model for 1-week-ahead forecasts, there was good evidence of miscalibration for longer forecast horizons. There was some evidence of miscalibration of the unfocused null model, which assumes that the same number of cases will be reported in the weeks following the week during which the forecast was made, for 1 week ahead and good evidence of miscalibration beyond. Calibration of the deterministic null model was poor for all forecast horizons.
The semi-mechanistic and deterministic models showed a tendency to overestimate the predicted number of cases, while the autoregressive and null models tended to underestimate (Fig 3B and and Table 2). This bias increased with longer forecast horizons in all cases. The best calibrated semi-mechanistic model variant progressed from a 12% bias at 1 week ahead to 20% (2 weeks), 30% (3 weeks), 40% (4 weeks) and 44% (5 weeks) overestimation. At the same time, this model showed rapidly decreasing sharpness as the forecast horizon increased (Fig 3C and and Table 2). This is reflected in the proper scoring rules that combine calibration and sharpness, with smaller values indicating better forecasts (Fig 3D and 3E and and Table 2). At 1-week ahead, the mean RPS values of the autoregressive, unfocused and best semi-mechanistic forecasting models were all around 30. At increasing forecasting horizon, the RPS of the semi-mechanistic model grew faster than the RPS of the autoregressive and unfocused null models. The DSS of the semi-mechanistic model, on the other hand, was very similar to the one of the autoregressive and better than that of the other null models at a forecast horizon of 1 week, with the autoregressive again performing best at increasing forecast horizons.
Focusing purely on the median forecast (and thus ignoring both calibration and sharpness), the absolute error (AE, Fig 3F and Table 2) was lowest (42) for the best semi-mechanistic model variant at 1-week ahead forecasts, although similar to the autoregressive and unfocused null models. With increasing forecasting horizon, the absolute error increased at a faster rate for the semi-mechanistic model than for the autoregressive and unfocused null models.
We lastly studied the calibration behaviour of the models over time; that is, using the data and forecasts available up to different time points during the epidemic (Fig 4). This shows that from very early on, not much changed in the ranking of the different semi-mechanistic model variants. Comparing the best semi-mechanistic forecasting model to the null models, again, for almost the whole duration of the epidemic calibration of the semi-mechanistic model was best for forecasts 1 or 2 weeks ahead.
Probabilistic forecasts aim to quantify the inherent uncertainty in predicting the future. In the context of infectious disease outbreaks, they allow the forecaster to go beyond merely providing the most likely future scenario and quantify how likely that scenario is to occur compared to other possible scenarios. While correctly quantifying uncertainty in predicted trajectories has not commonly been the focus in infectious disease forecasting, it can have enormous practical implications for public health planning. Especially during acute outbreaks, decisions are often made based on so-called “worst-case scenarios” and their likelihood of occurring. The ability to adequately assess the magnitude as well as the probability of such scenarios requires accuracy at the tails of the predictive distribution, in other words good calibration of the forecasts.
More generally, probabilistic forecasts need to be assessed using metrics that go beyond the simple difference between the central forecast and what really happened. Applying a suite of assessment methods to the forecasts we produced for Western Area, Sierra Leone, we found that probabilistic calibration of semi-mechanistic model variants varied, with the best ones showing good calibration for up to 2-3 weeks ahead, but performance deteriorated rapidly as the forecasting horizon increased. This reflects our lack of knowledge about the underlying processes shaping the epidemic at the time, from public health interventions by numerous national and international agencies to changes in individual and community behaviour. During the epidemic, we only published forecasts up to 3 weeks ahead, as longer forecasting horizons were not considered appropriate.
Our forecasts suffered from bias that worsened as the forecasting horizon expanded. Generally, the forecasts tended to overestimate the number of cases to be expected in the following weeks, as did most other forecasts generated during the outbreak [29]. This is in line with previous findings where our model was applied to predict simulated data of a hypothetical Ebola outbreak [38]. Log-transforming the transmission rate in order to ensure positivity skewed the underlying distribution and made very high values possible. Moreover, we did not model a trend in the transmission rate, whereas in reality transmission decreased over the course of the epidemic, probably due to a combination of factors ranging from better provision of isolation beds to increasing awareness of the outbreak and subsequent behavioural changes. While our model captured changes in the transmission rate in model fits, it did not forecast any trends such as the observed decrease over time. Capturing such trends in the attempt to identify underlying causes would be an important future improvement of real-time infectious disease models used for forecasting.
There are trade-offs between achieving good outcomes for the different forecast metrics we used. Deciding whether the best forecast is the best calibrated, the sharpest or the least biased, or some compromise between the three, is not a straightforward task. Our assessment of forecasts using separate metrics for probabilistic calibration, sharpness and bias highlights the underlying trade-offs. While the best calibrated semi-mechanistic model variant showed better calibration performance than the null models, this came at the expense of a decrease in the sharpness of forecasts. Comparing the models using the RPS alone, the semi-mechanistic model of best calibration performance would not necessarily have been chosen. Following the paradigm of maximising sharpness subject to calibration, we therefore recommend to treat probabilistic calibration as a prerequisite to the use of forecasts, in line with what has recently been suggested for post-processing of forecasts [56]. Probabilistic calibration is essential for making meaningful probabilistic statements (such as the chances of seeing the number of cases exceed a set threshold in the upcoming weeks) that enable realistic assessments of resource demand, the possible future course of the epidemic including worst-case scenarios, as well as the potential impact of public health measures. Beyond the formal test for uniformity of the PIT applied here, alternative ways of assessing calibration can be used [47, 57]. Once a subset of models has been selected in an attempt to discard miscalibrated models, other criteria such as the RPS or DSS can be used to select the best model for forecasts, or to generate weights for ensemble forecasts combining several models. Such ensemble forecasts have become a standard in weather forecasting [58] and have more recently shown promise for infectious disease forecasts [12, 59, 60].
Other models may have performed better than the ones presented here. Because we did not have access to data that would have allowed us to assess the importance of different transmission routes (burials, hospitals and the community) we relied on a relatively simple, flexible model. The deterministic SEIR model we used as a null model performed poorly on all forecasting scores, and failed to capture the downturn of the epidemic in Western Area. On the other hand, a well-calibrated mechanistic model that accounts for all relevant dynamic factors and external influences could, in principle, have been used to predict the behaviour of the epidemic reliably and precisely. Yet, lack of detailed data on transmission routes and risk factors precluded the parameterisation of such a model and are likely to do so again in future epidemics in resource-poor settings. Future work in this area will need to determine the main sources of forecasting error, whether structural, observational or parametric, as well as strategies to reduce such errors [32].
In practice, there might be considerations beyond performance when choosing a model for forecasting. Our model combined a mechanistic core (the SEIR model) with non-mechanistic variable elements. By using a flexible non-parametric form of the time-varying transmission rate, the model provided a good fit to the case series despite a high levels of uncertainty about the underlying process. Having a model with a mechanistic core came with the advantage of enabling the assessment of interventions just as with a traditional mechanistic model. For example, the impact of a vaccine could be modelled by moving individuals from the susceptible into the recovered compartment [36, 37]. At the same time, the model was flexible enough to visually fit a wide variety of time series, and this flexibility might mask underlying misspecifications. Whenever possible, the guiding principle in assessing real-time models and predictions for public health should be the quality of the recommended decisions based on the model results [61].
Epidemic forecasts played a prominent role in the response to and public awareness of the Ebola epidemic [28]. Forecasts have been used for vaccine trial planning against Zika virus [62] and will be called upon again to inform the response to the next emerging epidemic or pandemic threat. Recent advances in computational and statistical methods now make it possible to fit models in near-real time, as demonstrated by our weekly forecasts [35]. Such repeated forecasts are a prerequisite for the use of metrics that assess not only how close the predictions were to reality, but also how well uncertainty in the predictions has been quantified. An agreement on standards of forecast assessment is urgently needed in infectious disease epidemiology, and retrospective or even real-time assessment should become standard for epidemic forecasts to prove accuracy and improve end-user trust. The metrics we have used here or variations thereof could become measures of forecasting performance that are routinely used to evaluate and improve forecasts during epidemics.
For forecast assessment to happen in practice, evaluation strategies must be planned before the forecasts are generated. In order for such evaluation to be performed retrospectively, all forecasts as well as the data, code and models they were based on should be made public at the time, or at least preserved and decisions recorded for later analysis. We published weekly updated aggregate graphs and numbers during the Ebola epidemic, yet for full transparency it would have been preferable to allow individuals to download raw forecast data for further analysis.
If forecasts are not only produced but also evaluated in real time, this can give valuable insights into strengths, limitations, and reasonable time horizons. In our case, by tracking the performance of our forecasts, we would have noticed the poor calibration of the model variant chosen for the forecasts presented to the public, and instead selected better calibrated variants. At the same time, we did not store the predictive distribution samples for any area apart from Western Area in order to better use available storage space, and because we did not deem such storage valuable at the time. This has precluded a broader investigation of the performance of our forecasts.
Research into modelling and forecasting methodology and predictive performance at times during which there is no public health emergency should be part of pandemic preparedness activities. To facilitate this, outbreak data must be made available openly and rapidly. Where available, combination of multiple sources, such as epidemiological and genetic data, could increase predictive power. It is only on the basis of systematic and careful assessment of forecast performance during and after the event that predictive ability of computational models can be improved and lessons be learned to maximise their utility in future epidemics.
|
10.1371/journal.pbio.1000554 | Domain Swapping in Allosteric Modulation of DNA Specificity | SgrAI is a type IIF restriction endonuclease that cuts an unusually long recognition sequence and exhibits allosteric self-modulation of cleavage activity and sequence specificity. Previous studies have shown that DNA bound dimers of SgrAI oligomerize into an activated form with higher DNA cleavage rates, although previously determined crystal structures of SgrAI bound to DNA show only the DNA bound dimer. A new crystal structure of the type II restriction endonuclease SgrAI bound to DNA and Ca2+ is now presented, which shows the close association of two DNA bound SgrAI dimers. This tetrameric form is unlike those of the homologous enzymes Cfr10I and NgoMIV and is formed by the swapping of the amino-terminal 24 amino acid residues. Two mutations predicted to destabilize the swapped form of SgrAI, P27W and P27G, have been made and shown to eliminate both the oligomerization of the DNA bound SgrAI dimers as well as the allosteric stimulation of DNA cleavage by SgrAI. A mechanism involving domain swapping is proposed to explain the unusual allosteric properties of SgrAI via association of the domain swapped tetramer of SgrAI bound to DNA into higher order oligomers.
| Restriction endonucleases protect their bacterial hosts from viral infection by cleaving any invading viral DNA. One such enzyme, SgrAI, cleaves DNA very slowly but can be activated to cleave DNA 200 times more rapidly. Activation occurs when the enzyme interacts with two or more copies of DNA containing its recognition sequence. We have recently discovered that this enzyme forms polymers when activated. Polymerization may function to sequester the activated enzyme away from, and thereby protect, the host DNA. We have determined the three-dimensional structure of two SgrAI enzymes, each bound to their recognition sequence, interacting in a way that may occur in the polymer. The observed interaction involves a very unusual swapping of parts of each enzyme, termed domain swapping, which has rarely been found in enzyme activation. In support of the idea that this interaction operates during polymer formation and activation of SgrAI, we have shown that mutations designed to interfere with the interaction eliminate both activation and polymerization of the enzyme.
| Domain swapping involves the exchange of identical folding motifs between two copies of the same polypeptide chain [1], and as a result, a tight oligomer is formed. Such swapping has been found in many oligomeric proteins, where the swapped form is the biologically natural form [2] and in some cases where binding to a receptor brings two copies of a polypeptide together [3]. Domain swapping can in principle also lead to aggregation, as may occur in some amyloid diseases [4]. Clear cases of reversible domain swapping, serving to alter natural functions such as enzyme activity or specificity, are less well known.
Sequence-specific endonucleases capable of cleaving longer recognition sequences are highly sought for use in genomic work, as longer sequences occur less frequently and allow the manipulation of larger DNA fragments. The type IIF restriction endonuclease SgrAI cleaves a relatively long cognate, primary site sequence, CR|CCGGYG (R = A or G, Y = C or T, | denotes cut site) [5]. However, SgrAI also exhibits unusual biochemical properties; under certain conditions SgrAI cleaves plasmids bearing two copies of its recognition sequence faster than those bearing only a single site [6]–[8]. Further, SgrAI will also cleave at secondary sites containing the sequences CR|CCGGY(A,C,T) and CR|CCGGGG but only appreciably when the plasmid contains a primary site [9],[10]. Secondary sites are distinct from star sites, in that secondary sites are cleaved under solution conditions that are also optimal for cognate sequence cleavage. In contrast, star site sequences are cleaved appreciably only under special reaction conditions, such as high enzyme concentrations or the presence of organic solvents or Mn2+, and are discriminated against under optimal enzyme conditions by 2–4 orders of magnitude [11].
Type II restriction endonucleases typically bind and recognize palindromic sequences as dimers [11],[12], but the unusual biochemical properties exhibited by SgrAI suggest the formation of a higher order oligomer, containing altered enzymatic properties. For example, at low enzyme concentrations SgrAI cleaves plasmids bearing one or two sites at equal rates, but higher concentrations of enzyme result in the faster cleavage of the two site plasmid [7],[8]. This suggests that SgrAI forms a tetramer, or higher order species, on the DNA and cleaves two SgrAI sites in a concerted manner, similarly to that reported for the pseudodimer Sau3AI [13]. The event or process that leads to stimulation of DNA cleavage activity must occur through three-dimensional space, as the accelerated and concerted cleavage also occurs with plasmids each bearing a single site but connected by catenation [8]. Cleavage at primary site sequences in plasmids can also be stimulated by the addition of oligonucleotides containing the primary site sequence, intact or mimicking the cleavage products of SgrAI [8]–[10]. Analytical ultracentrifugation (AUC) shows that SgrAI exists as a dimer in the absence of DNA but forms both DNA bound dimers and high molecular mass aggregates in the presence of a 20 base pair DNA containing its recognition site [7]. The stoichiometry of this mixture of species has been determined by titration of DNA with SgrAI in AUC sedimentation velocity experiments showing one dimer of SgrAI per DNA duplex [7]. The DNA cleavage turnover number, kcat, of SgrAI with its cognate sequence shows a sigmoidal dependence on SgrAI concentration, consistent with the formation of an activated oligomer at the higher enzyme concentrations [7],[8]. The DNA cleavage rate also shows a sigmoidal dependence on DNA concentration, suggesting that DNA binding stimulates the formation of the activated conformation, which is presumably a tetramer or higher molecular weight species [10].
In addition to the stimulation of cleavage at cognate sequences, CR|CCGGYG, cleavage at secondary sites (CR|CCGGY(A,C,T) and CR|CCGGGG) by SgrAI can also be stimulated using the appropriate conditions. The cleavage at secondary sites is 200-fold slower relative to cognate [9], but this difference is reduced to only 10-fold when the secondary sites are adjacent to DNA ends simulating the products from cognate DNA cleavage [9]. The stimulation also involves an interaction in three dimensions, as stimulation of cleavage at secondary sites can be induced on a plasmid catenated to a plasmid containing the cognate sequence [8].
The related enzymes, Cfr10I [14],[15] and NgoMIV14, which cleave sequences R|CCGGY and G|CCGGC, respectively, form stable tetramers in both the presence and absence of DNA. Yet the crystal structures of SgrAI bound to cognate DNA CACCGGTG determined previously show only a dimer of SgrAI bound to a single duplex of DNA [16]. In this structure the central CCGG sequence is recognized by SgrAI using the same side chain-DNA contacts found in the structure of NgoMIV bound to DNA. The degenerate base pairs of the sequence, in the second and seventh positions (CRCCGGYG), appear to be recognized by indirect readout, and the outer base pair (CRCCGGYG) is recognized by a single contact from an arginine side chain to the G. However, these structures did not shed light on the mechanism of activation and sequence modulation seen in the SgrAI cleavage studies. We have recently shown that SgrAI forms oligomers of DNA bound dimers with primary site DNA at sufficient concentrations of enzyme and DNA [17]. These high molecular weight species (HMWS) occur under nearly identical conditions as the stimulation of SgrAI mediated primary and secondary site DNA cleavage. We have proposed that the HMWS are the activated form of SgrAI, or are at least a pre-requisite for the stabilization of the activated form. Here we present a new structure of SgrAI bound to DNA showing the close interaction of two DNA bound dimers. This tetrameric form is completely unlike those of Cfr10I or NgoMIV, as the tetrameric interface is at the opposite side of the dimer and is stabilized by swapping of the amino terminal 24 amino acid residues. A mechanism for the modulation of specificity is postulated and tested by analysis of the effects of mutations designed to destabilize the domain swapped tetramer. The mutant enzymes show complete loss of allosteric stimulation, as well as the inability to form HMWS.
The structure of SgrAI bound to DNA has been refined to 2.03 Å (Table 1) with Rcryst of 20.3% and Rfree of 25.2% and deposited in the RCSB Protein Data Bank with id 3MQ6. The structure shows the same global conformation of the protein as described previously [16], however the amino terminal 24 residues of each subunit appear to be swapped with a subunit of a neighboring dimer (space filling spheres, Figure 1A). The domain swapping, along with other contacts between the two dimers (Figure 1B), create a tetramer of SgrAI bound to two DNA duplexes. The crystallographic asymmetric unit contains two such tetramers. Residues 25–30 of each subunit comprise a linker, or hinge loop as it is referred in domain swapped structures, in the domain swapping, as these residues take on a different conformation in the swapped structures than in the previously described unswapped structures [16]. Figure 2 shows electron density in the vicinity of the swapped domains of subunits B and G, with the trace following a swapped (Figure 2A) or unswapped route (Figure 2B). Electron density for the hinge loop residues of subunits A, D, and E was poor and these residues could not be modeled.
The domain swapped tetramer of SgrAI is completely unlike the tetramers of NgoMIV or Cfr10I. The SgrAI dimers interact across the DNA binding face using the swapped amino terminal domains (Figure 1A shows the unswapped dimer of SgrAI, Figure 1B shows the domain swapped tetramer of SgrAI). In addition to the swapping interaction, approximately 400 Å2 is buried between non-swapping subunits of the tetramer from different dimers (for example, subunits B, salmon, and H, sand). Figure 1C shows a tetramer of NgoMIV bound to DNA, with the subunits of the top dimer (salmon, teal; Figure 1C) oriented as those of the bottom dimer of SgrAI (salmon, teal; Figure 1B), illustrating the different interfaces between dimers in the two tetramers. The contacts to the recognition sequence, CACCGGTG, by SgrAI are identical to those described previously [16]. The crystals were grown in a solution containing 50 mM CaCl2, and two Ca2+ are bound in each active site at positions found previously [16].
Analysis of the 2-fold axes in the two tetramers of the asymmetric unit reveals a slight difference, corresponding to a 2.5° rotation of one dimer relative to the other. The symmetry within each tetramer shows a small deviation from perfect 222 symmetry, where a total of three 2-fold axes occur orthogonal to one another. The deviation occurs in that the 2-fold axes that relate the two subunits of each dimer in the same tetramer are not coincident and instead are 10° apart (Figure 3A–B). This leads to a slight tilting of one dimer relative to the other in each tetramer (Figure 3B). A difference also exists in the hinge loops (residues 25–30) that connect the swapped domain (residues 1–24) with the rest of each subunit (31–339). In both tetramers, composed of A, B, G, and H in tetramer 1 and E, F, C, and D in tetramer 2, the hinge loops are better ordered in one swapped pair than the other. In tetramer 1 (Figure 3A–B), the hinge loops of swapping pairs B and G are relatively well ordered, while those of A and H are not, and residues 25–30 of subunit A could not be modeled. In tetramer 2, the hinge loops of swapping pairs C and F are well ordered, while those of D and E are not, and residues 25–30 could not be modeled in either subunit. In both tetramers, the better ordered hinge loops occur on the same face of the slightly asymmetric tetramer (Figure 3B). Although the electron density for the hinge loop of subunit A could not be modeled, the electron density for the hinge loop of subunit H allowed modeling in the swapped conformation, suggesting that swapping does occur between subunits A and H. In the absence of defined electron density for the hinge loops of subunits D and E, the possibility exists that these subunits are not domain swapped, yet it should be noted that the hinge loops of the unswapped forms do not show the same degree of disorder [16].
Only the conformations of residues 25–30 differ in the swapped and unswapped conformations of SgrAI (Figure 4); therefore, to test the role of the domain swapped form in the allosteric activity exhibited by SgrAI, two single site substitutions, P27G and P27W, were designed and prepared. First, as a control, the equilibrium dissociation binding constants (KD) of purified P27W and of P27G SgrAI were measured using a fluorescence polarization assay (FPA) and fluorophore labeled DNA (Table 2). The assays were performed at 4°C in buffer containing Ca2+ ions (20 mM Tris-OAc pH 8.0, 50 mM KOAc, 10 mM Ca(OAc)2, 1 mM DTT), which inhibit DNA cleavage while enhancing DNA binding affinity. All data fit well to a 1∶1 binding model without cooperativity. Wild type SgrAI binds to an 18 bp DNA containing a primary site sequence (18-1) with a KD of 0.6±0.2 nM [17]. The mutations P27G and P27W SgrAI weaken the affinity of SgrAI for this DNA, however the affinities are still quite tight (KD = 17±5 nM in the case of P27G, 4.0±0.8 nM in the case of P27W). Binding to PCP (precleaved primary site DNA) appears not to be diminished by any detectable amount by these mutations, with a KD of 6±2 nM for the wild type enzyme [17], compared to 5±2 nM and 9±2 nM for the P27G and P27W SgrAI enzymes, respectively. The binding affinity of the mutant enzymes to secondary site DNA (18-2) was not measured, since native PAGE (see below) indicated very weak binding (KD>1 µM).
Single turnover DNA cleavage rates were measured for P27W and P27G SgrAI with 18 base pair duplex oligonucleotides containing a primary site (18-1) sequence (Table 3). The assays were conducted at 37°C with 1 nM 32P labeled 18-1 DNA and 1 µM enzyme, with varying concentrations of added precleaved primary site DNA (PCP). They were also performed side-by-side with those for the wild type enzyme, with careful attention paid to the possible dissociation of PCP into single stranded DNA from repeated freeze thawing. The results show rate constants similar to wild type SgrAI in the absence of PCP (wild type: 0.094±0.015 min−1; P27G: 0.06±0.02 min−1; P27W: 0.037±0.005 min−1). However, while the wild type enzyme is stimulated >200-fold with 1 µM PCP, the mutant enzymes are hardly stimulated at all, 2–3-fold at best. The cleavage rate constants of secondary site DNA by the two mutant enzymes was not measured, as cleavage was undetectable likely due to very weak binding.
We have used native PAGE to separate different forms of enzyme bound DNA from free DNA and have found two different sizes of enzyme/DNA complexes [17]. The assay utilizes Ca2+ in the place of Mg2+ to facilitate tight DNA binding without cleavage, just as in the DNA binding affinity measurements, and is also performed at 4°C. We identify the faster migrating species as DNA bound dimer (DBD), and the slower as HMWS, composed of oligomers of DBD [17]. A titration using unlabeled PCP with 1 nM 32P labeled 18-1 or 18-2 and 1 µM wild type SgrAI indicates that HMWS forms appreciably at and above 100 nM PCP, and all DBD is shifted to HMWS at 1,000 nM PCP (Figure 5A). However, no HMWS are detected with either mutant enzyme, P27G or P27W SgrAI, under the same conditions (Figure 5B–C). In addition, the mutants appear to bind only weakly to 18-2. The lack of HMWS formation cannot be due to weak PCP binding, as both mutants bind PCP as tightly as wild type (Table 2). With P27W SgrAI, a slowly migrating band is found in lanes with low PCP concentration (Figure 5B) that disappears with higher PCP, hence showing behavior opposite to the HMWS formation by wild type SgrAI. This band may be a result of aggregation of additional SgrAI dimers on the 1∶1 SgrAI/DNA complex, as it disappears with additional PCP where more of the enzyme is expected to be bound to DNA.
Several structures of SgrAI bound to cognate (CACCGGTG) and noncognate (GACCGGTG) DNA, with Ca2+ or Mn2+, have been determined (Dunten et al. 2008 [16] and current work). In all, SgrAI forms a dimer very similar to those of Cfr10I and NgoMIV. Alignments of the structures show that SgrAI is more similar to Cfr10I, having some small deletions, and several insertions relative to Cfr10I [16]. NgoMIV and Cfr10I form tetramers in the crystal structures, with the tetrameric interface on the side of the dimer opposite to that of the DNA binding site (i.e. tail-to-tail). The new structure of SgrAI described here shows a tetramer that is unlike the NgoMIV and Cfr10I structures, with the tetrameric interface of SgrAI at the DNA binding face of the dimer (i.e. head-to-head) stabilized by the swapping of the amino-terminal 24 residues of each subunit (space filling spheres, Figure 1A). Residues 25–30 comprise the hinge loop that adopts a different conformation in the swapped form (Figures 2, 4). The SgrAI “swapping” domain is absent in NgoMIV and Cfr10I.
The biochemical data suggest that SgrAI can exist in at least two conformations, with one possessing an inherently greater DNA cleavage activity than the other. The observed stimulation of DNA cleavage activity could be accomplished by shifting the equilibrium from the low to the high activity form, possibly stabilized by higher order oligomers that favor the high activity conformation. The rate of DNA cleavage could be controlled by the positioning of groups in the active site, where the optimal alignment results in faster DNA cleavage kinetics. Analysis of the active sites of all SgrAI structures solved to date (Dunten et al. 2008 [16] and current work) shows very similar placement of all groups including the DNA in the various crystal structures, indicating that only a single conformation of the enzyme has been determined, which we have argued to be the low activity conformation [16].
To test the relevance of the domain swapped tetramer in the biochemical activity of SgrAI, two mutants were designed, P27G and P27W, predicted to destabilize the swapped conformation (Figure 4B), through introducing either increased flexibility with the glycine residue or steric conflicts with the large bulky tryptophan side chain. We found that both mutations disrupted the allosteric stimulation of DNA cleavage by SgrAI, without affecting the unstimulated DNA cleavage rate on the primary site sequence (Table 2), and without appreciably affecting binding affinity to uncleaved or precleaved primary site (Table 3). In addition, the activity of P27W SgrAI on plasmids containing either one or two primary site sequences shows that the presence of a second primary site does not appreciably accelerate DNA cleavage, as it does for the wild type enzyme (Text S1, Figures S1–S2). Further, the cleavage pattern of P27W SgrAI does not involve concerted cleavage of the two primary site sequences (Figure S2). Thus the plasmid assays also indicate that P27W SgrAI does not form the activated oligomer proposed to explain the fast, concerted cleavage by wild type SgrAI [6]–[8].
In addition to diminishing the ability of the SgrAI enzyme to be activated in DNA cleavage, the mutations were also found to eliminate the formation of HMWS under conditions where HMWS are formed by wild type enzyme (Figure 5). These results support our previous hypothesis that the HMWS is the activated form of SgrAI [17]. They also support a role for the interface between DBD seen in the structure of the domain swapped tetramer presented here in forming HMWS. Although species as small as tetramers are suggested by the accelerated cleavage of plasmids containing two primary site DNA sequences [6]–[8], our measurement of HMWS formed by wild type SgrAI and primary site containing DNA indicates species much larger than tetramers are formed [17]. Therefore if the tetramer found in the crystal structure is a building block of the HMWS, a second interface between the DNA bound dimers in addition to the domain swapped interface must exist, in order to form run-on oligomers of the size and heterogeneity seen in HMWS; an attractive possibility is the interface used by NgoMIV and Cfr10I (Figure 1B–C).
The effect of the mutations on binding to secondary site DNA was unexpected. Wild type SgrAI binds to both primary and secondary site DNA very tightly, with slightly tighter affinity (∼4-fold) for the primary sequence [17]. Therefore wild type SgrAI seems to discriminate very little between the two sequences at the binding level. Yet these single site substitutions, P27W and P27G, affect affinity very strongly for the secondary, but not the primary, sequence where the KD is shifted from nanomolar to micromolar. The origin of this effect is unknown and requires further investigation.
The SgrAI biochemical and structural data have some similarities to those of another type IIF restriction endonuclease, SfiI [18],[19]. SfiI is a tetramer in solution [19] that cleaves two copies of its recognition sequence in a concerted manner. The crystal structure of SfiI with its recognition sequence DNA show a tetrameric arrangement similar to that of NgoMIV, although the subunit structure is more like the dimeric BglI [18]. The conformation is identified to be in an inactive state since the DNA is mispositioned in the active site and only one of the predicted two divalent cations (Ca2+ or Mg2+ in the crystal structure) is bound. The low pH of the crystallization conditions may be responsible for the lack of the second divalent cation binding [18]. However, DNA cleavage data show that three recognition sites on the same DNA molecule are cleaved before enzyme dissociation, rather than the predicted two [20]. These data were interpreted as dissociation of one of the two sites cleaved concertedly followed by reassociation and cleavage of the third site prior to enzyme dissociation. Given the model for SgrAI, it is tempting to speculate whether SfiI is fully active also only in oligomers higher order than tetramers, explaining the concerted cleavage of three sites and the inactive conformation of the tetrameric species solved in the crystal structure. However, no direct evidence of oligomerization beyond tetrameric species has been reported for SfiI.
The allosteric communication network has been investigated in Bse634I, another type IIF endonuclease that bears very close structural similarity to SgrAI [21],[22]. Bse634I is a tetramer in solution and cleaves DNA fastest when both DNA binding sites are occupied with its recognition sequence. However, when only a single site is occupied, the DNA cleavage rate is reduced. Hence it possesses both auto-inhibition and stimulation capacities. While we have shown that DNA cleavage by SgrAI is stimulated (>200-fold, 4-fold greater than the 50-fold stimulation of Bse634I), it is not known if auto-inhibition also occurs. For auto-inhibition like that in Bse634I to occur, SgrAI dimers not bound to DNA would need to associate with DNA bound SgrAI dimers and decrease the DNA cleavage rate. Although SgrAI is dimeric in the absence of DNA binding [7], we have shown by gel shift measurements that oligomerization of the SgrAI dimers occurs significantly only with significant concentration of DNA bound dimers (i.e. above 100 nM) and not with excess SgrAI that is not bound to DNA [17]. However, the measurements of the stoichiometry of DNA binding by SgrAI performed with fluorescence anisotropy are suggestive of a second SgrAI dimer binding to the DNA bound SgrAI dimer. The single turnover DNA cleavage assays reported for SgrAI [17] have been done with a substantial excess of SgrAI over the DNA, and if a second SgrAI dimer (without bound DNA) binds to the DNA bound SgrAI dimer, then all reported rate constants have been performed with this additional dimer associated with the enzyme-DNA complex, and with any concomitant auto-inhibition. Investigation of auto-inhibition awaits measurements done with 1∶1 ratios of SgrAI and DNA.
To our knowledge, this is the first clear example of reversible domain swapping functioning to modulate the natural biological activity and specificity of an enzyme. Among previously reported examples [2], that of the RNase enzymes is strongest. RNase A, from bovine pancreas, forms oligomers during lyophilization in acetic acid [23],[24], and the dimers and trimers have been shown to be domain swapped [25]–[27]. Although the conditions for forming the oligomers are artificial, dimerization has been observed at pH 6.5 and 37°C with an equilibrium dissociation constant of 2 mM [28]. The enzymatic activities of the oligomers indicate that hydrolysis of double stranded RNA is faster with the oligomeric forms than with the monomeric, however virtually no difference is seen in the activities of the dimeric and monomeric species [29]. The related bovine seminal RNase, BS-RNase, exists as two interconverting dimers with only one stabilized by domain swapping. The enzyme exhibits cooperativity, but only at very high substrate concentrations (0.3 mM) and the effects are relatively small (1.2–1.3-fold) [30]. The domain swapped form is required for immunosuppression activity, but this activity is not a natural biological function of the enzyme [31]. Therefore, the potential use of domain swapping by SgrAI in a natural function of DNA cleavage rate stimulation and DNA sequence modulation may be the first clear case of a reversible domain swapping used to alter biological activity. This would also be the first case where DNA stimulates such domain swapping.
The unusual DNA cleavage activity of SgrAI may be a consequence of the large genome of Streptomyces griseus, from which it is derived. Restriction endonucleases are always coexpressed with a methyltransferase enzyme having the same sequence specificity, which functions to protect the host genome from the cleavage activity of the endonuclease. Hence the SgrAI methyltransferase must methylate all SgrAI recognition sequences within the genome before cleavage by the endonuclease can occur, and this requirement may be difficult due to the large size of the genome (over 8 million bp). The relatively long sequence recognized by SgrAI, 8 bp versus the usual 4–6, may have evolved due to this pressure, since the longer sequence greatly reduces the number of sites to be methylated in the host DNA. In addition, the inherently low cleavage activity of SgrAI in the absence of significant concentrations of unmethylated primary site DNA also reduces the pressure on host DNA, as well as the methyltransferase enzyme. However, such a long recognition sequence will also occur far less frequently in the phage DNA and hence place selective pressure on the enzyme for increased activity in order for adequate protection of the host from phage infection. It appears that one way in which the SgrAI enzyme activity is increased is through the stimulation of its cleavage activity with sufficient concentrations of unmethylated primary site DNA. Another way is through its secondary site cleavage activity, which will induce more cleavages in the phage DNA than at the primary sites alone, and hence could better protect the host. However, to protect against cleavage of the secondary sites in the host genome, the oligomerization may function to sequester activated SgrAI enzymes on the phage DNA and away from the host genome. It may also have an important role in sequestering the phage DNA itself or in rapidly communicating positive allosteric signals to multiple binding sites.
A modified USER-friendly DNA engineering method [32] was used to introduce P27W/G codon change into the sgrAIR gene. The original USER-friendly DNA mutagenesis technique employs two tail-to-tail overlapping primers, which prime template in the proximity of targeted mutation so that desired nucleotide changes can be incorporated into the primer sequences. The overlapping primers contain a single deoxyuracil (dU) residue flanking the overlap sequence on the 3′ side. After amplification, the dU is excised with the USER enzyme resulting in the PCR product flanked by complementary 3′ single-stranded extensions, which can reanneal to form a recombinant molecule26. Archaeal proofreading DNA polymerases are inhibited by dU in the primers; therefore the USER technique is compatible only with PfuTurbo Cx Hotstart DNA polymerase (Stratagene), which possesses a genetically modified uracil-binding pocket to overcome inhibition by dU [33]. (Taq DNA polymerase is not inhibited by dU, however it is not a proofreading polymerase.) Based on the structural organization of the uracil-binding pocket [33], we rationalized that 5-hydroxymethyluracil (5 hmU) would be prevented from entering the pocket due to the steric clashes with the 5-OH group. Therefore, 5 hmU could, in principle, be used in the primers for DNA amplification with archaeal proofreading DNA polymerases and afterwards be excised from PCR product by human SMUG1 DNA glycosylase, which is specific for 5 hmU [34].
Two overlapping primers, 5′ATGCGTGGGXGCGAAATCGTTCCAC and 5′ACCCACGCAXTTCGAATATCTTGGATGC, were used to introduce P27W (CCA→TGG) codon change into the sgrAIR gene. Likewise, the overlapping primers 5′ATGCGGGAGXGCGAAATCGTTCCAC and 5′ACTCCCGCAXTTCGAATATCTTGGATGC were used to introduce P27G (CCA→GGA) codon change. Each primer codes for the targeted codon change (underlined) and contains a single 5 hmU residue (marked as “X”) flanking the overlap sequence on the 3′ side (the overlap is shown in italic). The entire pET21a_SgrA1R plasmid was amplified as a 7548 bp linear fragment using Phusion DNA polymerase and the corresponding pair of overlapping primers. A 50 µl PCR reaction contained 10 ng of pET21a_SgrA1R template DNA, 0.2 mM dNTPs, 0.2 µM each primer, 3% DMSO, and 0.5 µl of Phusion Hotstart High-fidelity DNA polymerase (New England Biolabs). The pET21a_SgrAI was amplified for 30 cycles using cycling protocol as follows: initial denaturation is 30 s at 98°C; denaturation for 10 s at 98°C, annealing for 20 s at 65°C, polymerization for 4 min at 72°C; and final polymerization is 5 min at 72°C. After completion of the amplification reaction, a 5 µl PCR product aliquot was directly supplemented with 1 µl of 10X NEBuffer 1, 1 µl (20 units) of DpnI restriction endonuclease, and the reaction volume was adjusted to 10 µl with H2O. Restriction digestion was carried out for 1 h at 37°C and then reaction was incubated for 20 min at 80°C to inactivate DpnI. Ten units (1 µl) of EndoVIII DNA glycosylase and 10 units (2 µl) of SMUG1 DNA glycosylase (both enzymes from New England Biolabs) were added to the reaction and incubated for 15 min at 37°C to excise 5 hmU residues from PCR product, and then incubated an additional 15 min at room temperature to allow annealing of complementary extensions. Escherichia coli T7 Express Iq competent cells (New England Biolabs) were transformed with 5 µl of the annealing reaction. Recombinants were selected by plating 50 µl of transformation reaction on LB plates containing 0.1 mg/ml ampicilin. To confirm nucleotide sequence, plasmid DNA was purified from four individual recombinant colonies and sequenced across the sgrAIR gene. No sequence changes, except for the anticipated codon change, were observed.
Wild type and mutant SgrAI enzymes were prepared as described [16]. Briefly, the enzymes were expressed in E. coli strain ER2566 in the presence of the MspI methyltransferase (New England Biolabs). The enzymes were purified using FPLC (Pharmacia) chromatography and the following chromatographic resins: Heparin FF Sepharose (Pharmacia), SP FF Sepharose (Pharmacia), Q FF Sepharose (Pharmacia), and then a second Heparin FF Sepharose (Pharmacia) chromatographic step. Enzymes were dialyzed into storage buffer (20 mM Tris-OAc, pH 8.0, 50 mM KOAc, 0.1 mM EDTA, 1 mM DTT, 50% glycerol), aliquoted into small single use quantities, flash frozen in liquid nitrogen, and stored at −80°C until used.
The oligonucleotides (Figure 6) were made synthetically and purified using C18 reverse phase HPLC [35]. The concentration was measured spectrophotometrically, with an extinction coefficient calculated from standard values for the nucleotides [36], and fluorophore where appropriate. Fluorophore labeled DNA included with FLO (6-(3′,6′-dipivaloylfluoresceinyl-6-carboxamido)-hexyl group attached to the 5′ phosphate of the top strand only of PCP) or HEX (6-(4,7,2′,4′,5′,7′-hexachloro-(3′,6′-dipivaloylfluoresceinyl)-6-carboxamido)-hexyl group attached to the 5′ phosphate of both strands of 18-1) were obtained from a commercial synthetic source (Sigma Genosys) and contain a six carbon spacer between the fluorophore and the 5′ phosphate. The self-complementary DNA, or equimolar quantities of complementary DNA, were annealed by heating to 90°C for 10 min at a concentration of 1 mM, followed by slow-cooling to 4°C over 4–5 h in a thermocycler.
DNA used in the crystals have the self-complementary sequence 5′-AAGTCCACCGGTGGACT-3′, identical to 18-1 but one nucleotide shorter on the 3′ side leaving a 5′A overhang. Because freeze-thawing altered the concentration of double stranded DNA used in the assays, DNA used for stimulation of HMWS formation or in single turnover assays was treated very carefully to minimize this problem. Such DNA samples were either reannealed immediately prior to the assay or carefully annealed, assessed for concentration, aliquoted into small amounts, flash frozen in liquid nitrogen, stored at −20°C in water, and used only once after removing from the freezer. DNA was 5′ end labeled with 32P using T4 polynucleotide kinase (New England Biolabs) and [γ-32P]-ATP (Perlin-Elmer, Inc.), and excess ATP removed using G-30 spin columns (Biorad Laboratories, Inc.).
Crystals were prepared with SgrAI and DNA using 1.5 to 3.0 µl of the protein:DNA mixture with 1.0 to 1.5 µl of the precipitating solution (14% PEG 4K, 0.1 M Imidazole (pH 6.5), 0.15 M NaCl, 0.01 M NaNO3, 0.05 M CaCl2) per drop and placed over 1 ml of the precipitating solution. The SgrAI concentration varied between 10 and 30 mg/ml and was mixed with DNA to give a 1∶2 molar ratio of SgrAI dimer:DNA duplex. Crystals grow overnight to 1 wk at 17°C. The crystals were then exchanged into a cryoprotection solution (25% PEG 4K, 0.1 M Imidazole (pH 6.5), 0.3 M NaCl, and 30% glycerol) and flash-frozen in liquid nitrogen. X-ray diffraction was measured using synchrotron radiation at the Stanford Synchrotron Light Source (SSRL) BL9-2. Data collection was performed while maintaining the crystal at 100K. Image processing and data reduction were performed with HKL2000 (HKL Research, Inc.). The structure was solved by molecular replacement using PHASER [37],[38] and refined using CNS [39], PHENIX [40], REFMAC [41], and the model building program XtalView [42]. Symmetry relations between subunits were determined using LSQKAB [43] as found in CCP4 [44], and the alpha carbon atoms of residues 31–339 of each subunit. The 2Fo-Fc SA omit electron density map was calculated by first deletion of residues 23–31 from each subunit, then performing simulated annealing with a starting temperature of 2,000K in PHENIX [40]. All structure figures were prepared using PYMOL [45].
The equilibrium dissociation constant KD of SgrAI-DNA complexes was measured using a fluorescence polarization anisotropy technique [46]. DNA oligonucleotides (1 nM in 2 mL binding buffer: 20 mM Tris-OAc pH 8.0, 50 mM KOAc, 10 mM Ca(OAc)2, 1 mM DTT, 10% glycerol) containing a fluorophore (HEX or FLO) ligated to the 5′ end were titrated with increasing amounts of SgrAI enzyme (1 nM–1 µM), and the polarization recorded. Excitation occurred at 537 nm (HEX) or 495 nm (FLO) in a PC1 (ISS instrument) fluorimeter with T format, automatic polarizers, and temperature control. The emitted intensities were measured using a 50.8 mm diameter 570 nm cut-on filter with 580–2,750 nm transmittance range (ThermoOriel Inc., no. 59510) and 1 mm slit widths. The polarization of the emitted light as a function of added enzyme was fit to 1∶1 binding using Kaleidagraph software (Synergy Software) and the following [46]:where A is the polarization at a given protein concentration, Amax is the predicted polarization of fully bound DNA, Amin is the polarization with no protein binding, PT is the total concentration of protein, OT is the total concentration of the DNA, and KD is the dissociation constant to be determined.
Single turnover measurements of DNA cleavage were performed using chemical rapid quench techniques and 5′-end 32P labeled oligonucleotide substrates (typically 1 nM), under conditions of enzyme excess (1 µM), with and without the addition of unlabeled DNA. All reactions were performed at 37°C. For sampling by hand, 5 µl aliquots were withdrawn at specific time intervals after mixing the enzyme and labeled DNA (50 µl each), quenched by addition to 5 ul of quench (80% formamide, 50 mM EDTA), and electrophoresed on 20% denaturing polyacrylamide (19∶1 acrylamide:bisacrylamide, 4 M urea, 89 mM Tris, 89 mM boric acid, 2 mM EDTA) gels. Autoradiography of gels was performed without drying using a phosphor image plate, and exposing at 4°C for 12–17 h. Densitometry of phosphor image plate was performed with a Typhoon Scanner (GE Healthcare Life Sciences) and integration using ImageQuant (GE Healthcare Life Sciences) or ImageJ [47]. The percent of product formed as a function of time was determined by integrating both cleaved and uncleaved DNA bands. The single turnover DNA cleavage rate constant was determined from the data using a single exponential function:where C1 is a constant fitting the baseline, C2 is the total percent of DNA predicted to be cleaved by SgrAI, k is the rate constant, and t is the length of incubation in minutes.
Formation of HMWS was monitored in native gels (8% 29∶1 acrylamide:bisacrylamide in 89 mM Tris base, 89 mM boric acid, and 10 mM Ca2+). The electrophoresis running buffer was 89 mM Tris base, 89 mM boric acid, and 10 mM Ca2+ and was recirculated during electrophoresis. Gels were electrophoresed in a cold room (4°C) using 190 V. Gels were loaded while undergoing electrophoresis at 400 V, and the voltage returned to 190 V 5 min after the loading of the last sample. Electrophoresis was continued for an additional 2 h. Samples were prepared with 1 µM SgrAI, 1 nM 32P labeled DNA, and varied concentrations of unlabeled DNA in binding buffer (20 mM Tris-OAcpH 8.0, 50 mM KOAc, 10 mM Ca(OAc)2, 1 mM DTT, 10% glycerol) and incubated for 30 min at 4°C prior to electrophoresis. Autoradiography of gels was performed without drying using a phosphor image plate and exposing at 4°C for 12–17 h. Densitometry of phosphor image plate was performed with a Typhoon Scanner (GE Healthcare Life Sciences) and integration using ImageQuant (GE Healthcare Life Sciences) or ImageJ [47]. Integrated band intensities were normalized using the sum of the DNA bound species (DBD and HMWS) to determine the percent HMWS and then plotted versus PCP concentration using Kaleidagraph (Synergy Software).
|
10.1371/journal.ppat.1001274 | HTLV-1 bZIP Factor Induces T-Cell Lymphoma and Systemic Inflammation In Vivo | Human T-cell leukemia virus type 1 (HTLV-1) is the causal agent of a neoplastic disease of CD4+ T cells, adult T-cell leukemia (ATL), and inflammatory diseases including HTLV-1 associated myelopathy/tropical spastic paraparesis, dermatitis, and inflammatory lung diseases. ATL cells, which constitutively express CD25, resemble CD25+CD4+ regulatory T cells (Treg). Approximately 60% of ATL cases indeed harbor leukemic cells that express FoxP3, a key transcription factor for Treg cells. HTLV-1 encodes an antisense transcript, HTLV-1 bZIP factor (HBZ), which is expressed in all ATL cases. In this study, we show that transgenic expression of HBZ in CD4+ T cells induced T-cell lymphomas and systemic inflammation in mice, resembling diseases observed in HTLV-1 infected individuals. In HBZ-transgenic mice, CD4+Foxp3+ Treg cells and effector/memory CD4+ T cells increased in vivo. As a mechanism of increased Treg cells, HBZ expression directly induced Foxp3 gene transcription in T cells. The increased CD4+Foxp3+ Treg cells in HBZ transgenic mice were functionally impaired while their proliferation was enhanced. HBZ could physically interact with Foxp3 and NFAT, thereby impairing the suppressive function of Treg cells. Thus, the expression of HBZ in CD4+ T cells is a key mechanism of HTLV-1-induced neoplastic and inflammatory diseases.
| Human T-cell leukemia virus type 1 (HTLV-1) is the first retrovirus that is associated with human diseases including an aggressive leukemia derived from CD4+ T cells, adult T-cell leukemia (ATL), and chronic inflammatory diseases of the central nervous system, lung, or skin. However, it remains to be elucidated how HTLV-1 induces these diseases. A viral gene, tax, has been considered as a critical player in HTLV-1 pathogenesis, yet Tax expression is frequently lost in ATL cells. Another viral gene, HBZ, is constitutively expressed in both HTLV-1 infected cells and ATL cells. However, it remains unknown how HBZ functions in the HTLV-1-related diseases. We show here that the HBZ induced T-cell lymphoma and chronic inflammation in vivo similar to those in HTLV-1 infected individuals, indicating an important role of HBZ in HTLV-1 associated human diseases. As observed in HTLV-1 infected individuals, effector/memory and regulatory CD4+ T cells were increased in the HBZ-transgenic mice. Further, HBZ could interact with host transcription factors, Foxp3 and NFAT, leading to dysregulation of Treg function. The Treg dysregulation induced by HBZ is thought to be a critical mechanism of the HTLV-1 pathogenesis. This study sheds light on the HTLV-1 associated pathogenesis and provides an important clue to prevent or treat the human diseases.
| Human T-cell leukemia virus type 1 (HTLV-1) was the first human retrovirus associated with human diseases including adult T-cell leukemia (ATL) [1], [2] and HTLV-1 associated myelopathy/tropical spastic paraparesis (HAM/TSP)[3], [4]. One of the virological attributes of HTLV-1 is that it transmits mainly by cell-to-cell contact [5], [6]. Therefore, HTLV-1 induces the proliferation of infected CD4+ T cells to increase further transmission [7]. HTLV-1 encodes several regulatory and accessory genes in the pX region located between the env gene and the 3′ LTR [7], [8]. Among the viral genes, tax possesses in vitro transforming activity and can induce cancers in transgenic (Tg) animals via its pleiotropic actions [9], [10]. Yet the expression of Tax is frequently disrupted in ATL [7]. In contrast, the HTLV-1 bZIP factor (HBZ) gene, which is encoded in the minus strand of the HTLV-1 genome [11], [12], is transcribed in all ATL cases [13]. Recently, it has been reported that APOBEC3G generates nonsense mutations in all HTLV-1 genes except HBZ [14], suggesting that the HBZ gene is indispensable for the growth and/or survival of ATL cells and HTLV-1 infected cells. The HBZ gene product promotes the proliferation of ATL cells [13], [15]. Further, HBZ mRNA expression in HAM/TSP patients was well correlated with disease severity [16]. These findings suggest that HBZ has a critical role in the development of ATL and HAM/TSP.
It has been shown that ATL cells functionally and phenotypically resemble Foxp3+ CD25+CD4+ regulatory T (Treg) cells, which control immune responses against self- and non-self-antigen [17]. ATL cells constitutively express CD25 and scarcely produce interleukin-2 (IL-2)[18], [19]. Furthermore, two thirds of ATL cases harbor leukemic cells expressing FoxP3 [20], [21], a key transcription factor for the generation and function of Treg cells [22], [23], [24]. In HTLV-1 carriers, HTLV-1 provirus is detected mainly in CD4+ effector/memory T cells and Treg cells [25], [26], [27]. Thus, HTLV-1 favors Treg cells and effector/memory T cells in vivo, and transforms them. However, how HTLV-1 targets these T cell subpopulations remains to be elucidated.
In this study, we show that transgenic expression of HBZ increases Foxp3+ Treg cells and effector/memory T cells, leading to development of T-cell lymphomas and systemic inflammatory diseases. In addition, the suppressive function of Treg cells is severely impaired in HBZ transgenic mice. At the molecular level, we show that HBZ interacts with Foxp3 and NFAT, interrupting the function of each molecule, and leading to the deregulation of Foxp3-mediated transcriptional control of the genes associated with Treg functions. These results indicate that HBZ plays a critical role in neoplastic and inflammatory diseases arising from HTLV-1 infection.
Since HTLV-1 mainly infects CD4+ T cells in vivo, we generated Tg mice expressing the HBZ gene under the control of the murine CD4-specific promoter/enhancer/silencer (Figure S1) [13]. We analyzed the HBZ transgenes (Figure S1) and their expression in the three lines generated. HBZ gene expression was specifically detected in CD4+ T cells (Figure 1A). HBZ protein was also detected in these transgenic mice (Figure 1B). The level of HBZ gene transcripts in line 12 was the most abundant but similar to that of endogenous expression of the HBZ gene in ATL cell lines (Figure 1C). Therefore, unless specifically described, we used line 12 in this study. Notably, the majority of HBZ-Tg mice developed skin lesions by 18 weeks of age, in contrast with no disease in non-transgenic littermates (non-Tg mice) (Figure 1, D and E). Histological analyses revealed infiltration of CD3+CD4+ T cells into the dermis and epidermis, and also the alveolar septa of the lung (Figure 1, F, G and S2), whereas no obvious evidence of inflammation in other tissues, including liver, kidney, muscle, heart, stomach, spinal cord, intestines and brain. Since massive infiltration of lymphocytes in the skin and lung was observed in line 9 and 12, but not in line 2, level of HBZ expression is likely associated with these phenotypes. Thus, HBZ-Tg mice spontaneously developed dermatitis and alveolitis. Similar lesions have been observed in HTLV-1 carriers, especially in those harboring large numbers of infected cells [28], [29].
To study the growth-promoting activity of the HBZ gene, we assessed the proliferation of CD4+ T cells in HBZ-Tg mice by incorporation of bromodeoxyuridine (BrdU), and found that the proliferation was three fold-higher than in non-Tg mice, whereas the proliferation of CD8+ T cells or B cells was not altered (Figure 2A, Table S1A). Transgenic expression of HBZ enhances the in vivo proliferation of mouse T cells, as ectopic expression of HBZ enhances the proliferation of human T cells [13], [15]. It is known that HTLV-1 transforms CD4+ T cells after a long latent period in a fraction of asymptomatic carriers [7]. Analogous to the development of ATL in humans, 14 of 37 (37.8%) HBZ-Tg mice of all three-founder lines developed T-cell lymphomas after 16 months, in contrast with 2 of 27 non-Tg mice (7.4%) (P<0.001 by the logrank test) (Figure 2B). In some transgenic mice, lymphoma cells infiltrated various organs including the lung, bone marrow, spleen and liver (Figure 2C). All of the lymphomas in HBZ-Tg mice were CD3+ and CD4+ by immunohistochemical analyses when examined before the mice became moribund (Figure 2D). Lymphoma cells also expressed αβT cell receptors on their surfaces (Figure S3). Monoclonal proliferation of these lymphoma cells was shown by single strand conformation polymorphism in Vγ2-Jγ1 junction region of T cell receptor γchain gene (Figure S4). Notably, the primary lymphoma cells expressed Foxp3 at various intensities in the majority of cases (Figure 2E, Table 1), exhibiting a similar FoxP3 staining pattern to that in lymph nodes in human ATL cases (Figure S5). Thus, the T-cell lymphomas in HBZ-Tg mice phenotypically resemble ATL, suggesting that HBZ promotes proliferation of CD4+ T cells and predisposes expressing cells to transform in due course.
To study the cellular basis of the lymphomagenesis and inflammation in HBZ-Tg mice, we analyzed the phenotype and function of T cells, especially Treg cells, in 3-month-old HBZ-Tg line 12 mice before their pathological manifestations. CD44high CD62Llow effector/memory CD4+ T cells increased in HBZ-Tg mice (Figure 3A). CD4 single positive T cells also increased in the thymus (Figure S6). Further, not only the ratio but also the absolute number of Foxp3+ T cells was markedly increased in HBZ-Tg mice compared with non-Tg mice, while the numbers of Foxp3− T cells were equivalent (Figure 3, B and C). Increased Treg cells were also observed in thymus, lymph node and peripheral blood mononuclear cells (Figure 3D and Figure S7). We also observed the increased Treg cells and effector/memory T cells in the HBZ-Tg line 2 (Figure S8), which showed quite lower expression of HBZ than line 12 (Figure 1C). The proportion of Treg cells in skin and lung was rather low compared with that in spleen (Figure 3B and S2), indicating that Foxp3− T cells are predominant in the infiltrating T cells.
This result indicates that transgenic expression of HBZ induces systemic inflammation despite an increase in Foxp3+ Treg cells. It has been reported that IL-2 is critical in the homeostasis of Treg cells [30]. To study mechanisms by which HBZ expression increases Treg cells, we analyzed IL-2 production in the CD4+ T cells of HBZ-Tg mice after stimulation by PMA and ionomycin. IL-2 production was not augmented in either the Foxp3+ or Foxp3− populations from HBZ-Tg mice (Figure S9), indicating that the increase in the number of Treg cells was not due to enhanced IL-2 production.
Previous studies showed that Tax is a critical viral protein for the pathogenesis of HTLV-1. Therefore, we generated Tax transgenic (tax-Tg) mice using the same promoter/enhancer/silencer. In the tax-Tg mice, we did not observe increased effector/memory T cells or Treg cells (Figure S10). Thus, this increase in effector/memory T cells and Treg cells was specific to HBZ and not associated with similar transgenic expression of tax in this transgenic model system.
We next analyzed the phenotype and function of the increased Foxp3+ Treg cells in HBZ-Tg mice. CD4+Foxp3+ T cells of HBZ-Tg mice expressed Treg-associated molecules, such as cytotoxic T-lymphocyte associated antigen-4 (CTLA-4), glucocorticoid-induced TNF receptor family-related-protein (GITR), CD103, and CD25 [31]; yet the expression levels of CTLA-4, GITR and CD25 were lower than those of Foxp3+ T cells in non-Tg mice (Figure 3, B and E, Table S1B). In contrast, both Foxp3+ and Foxp3− CD4+ T cells of HBZ-Tg mice expressed CCR4 and CD103 at higher levels than those in non-Tg mice, suggesting that this might contribute to the migration and infiltration of HBZ-Tg CD4+ T cells into the skin (Figure 1F) [32], [33]. Further, it is of note that the in vitro suppressive function of HBZ-Tg Treg cells was severely impaired. When CD4+GITRhigh T cells, which were >90% Foxp3+ [23], from HBZ-Tg or non-Tg mice were co-cultured with CD4+CD25− T cells from wild-type mice and stimulated with Con A or anti-CD3 antibody, HBZ-Tg Treg cells were much less suppressive (Figure 3F). These results indicate that HBZ expression increases functionally impaired Treg cells.
Next, we assessed the proliferation of CD4+ T cells in HBZ-Tg mice. BrdU incorporation of Foxp3+ as well as Foxp3−CD4+ T cells from HBZ-Tg mice was also significantly higher than those in non-Tg mice (Figure 3G). In general, proliferation of Treg cells in response to mitogenic stimulation is suppressed in vitro. However, Foxp3+ T cells from HBZ-Tg mice proliferated more vigorously in vitro in response to anti-CD3 antibody than did non-Tg Foxp3+ T cells (Figure 3H). Thus, transgenic expression of HBZ in CD4+ T cells induces the expansion of Foxp3+ Treg cells, yet impairs their suppressive function.
To study whether HBZ increases Foxp3+ Treg cells in a cell intrinsic manner, we expressed HBZ in naive CD4+ T cells in vitro using a retrovirus vector (Figure 4A). Interestingly, HBZ induced Foxp3 expression in 16.8% of HBZ expressing T cells, which is a similar enhancement to that due to TGF-β treatment (14.8%). The expression was markedly augmented in HBZ expressing T cells treated with TGF-β (72.2%) (Figure 4B). A reporter assay using the enhancer and promoter of the Foxp3 gene [34] demonstrated that HBZ induced transcription of the Foxp3 gene (Figure 4C), which was enhanced in the presence of TGF-β. Thus, HBZ-induced Foxp3 expression could be a mechanism for the increase of Foxp3+ T cells in HBZ-Tg mice.
Previous studies have shown that Foxp3 controls Treg function by cooperating with transcription factors including NFAT [35] and AML-1/Runx1[36]. Impaired interactions of Foxp3 with these factors not only alter the suppressive function of Treg cells but also suppress the expression of Treg associated molecules, such as CD25, CTLA-4, and GITR [23], [35], [36], [37], which is similar to the phenotype observed in HBZ-Tg mice (Figure 3, B and E). These findings prompted us to assess the possibility that HBZ might be involved in Foxp3-dependent transcriptional regulation. To address this, we first examined direct interaction among HBZ, NFAT and Foxp3. Immunoprecipitation experiments showed that HBZ physically interacted with both NFAT and Foxp3 (Figure 5A). Moreover, to study the interaction of endogenous HBZ and Foxp3, we performed immunoprecipitation using ATL-43T, a Foxp3-expressing ATL cell line. An anti-HBZ antibody co-precipitated endogenous Foxp3 in the ATL-43T cells, demonstrating that the interaction occurs not only in an enforced over-expressed state but also under physiological conditions (Figure 5B). It has been previously reported that human FoxP3 protein migrates as a doublet, which coincides with this result [38]. Analyses using HBZ deletion mutants showed that the central domain of HBZ interacted with Foxp3 (Figure 5C). Experiments with Foxp3 deletion mutants revealed that HBZ interacted with the forkhead (FH) domain of Foxp3 (Figure 5D). It has been reported that the region between the forkhead domain and the leucine zipper domain of Foxp3 interacted with AML-1 [36]. HBZ did not inhibit the binding between Foxp3 and AML-1 nor the suppressive effect of Foxp3 on AML-1-mediated transcription from the IL-2 gene promoter (Figure S11), indicating that HBZ does not influence Foxp3/AML1 mediated gene regulation.
To study whether HBZ independently interacts with Foxp3 and NFAT or, alternatively, if these molecules form a ternary complex, we studied the effect of the DNA intercalator ethidium bromide (EtBr) on their interactions. As shown in Figure 5E, the interactions of HBZ with Foxp3 or NFAT were not affected by EtBr while the interaction between NFAT and Foxp3 was diminished by EtBr as reported previously [35]. These findings suggest that the interactions of HBZ with NFAT and Foxp3 are independent of DNA while the interaction between NFAT and Foxp3 requires the presence of DNA.
In HBZ-Tg mice, the expression of Treg-associated molecules including CTLA-4, GITR and CD25 was suppressed when compared with their expression in Treg cells from non-Tg mice (Figure 3B and E). This finding may account for the impaired function of Treg cells since these molecules, in particular CTLA-4, play a critical role in Treg-mediated suppression [39]. To further study the effect of HBZ on the expression of Treg-associated molecules, we transduced HBZ along with Foxp3 into naive CD4+ T cells in vitro using retrovirus vectors (Figure 4A). HBZ expression suppressed Foxp3-induced GITR and CTLA-4 expression whereas it did not inhibit CD25 expression (Figure 6A). Expression of HBZ alone increased CD25 expression (Figure 6A), which might obscure the suppressive effect of HBZ under these conditions. Suppression of GITR and CTLA-4 expression required both the activation and the central domains of HBZ (Figure 6, B and C), which correspond to the binding sites of HBZ to Foxp3 (Figure 5C) and NFAT (Figure S12). Since both Foxp3 and NFAT are critical for Treg function [35], it is likely that HBZ suppresses the expression of GITR and CTLA-4 by interacting with Foxp3 and NFAT and thereby interfering with their transcriptional regulation in Treg cells. To examine suppressive effect of HBZ on expression of GITR, CTLA-4 and CD25, we isolated Treg cells from wild type mice and expressed HBZ using retroviral vectors. As shown in Figure 6D, HBZ suppressed endogenous expression of CD25, GITR and CTLA-4 in Treg cells, confirming that HBZ is responsible for suppressed expression of these molecules.
HTLV-1 targets CD4+ T cells; cell central to immune regulation. In contrast to human immunodeficiency virus, which destroys CD4+ T cells, HTLV-1 increases its copy number by inducing clonal proliferation of infected cells in vivo [40], [41]. Since HTLV-1 spreads mainly by cell-to-cell transmission [5], increased number of infected cells facilitates transmission of HTLV-1 to new cells. Recent studies showed that glucose transporter 1, heparan sulfate proteoglycans and neuropilin-1 are important for the entry of HTLV-1[42], [43], [44], consistent with the finding that this virus can infect a variety of cell types [45], [46]. However, HTLV-1 provirus is detected mainly in the regulatory and effector/memory CD4+ T cells of HTLV-1 carriers (Figure S13) [25], [26], [27], which indicates that HTLV-1 favors these specific subpopulations of CD4+ T cells. These findings suggest that HTLV-1 induces proliferation and/or facilitates survival of the regulatory and effector/memory CD4+ T cells. The mechanism(s) by which HTLV-1 targets Treg cells, however, remained unclear until now. In this study, we showed that HBZ could enhance transcription of the Foxp3 gene, and also promote proliferation of Foxp3+CD4+ T cells in transgenic mice, indicating that HBZ enhances both the generation and proliferation of Foxp3+ T cells. Impaired Foxp3 function is associated with proliferation of Treg cells [37], so the HBZ-mediated Treg dysfunction may also contribute to Treg proliferation in addition to direct growth proliferation by the HBZ transcript [13]. Another possible explanation is that Treg cells might be more susceptible to HTLV-1 infection, since Treg cells proliferate vigorously in vivo presumably by recognizing self-antigen and commensal microbes [47]. With these strategies, HTLV-1 likely targets this specific T-cell population as its host, which might be beneficial for their survival.
As mechanisms of the HBZ-mediated effect on Foxp3 functions, we demonstrated that HBZ physically interacted with Foxp3 and impaired its function in vitro. HBZ lacking the Foxp3-binding region showed a slight inhibitory effect on Foxp3 function, indicating that direct interaction between HBZ and Foxp3 is, at least in part, responsible for suppression. The results of immunoprecipitation analyses using Foxp3 mutants showed that the forkhead domain of Foxp3 was responsible for the molecular interaction between HBZ and Foxp3. Since the forkhead domain is the DNA-binding domain of Foxp3 [17], HBZ might inhibit the transcriptional function of Foxp3 by interfering with the DNA binding activity. Foxp3 play a key role in the function and homeostasis of Treg cells [22], [23], [24], indicating that HBZ-mediated dysfunction of Foxp3 contributes to impaired Treg function in HBZ-Tg mice. This impaired Treg function allows non-regulatory T cells to become hyper-reactive to commensal microbes and self-antigens, provoking enhanced proliferation of non-regulatory T cells and T cell-mediated autoimmune/inflammatory disease. These data collectively suggest that the viral protein HBZ hijacks the transcriptional machinery of host Treg cells leading to inflammatory disorders in the host. Conversely, Tax, another HTLV-1 protein, has been reported to suppress FoxP3 expression in human T cells in vitro [48]. Therefore, it is likely that both viral proteins target Foxp3 albeit with apparently different effects. Considering that HBZ is consistently expressed while Tax expression is sporadic, Tax might control excess expression of Foxp3 in HTLV-1 infected cells.
In this study, we demonstrated that the characteristics of CD4+ T cells in HBZ-Tg mice resemble those of human ATL cells or HTLV-1 infected cells in carriers. First, the frequency of Foxp3 positive cells in T-cell lymphomas was similar in HBZ-Tg mice and in ATL [20]. Second, the suppressive function of Foxp3+ T cells was impaired in both ATL and HBZ-Tg mice [49]. Third, CD4+ T cells in HBZ-Tg mice, HTLV-1-infected cells in carriers, and ATL cells possess similar effector/memory and regulatory phenotypes [25], [27]. As shown in this study, transgenic mice expressing Tax under the same promoter as the HBZ-Tg mice did not show any changes in the number of Foxp3+ Treg cells or effector/memory T cells. These data suggest that HBZ, rather than Tax, is responsible for conferring the specific phenotype of HTLV-1 infected cells and ATL cells.
It has been reported that tax transgenic animals develop tumors [50], [51], [52]. In these reports, Tax induced tumors, the type of which depends on the promoter used. However, irrespective of the possible oncogenic activity of Tax, leukemic cells in ATL patients frequently lose Tax expression [7], whereas HBZ expression has been detected in all ATL cases studied so far [13]. We reported that the HBZ gene transcript itself has growth-promoting activity in vitro [13]. Taken together, our results suggest that HBZ is responsible for the specific phenotype, function and proliferation of HTLV-1-infected CD4+ T cells and ATL cells, and that HBZ plays important roles for the oncogenic activity of HTLV-1 in addition to Tax. Further, the long latent period before the onset of T-cell lymphomas in HBZ-Tg mice suggests that additional genetic and/or epigenetic alterations in CD4+ T cells are necessary for the development of T-cell lymphomas in HBZ-Tg mice as well as for ATL.
In conclusion, the HBZ-mediated dysregulation of Treg function and proliferation that we report here provides novel insights into the interaction between the host and the virus and may be exploited to treat and prevent HTLV-1-induced diseases.
This study was conducted according to the principles expressed in the Declaration of Helsinki. The study was approved by the Institutional Review Board of Kyoto University (E921). All patients provided written informed consent for the collection of samples and subsequent analysis. Animal experimentation was performed in strict accordance with the Japanese animal welfare bodies (Law No. 105 dated 19 October 1973 modified on 2 June 2006), and the Regulation on Animal Experimentation at Kyoto University. The protocol was approved by the Institutional Animal Research Committee of Kyoto University (Permit Number: D09-3). All efforts were made to minimize suffering.
C57BL/6J mice were purchased from CLEA Japan. The HBZ cDNA was cloned into the SalI site of the H/M/T-CD4 vector, which was designed for restricted expression of a transgene in CD4+ cells. The purified fragment containing the HBZ transgene was microinjected into C57BL/6J F1 fertilized eggs. Transgenic founders were screened for the integration of transgenes in their genomic DNA by PCR and mated with C57BL/6J mice to generate transgenic progeny [13], [15]. All HBZ-Tg mice were heterozygotes for the transgene. The phenotype of HBZ-Tg mice was stable in the different generations. They express the spliced HBZ gene under the control of the CD4-specific promoter/enhancer/silencer. All mice were used at 10-16 weeks of age unless specifically described.
The human embryonic kidney cell line, 293FT, was cultured in DMEM containing 10% FCS and G418 (500 µg/ml). The 293FT cell line is derived from the 293F cell line and stably expresses the SV40 large T antigen. 293FT cell line was purchased from Invitrogen. The packaging cell line, Plat-E (kindly provided by T. Kitamura, Tokyo University) was cultured in DMEM supplemented with 10% FCS containing 10 µg/ml blasticidin and 1 µg/ml puromycin. ATL-43T(−) (kindly provided by M. Maeda, Kyoto University) and MT-1 cells were derived from ATL cells, and cultured in RPMI containing 10% FCS and antibiotics (penicillin and streptomycin). A mouse T-cell lymphoma line, EL4 cells, were cultured with RPMI1640 containing 10% FCS, antibiotics, and 50 µM 2-mercaptoethanol (2-ME; Invitrogen).
In order to construct the vectors expressing tagged spliced HBZ and its mutants, their coding sequences were amplified by PCR, and cloned into the expression vector, pcDNA 3.1(−)/myc-His (Invitrogen). A cDNA clone that contains NFATc2 coding sequence was kindly provided by Kazusa DNA Research Institute. To construct the FLAG-tagged NFATc2 expression vector, its coding region was cloned into pCMV-Tag2 (Stratagene). pCMV-HA (Clontech) was used to generate HA-tagged Foxp3 expression vectors. The vectors expressing Flag-tagged Foxp3 mutants were also used for immunoprecipitation.
The following antibodies were used for immunoprecipitation and Western blotting: mouse anti-Flag (clone M2; Sigma, Saint Louis, MO), mouse anti-c-myc (clone 9E10; Sigma), mouse anti-HA (clone HA-7; Sigma), rabbit anti-His polyclonal antibody (MBL), rabbit anti-FOXP3 (polyclonal antibody; Abcam), and rabbit anti-HBZ polyclonal antisera [15].
The following antibodies were purchased from BD PharMingen; purified monoclonal antibody (mAb) for mouse CD4 (RM4-5), CD8α (53-6.7), CD25 (PC61), CD44 (IM7), CD103 (M290), and IL-2 (JES6-5H4). Purified monoclonal antibodies for mouse GITR (DTA-1), CTLA-4 (UC10-4B9), CD62L (MEL-14), TCRβ (H57-597), TCRγδ (eBioGL3) and Foxp3 (FJK-16s) or human FoxP3 (236A/E7) were purchased from eBioscience. Anti-mouse CCR4 antibody (polyclonal antibody; Capralogics) and FITC-labeled anti-goat IgG antibody (Santa Cruz Biotechnology) were used for the detection of mouse CCR4. The following reagents were used for cell culture: anti-CD3ε antibody (145-2C11; R&D systems), Con A (Sigma), PMA (Sigma), and ionomycin (Sigma).
cDNAs were synthesized from 1 µg total RNA of purified mouse CD4+ T cells by a reverse transcriptase SuperScript III and random primers according to the manufacturer's instructions (Invitrogen). Spliced HBZ and GAPDH transcripts were quantified using RT-PCR. The primers used were as follows: sHBZ gene: 5′-TAAACTTACCTAGACGGCGG-3′ (sense), 5′-CTGCCGATCACGATGCGTTT -3′ (antisense); GAPDH gene: 5′-GTGGAGA TTGTTGCCATCAACG -3′ (sense) and 5′-AGAGGGGCCATCCACAGTCTT-3′ (antisense). PCR was performed in a PC-808 (Astec) under the following conditions: HBZ: 2 minutes at 95°C, followed by 26 cycles of 30 seconds at 95°C, 30 seconds at 59°C and 60 seconds at 72°C; GAPDH: 3 minutes at 95°C, followed by 35 cycles of 30 seconds at 95°C, 30 seconds at 61°C and 30 seconds at 72°C.
To quantify the expression level of HBZ, a TaqMan probe and primers for HBZ were designed. The sequences of primers and probe for HBZ were as follows; HBZ primers; 5′-GGACGCAGTTCAGGAGGCAC-3′ (sense) and 5′-CCTCCAAGGATAATAGCCCG-3′ (antisense); HBZ probe; 5′-CCTGTGCCATGCCCGGAGGACCTGC-3′. We used the TaqMan Gene expression Assay for 18S rRNA (Applied Biosystems) as an internal control. Relative expression level of HBZ or IL-2 was calculated with the delta delta Ct method.
For retroviral gene transduction experiments, spliced HBZ cDNA was cloned into a retroviral vector, pMXs-Ig (a gift from T. Kitamura), to generate pMXs-Ig-HBZ. pGCSamIN (kindly provided from M. Onodera) and pGCSamIN-Foxp3 were used as previously described. Transfection of the packaging cell line, Plat-E, was performed as described. For retroviral transduction, CD25−CD4+ cells were enriched by a CD4 enrichment kit (BD Pharmingen) and were activated by 0.5 µg/ml anti-CD3 Ab and 50 U/ml rIL-2 in the presence of T-cell-depleted and x-irradiated (20Gy) C57BL/6J splenocytes as APCs in 12 well plates. After 16 hours, activated T cells were transduced with viral supernatant and 4 µg/ml polybrene, and centrifuged at 3,000 rpm for 60 min. Cells were cultured in medium supplemented with 50 U/ml rIL-2. Activation of naïve T cells by anti-CD3 antibody influenced expression of these molecules. Therefore, we analyzed their expression after influence by activation was lost [35]. Two days later, Foxp3-mediated CTLA-4 expression was detected by a flow cytometry, and five days later, expression of GITR or CD25 was analyzed. After two days, we stimulated the transduced cells with 50 ng/ml PMA and 1 µg/ml ionomycin in the presence of protein transport inhibitor (BD PharMingen) for 6 hours, and then analyzed intracellular IL-2 expression using intracellular cytokine staining kits (BD Pharmingen) according to the manufacturer's instructions.
To elucidate the effect of HBZ on endogenous expression of Treg associated molecules, we transduced HBZ into CD4+Foxp3+ cells purified from mouse splenocytes. Three days after transduction, the expression levels of Treg associated molecules were evaluated by a flow cytometry.
Cell suspensions were prepared from murine spleens by forcing the organs through a nylon mesh, and splenic erythrocytes were eliminated with NH4Cl. Proliferation of murine cells was measured by 3H-thymidine uptake after 3 days of incubation in RPMI1640 medium supplemented with 10% FCS and 50 µM 2-ME. Flow cytometric analyses and cell sorting were carried out using a FACS CantoII or FACS Aria with Diva Software (BD Pharmingen) and the data was analyzed by FlowJo software (Treestar). For cell surface staining, 106 cells were incubated with mAbs for 30 min at 4°C, and then analyzed. For intracellular staining, we used a mouse Foxp3 staining kit according to its protocol (eBioscience). To sort Foxp3+ cells, suspended splenocytes were stained with mAb for CD4 and GITR, and the CD4+GITRhigh fraction was sorted by FACS Aria. Purity of the sorted population was always >90% by re-analysis of Foxp3 staining. For the ex vivo proliferation assay of Foxp3+ cells, carboxy-fluorescein diacetate, succinimidyl ester (CFSE)(Molecular Probe) was used according to the manufacturer's instructions. Foxp3+ T cells (2×104/well) were stimulated with anti-CD3 antibody (4 µg/ml) in round-bottomed 96-well plates in the presence of x-irradiated splenocytes as antigen presenting cells (APC; 5×104/well) for 96 hours. Then, cells were permeabilized, and stained with anti-Foxp3. CFSE dilution was analyzed by flow cytometry. To evaluate the suppressive activity of Foxp3+ T cells sorted from HBZ-Tg or non-Tg mice, Foxp3+ T cells (2×104/well) were cultured with CD25−CD4+ cells (2×104/well) and APCs (5×104/well) from wild-type mice for 72 h in the presence of soluble anti-CD3 (4 µg/ml) or Con A (1 µg/ml), and then [3H] thymidine incorporation was measured.
In vivo proliferation was measured by BrdU incorporation. BrdU (Nacalai Tesque) was dissolved in PBS (3 µg/ml), and then 200 µl was injected intraperitoneally into HBZ-Tg and non-transgenic mice twice a day for three days as reported previously [53]. BrdU incorporation in CD4+, CD8+, or B220+ splenocytes was detected using FITC BrdU Flow Kits (BD Pharmingen) according to the manufacturer's instructions. Flow cytometric analyses were performed on a FACS CantoII with Diva Software (BD Pharmingen).
We constructed Foxp3 promoter and enhancer reporter plasmids as the previous report [34]. A murine T-cell line, EL4 cells (1×107), were transiently cotransfected by electroporation with the following plasmid DNAs: Foxp3 reporter plasmid, Renilla luciferase control vector (pRL-TK), and HBZ expression vector (pME18SneoHBZ). Cells were cultured with or without TGF-β (2 ng/ml). Firefly and Renilla luciferase activities were measured using the Dual-Luciferase Reporter Assay System (Promega). Relative luciferase activities were calculated as the ratio of firefly and Renilla luciferase activities. The luciferase values are shown as relative values. Values represent means plus standard deviations (error bars) (n = 3).
The study of clinical samples was approved by the local research ethics committee of the appropriate hospital. Tissue samples were fixed in 10% formalin in phosphate buffer and then embedded in paraffin. Haematoxylin and eosin (H&E) staining was performed according to standard procedures. Images were captured using a Provis AX80 microscope (Olympus) equipped with OLYMPUS DP70 digital camera, and detected using a DP manager system (Olympus).
For analysis of tumors, mice that became immobilized were sacrificed and subjected to autopsy. Tissue samples were surgically removed and fixed in 10% formalin in phosphate buffer and embedded in paraffin. Sections were stained with H&E for histopathologic examination. After we obtained informed consent, tissue samples from patients who were diagnosed as lymphoma-type ATL were analyzed by immunohistochemical methods to determine FoxP3 expression. Monoclonal antibodies for CD3ε(500A2; BD Pharmingen), B220 (RA3-6B2; BD Pharmingen), and Foxp3 (FJK-16s; eBioscience) were used for immunohistochemistry. We judged CD3+B220+ cases to be T-cell lymphomas since some activated T cells and T cells of the lpr/lpr mutant mouse expressed B220 [54], [55].
To investigate clonality of lymphoma cells observed in HBZ-Tg mice, lymphoma tissue samples of HBZ-Tg were analyzed for the clonality of T-cell receptor (TCR) γ locus using PCR-SSCP analysis of the TCR γ-gene. Genomic DNA was subjected to PCR amplification using primers for the Vγ2 gene and the Jγ1. The primers used were as follows: Vγ2: 5′-ACCAAGAGATGAGACTGCACAA-3′ (sense), Jγ1: 5′-GCGTCTGATCCTCAAAATAACTTCC-3′ (antisense); PCR was performed in a PC-808 (Astec) under the following conditions: 3 minutes at 95°C, followed by 35 cycles of 30 seconds at 95°C, 30 seconds at 55°C and 30 seconds at 72°C. We used EL-4 as a positive control and splenic DNA from young non-Tg or HBZ-Tg mice as negative control. PCR products were run on a 6% polyacrylamide gel and visualized by staining with DNA Silver Staining Kit (GE Healthcare).
Expression vectors for the relevant genes were transiently cotransfected into 293FT cells using the TransIT-LT1 reagent (Mirus Bio). 24 hours later, transfected cells were stimulated with 50 ng/ml PMA and 1 µg/ml ionomycin for another 6 hours. Coimmunoprecipitation assays were performed using the Nuclear Complex Co-IP Kit (Active motif). Briefly, the nuclear extracts of transfected cells were prepared in the presence or absence of ethidium bromide (10 µg/ml). They were precleared with Protein G Sepharose 4 Fast Flow (GE Healthcare), and their supernatants were incubated with anti-myc tag (clone 9E10, Sigma) or anti-Flag tag (M2, Sigma) antibody overnight at 4°C. The immunocomplexes were precipitated with Protein G Sepharose 4 Fast Flow, fractionated in SDS-PAGE, and transferred to PVDF membranes. HBZ-myc-His was detected with horseradish peroxidase (HRP)-conjugated anti-His tag (MBL) antibody. HRP-conjugated anti-Flag tag (Sigma) and anti-HA tag (Sigma) antibodies were used to detect Flag-tagged and HA-tagged proteins, respectively. To detect endogenous interaction between HBZ and FoxP3, immunoprecipitation was performed using an ATL cell line, ATL-43T(-), as described above with anti-HBZ antisera and anti-FOXP3 antibody (Abcam). To examine the expression of HBZ in transgenic mice, CD4+ splenocytes from wild type or HBZ-Tg mice were enriched by a mouse CD4 T lymphocyte enrichment set (Pharmingen). Whole cell extracts were prepared with the lysis buffer (50 mM Tris-HCL, PH 7.5, 150 mM NaCl, 1% NP-40), and analyzed by western blotting with anti-HBZ antisera.
A previous report demonstrated that Tax expression could not be detected in freshly isolated PBMC from HTLV-1 infected carriers but could be detected when they were cultivated ex vivo for 12 hours [56]. We cultured PBMCs from asymptomatic HTLV-1 carriers for 12 hours and stained with monoclonal antibodies against FoxP3 or Tax (MI-73), and then analyzed by flow cytometry.
For in vitro experiments, multiple data comparisons were performed using Student's unpaired t-test. Statistical differences in the incidence of T-cell lymphoma were analyzed using a logrank test.
|
10.1371/journal.pntd.0000236 | The Effect of Azithromycin on Ivermectin Pharmacokinetics—A Population Pharmacokinetic Model Analysis | A recent drug interaction study reported that when azithromycin was administered with the combination of ivermectin and albendazole, there were modest increases in ivermectin pharmacokinetic parameters. Data from this study were reanalyzed to further explore this observation. A compartmental model was developed and 1,000 interaction studies were simulated to explore extreme high ivermectin values that might occur.
A two-compartment pharmacokinetic model with first-order elimination and absorption was developed. The chosen final model had 7 fixed-effect parameters and 8 random-effect parameters. Because some of the modeling parameters and their variances were not distributed normally, a second mixture model was developed to further explore these data. The mixture model had two additional fixed parameters and identified two populations, A (55% of subjects), where there was no change in bioavailability, and B (45% of subjects), where ivermectin bioavailability was increased 37%. Simulations of the data using both models were similar, and showed that the highest ivermectin concentrations fell in the range of 115–201 ng/mL.
This is the first pharmacokinetic model of ivermectin. It demonstrates the utility of two modeling approaches to explore drug interactions, especially where there may be population heterogeneity. The mechanism for the interaction was identified (an increase in bioavailability in one subpopulation). Simulations show that the maximum ivermectin exposures that might be observed during co-administration with azithromycin are below those previously shown to be safe and well tolerated. These analyses support further study of co-administration of azithromycin with the widely used agents ivermectin and albendazole, under field conditions in disease control programs.
| This paper describes the use of a modeling and simulation approach to explore a reported pharmacokinetic interaction between two drugs (ivermectin and azithromycin), which along with albendazole, are being developed for combination use in neglected tropical diseases. This approach is complementary to more traditional pharmacokinetic and safety studies that need to be conducted to support combined use of different health interventions. A mathematical model of ivermectin pharmacokinetics was created and used to simulate multiple trials, and the probability of certain outcomes (very high peak blood ivermectin levels when given in combination) was determined. All simulated peak blood levels were within ranges known to be safe and well tolerated. Additional field studies are needed to confirm these findings.
| The operational efficiency of disease elimination programs in developing countries could be improved by integrating delivery of several interventions at local (village and district) levels [1]–[3]. In areas endemic for co-infection with filarial nematodes and Chlamydia trachomatis, one such integrated disease elimination strategy would be based on mass administration of a three-drug combination: ivermectin for onchocerciasis, albendazole for lymphatic filariasis and azithromycin for trachoma. Regular administration of this combination would also be predicted to reduce other infectious agents including soil transmitted nematodes and bacterial sexually transmitted diseases [4].
A recent pharmacokinetic study evaluated co-administration of azithromycin, ivermectin and albendazole [5], and showed that mean ivermectin pharmacokinetic parameters, area under the concentration-time curve (AUC) and maximum concentration (Cmax), were increased by 31% and 27%, respectively relative to a baseline period. The variability in this interaction was large, with two individuals having 3-fold increases in ivermectin AUC. Increased ivermectin exposures could potentially have safety implications, as high dose ivermectin animal studies and observations of human overdose have reported signs and symptoms of central nervous system (CNS) toxicity including emesis, mydriasis and ataxia [6]. However a recent safety study demonstrated no significant toxicity in the CNS or other body systems, with ivermectin doses up to 10 times the highest labeled dose of 200 µg/kg [7],[8].
The purpose of this analysis was to model the ivermectin pharmacokinetic data from the recently reported interaction study [5], to further characterize the interaction, and explore the sources of variabilities between subjects and across treatments. The model was also used to simulate the outcomes of 1000 trials, to ensure that peak ivermectin exposures seen during co-administration did not exceed those observed in the high dose safety and pharmacokinetic study [7].
Data from a historical Phase I study with intensive sampling in healthy subjects was used to develop a population pharmacokinetic model for ivermectin [5]. All subjects provided written informed consent according to local requirements before entering the study, and the protocol and Informed Consent Form were approved by the local Institutional Review Board. This was a three-arm crossover study, where subjects were administered single-dose regimens of the following treatments in random order: (i) azithromycin 500 mg; (ii) ivermectin 200 µg/kg of total body weight rounded to the nearest 3 mg plus albendazole 400 mg; and, (iii) all 3 drugs administered concurrently. All doses were administered with 240 mL of water and a standardized breakfast. Prior to dosing and breakfast, subjects fasted overnight and then abstained from any further food for 4 hours after study drug administration. Study arms were separated by washout periods of 3 weeks. Full details of the study are provided in [5].
Blood samples were collected predose and at 0.5, 1, 1.5, 2, 3, 4, 6, 8, 10, 12, 24, 36, 48, 72, 96, 120, 144, and 168 hours after drug administration during each of the study phases. Samples were collected into heparinized Vacutainers. Blood samples were centrifuged at 3000 rpm for 15 minutes and the plasma samples were collected in plain plastic tubes without anticoagulant and then stored at −80°C. Samples were shipped frozen overnight on dry ice to BAS Analytics (West Lafayette, IN) for sample analyses. Ivermectin is detected in the body as two metabolites (22,23-dihydroavermectin-B1a (H2B1a) and 22,23-dihydroavermectin-B1b (H2B1b), and these were assayed using a validated high performance liquid chromatography system with liquid chromatography/mass spectrographic detection. The assays were linear over the ranges of 2.5–1000.0 ng/mL and 2.5–20.0 ng/mL, respectively. The precision values for both assays were <10%. In terms of accuracy, while the bias was not exceeded (±15%) for H2B1b for either the high or low quality control (QC) samples, they were for H2B1a during long-term stability testing (−21.8% at the low QC and −17.3% for the high QC) (see [5]). Plasma concentration-time data were analyzed using standard noncompartmental analytical software (WinNonlin 4.1; Pharsight Corporation, Mountain View, CA), and key parameters are shown in Figure 1. The data analysis presented here is for ivermectin data from the ivermectin plus albendazole arm (Baseline Phase), and from the ivermectin, albendazole plus azithromycin arm (Interaction Phase).
Eighteen healthy Caucasian volunteers were enrolled in and completed this study (9 males and 9 females, mean [±SD] age, 39.4±10.5 years, weight 78.2±12.4 kg, ivermectin dose 15.5±2.6 mg).
All the data from both arms of the cross-over study were fitted simultaneously. The data set contained pooled pharmacokinetic, demographic/covariate, and dosing information. Data were analyzed using nonlinear mixed-effects modeling with the NONMEM software system, Version V, Level 1.1 (GloboMax LLC, Ellicott City, MD) with the PREDPP model library and NMTRAN subroutines. Computer resources included personal computers with Intel Pentium 4 processors, Windows XP Professional operating system, the GNU Fortran Compiler, GCC-2.95 (Win-32 version also known as G77; GNU Project, http://www.GNU.org/). Key pharmacokinetic parameters from the modeling are described in Figure 1.
The first-order conditional estimation method with η-ε interaction (FOCEI) was employed for all model runs. Individual estimates of pharmacokinetic parameters were obtained using POSTHOC (an empirical Bayesian estimation method). The random effect models sufficiently described the error distributions. For this analysis all interindividual errors were described by exponential error models on selected parameters (Equation 1).(1)where: Pi is the true parameter value for individual i, is the typical population value (geometric mean) of the parameter, ηPi are individual-specific interindividual random effects for individual i and parameter P and were assumed to be independently and identically distributed following a normal distribution with mean 0 and variance omega (ω) squared (η∼N(0, ω2)).
The data could not support a full covariance block for the OMEGA matrix. Modeling began with the assumption of no covariance between interindividual random effects (diagonal ω matrix). Later, the covariance between clearance (CL) and volume of distribution in the central compartment (Vc) was estimated. For pharmacokinetic observations in this analysis, the residual error model was described by a combined additive and proportional error model (Equation 2).(2)where: Cij is the jth measured observation (plasma concentration) in individual i, is the jth model predicted value (plasma concentration) in individual i, εpij and εaij are proportional and additive residual random errors, respectively, for individual i and measurement j and are assumed to be independently and identically normally distributed, following a normal distribution with mean 0 and variance sigma (σ) squared (ε∼N(0, σ2)). For each treatment arm, separate residual errors were explored. The pharmacokinetic models were evaluated for goodness of fit and were then subjected to predictive check model evaluation. For more detailed technical information on these methods, please see NONMEM user's guide [9].
After the structural pharmacokinetic model was established, known physiologic relationships were incorporated into the covariate-parameter models. For example, the change in physiologic parameters as a function of body size was both theoretically and empirically described by an allometric model (Equation 3) [10](3)where: the typical individual value of a model parameter (TVP) was described as a function of individual body weight (WTi), normalized by a reference weight (WTref), which was 70 kg. θTVP is an estimated parameter describing the typical pharmacokinetic parameter value for an individual with weight equal to the reference weight and θallo is an allometric power parameter (which can be estimated or fixed to a value of 0.75 for clearances, and a value of 1 for anatomical volumes).
Assessment of model adequacy and decisions about increasing model complexity were driven by the data and guided by goodness-of-fit criteria, including: (i) visual inspection of diagnostic scatter plots (observed vs. predicted concentration, residual/weighted residual vs. predicted concentration or time, and histograms of individual random effects; (ii) successful convergence of the minimization routine with at least 2 significant digits in parameter estimates; (iii) plausibility of parameter estimates; (iv) precision of parameter estimates; (v) correlation between model parameter estimation errors <0.95, and (vi) the Akaike Information Criterion (AIC), given the minimum objective function (OBJ) value and number of estimated parameters [9]. The criteria for successful runs were restricted to successful convergence using FOCE with interaction, good diagnostics for the model-fit for all data of the different treatment periods, and reasonable estimates for fixed and random effect parameters. Model evaluations included comparisons of the OBJ between hierarchical models. A decrease in OBJ corresponding to a chi-square distribution with α = 0.01 and degrees of freedom equal to the difference in the number of estimated parameters between the two models was used as the criterion for model comparisons.
Final model parameter estimates were reported with a measure of estimation uncertainty including the asymptotic standard errors (obtained from the NONMEM $COVARIANCE step). A limited covariate modeling approach emphasizing parameter estimation given the available data, rather than stepwise hypothesis testing, was implemented for this population pharmacokinetic analysis. The study population contained equal numbers of males and females. As such, age, weight and gender were explored as potential covariates. First, pre-defined covariate-parameter relationships were identified based on exploratory graphics, mechanistic plausibility of prior knowledge, and then a full model was constructed, with a fixed allometric relationship of body weight on clearance and volume parameters. Interindividual variability could not be incorporated on all fixed-effects parameters to get successful FOCE runs. For residual variance, a separate residual error was assigned for each of the treatment arms. A combined additive and proportional error model was used with 4 parameters to be estimated for the residual error. Various population models were evaluated, but only two models that best described the data (as determined by the log likelihood criterion and visual inspection) are presented. The first modeling approach was a population model that included all subjects. Because some of the modeling parameters and their variances were clearly not normally distributed, and showed asymmetric distribution, a mixture model was developed.
A second modeling approach was a population mixture model as it met our criteria for model adequacy and provided supporting evidence of the dichotomy of the observed individual data. Each subpopulation would have an associated submodel with different fixed or random effects. This model was adopted to accommodate the fact that only some of the individuals exhibited a pronounced increase in ivermectin bioavailability during the interaction arm of the study. It was preferred over a population model with and without outlier individuals, as it gave a better fit to the data as measured by change in OBJ, and met our criteria for a successful run in terms of a complete successful convergence with reasonable estimate for precision for both fixed and random effects.
Model development was guided by various goodness-of-fit criteria, including diagnostic scatter plots. Checking of the individual fits was also employed as part of judging the model performance for each patient. The final model and parameter estimates were then investigated with the predictive check method. This method was similar to the previously described posterior predictive check, but assumes that parameter uncertainty is negligible, relative to interindividual and residual variance [11]. The basic premise is that a model and parameters derived from an observed data set should produce simulated data that are similar to the original observed data. The predictive check is a useful adjunct to typical diagnostic plots, in that the predictive check provides information about the performance of random-effects parameter estimates, whereas typical diagnostic plots are primarily informative about the fixed-effects parameter estimates. The predictive check model evaluation step was performed by using the final model and its parameter estimates to simulate data under the same experimental design of the original data.
One thousand Monte Carlo simulation replicates of the original data set were generated using the final non-mixture and mixture population pharmacokinetic models. Distributions of Cmax across all data simulations were compared with Cmax distribution in the observed data set. The simulated data from each of the 1000 virtual trials (18000 subjects for each treatment period) were assembled, and the similarity between the actual observed data and simulated data was examined by comparing the 95% predictions intervals of the simulated data with the original observed data.
Assessment of the relationship between azithromycin and ivermectin by noncompartmental analysis showed that mean ivermectin AUC and Cmax was increased by 31% and 27%, respectively (see [5] for complete results). Visual inspection of the magnitude of ivermectin accumulation against azithromycin exposure in the interaction arm showed no obvious relationship (Figure 2), and a very low Pearson's r2 (0.03).
Ivermectin concentration-time data were best described by a two-compartment pharmacokinetic model with first-order elimination and absorption (Figure 3). Covariance between CL and Vc elements of the OMEGA matrix was incorporated in the model. The use of different residual variance models stratified by the treatment with and without shared additive components was explored and incorporated into the structural model. Inclusion of age or gender as covariates did not contribute additional information for explaining pharmacokinetic variability based on OBJ differences in hierarchical models, model convergence, as well as diagnostic graphics. Therefore, none of these covariates was included as a covariate in the final population pharmacokinetic model. Importantly, the available data for this investigation contained a relatively small number of subjects and a limited age range, and so formal hypothesis (significance) testing for covariate effects was not considered.
The final non-mixture model had 7 fixed-effect parameters and 8 random-effect parameters as shown in Table 1. Population pharmacokinetic parameters (CL, Vc, Q, Vp; see Figure 1) were standardized to a 70 kg person using the allometric size model [10]. In parametric nonlinear mixed effects modeling, the distribution of ηs is assumed to be normal (mean = 0, variance = ω2). With each model developed, we checked the distribution of ηs, and their mean values. The η distribution indicated a clear violation of the normality assumption. It was necessary to modify the original model to improve η distribution diagnostics. A mixture modeling approach was considered as the distribution of some of the pharmacokinetic parameters and inter-individual variabilities indicated a lack of homogeneity. The final mixture model had 9 fixed-effect parameters and 8 random-effect parameters as shown in Table 2. Goodness-of-fit plots for the final model are shown in Figure 4. The mixture model differed from the non-mixture model in only two parameters: one defining the difference between the two subpopulation in terms of bioavailability, the second defining the partition of the population between the two subpopulations. Using this approach, inter-individual variability distribution was modeled as two subpopulations (A and B). The unknown mixture distribution was estimated at an individual level. The estimate for each subpopulation included different fixed effects parameters, different variance parameters, estimation of fraction of individuals in each subpopulation, and each individual was assigned to the most likely subpopulation. The proportion of subjects in subpopulations A and B was estimated as 55% and 45%, respectively.
Both the population and individual predictions adequately described the AUC profiles for each subject (Figure 5), as displayed by the baseline and interaction phases for subpopulation B. A similar fit of individual data was observed for Subpopulation A (data not shown). Figure 6 displays median, 97.5th, and 2.5th quantiles of the simulated data as lines with the observed data plotted as individual points. Less than 5% of the observed data were outside these 95% prediction intervals. No biased pattern or any tendency for over- or underestimation was noted for the different treatment periods, or for the two subpopulations. This finding gives confidence in the model performance in predicting the expected ivermectin exposures under different circumstances.
Simulated maximum concentrations for each individual's Cmax values were summarized across 1000 simulation replicates of the original population pharmacokinetic database and plotted as box plots (Figure 7). The upper panel shows box plots of the observed ivermectin Cmax for baseline and interaction periods for all subjects, and for the two subpopulations. The lower panel shows box plots for ivermectin Cmax from 1000 simulated trials for the non-mixture model (all subjects), and the mixture model (subpopulations A and B). The mixture model pattern predictions for the two subpopulations were very consistent with the observed data [5]. Extreme values were: non-mixture model: 201.2 ng/mL; mixture model subpopulation A: 115.3 ng/mL; B: 175.5 ng/mL.
There are a number of interesting findings from this analysis of data from an interaction study of ivermectin and azithromycin. This is the first published population model of ivermectin pharmacokinetics. It demonstrates the utility of population mixture modeling as an approach to explore drug interactions, especially where there may be population heterogeneity. The mechanism for the interaction was identified (an increase in bioavailability in one subpopulation). The model was used to simulate multiple clinical trials, to identify the maximum exposures that might be observed during co-administration, which permits comparison with previously published safety and pharmacokinetic data.
Ivermectin has been approved for use in humans for 2 decades, yet relatively limited pharmacokinetic data have been published. Recent studies using modern assay methods have characterized its pharmacokinetics using noncompartmental methods in the context of drug combination studies for treatment of onchocerciasis and lymphatic filariasis [12]–[14], or in high doses for treatment of head lice [7]. The calculated model parameters are in close agreement with those determined using noncompartmental methods [5]. A two compartment model is consistent with the disposition of ivermectin in man and other species, with a high volume of distribution into a peripheral compartment [15]. Ivermectin is metabolized extensively in the liver via cytochrome P450 isozyme (CYP) 3A4 [16]. It is both a substrate for the transporter P-glycoprotein (Pgp) [17],[18], as well as a moderately potent Pgp inhibitor at concentrations consistent with clinical exposures in the present study (IC50 0.18–0.4 µM; [19],[20]).
The variability of the magnitude of change in ivermectin pharmacokinetics observed in the interaction phase [5] complicated the interpretation of the presence or absence of a drug interaction, as the response was very inconsistent among individuals. One of the objectives of this analysis was to explore how nonlinear mixed-effects modeling could be used to analyze such heterogeneous and highly variable experimental data from a relatively small number of subjects, with intensive pharmacokinetic sampling. The initial non-mixture model provided an adequate description of ivermectin pharmacokinetic data, however interindividual variability was not homogeneous and could not be explained by the available covariates. A mixture model was able to resolve this, and provided an explanation for the observed differences in bioavailability seen in the clinical study. Mixture modeling assumes two or more subpopulations exist, rather than a single homogeneous one [21], and the final model has two additional fixed parameters, one relating to subpopulation differences in ivermectin bioavailability, and the other defining the two subpopulations.
The final mixture model provided a good description of ivermectin data from both treatment periods. Goodness-of-fit criteria revealed that the final model was consistent with the observed data and that no systematic bias remained. The data points (Figure 4) are scattered closely and randomly around the line of identity, and the homogenous and random distributions of weighted residuals indicate the error model was suitable for describing the variance of the data. The model evaluation results provided evidence that both the fixed-effects and random-effects components of the final model were reflective of the observed data as well. The fact that less than 5% of the data were located outside the 2.5-97.5th quantile range suggests that the model accurately describes the central tendency and the variability of the data for the two subpopulations and for the two treatment periods, despite the large number of parameters and the low number of patients who participated in the study. The predictive check shows there is no bias at any phase of the pharmacokinetic profile, which makes the model useful in predicting ivermectin blood concentrations, when given alone or co-administered with azithromycin.
Typically, a mixture modeling approach would not be considered at the outset of a population pharmacokinetic analysis. Because of the unexplained remaining variability (see above), in the present analysis, the following decision rules were used in the evaluation of the mixture model: (i) The Estimation step and Covariance step terminated successfully; (ii) 95% CI for Mixture partition did not include 0 nor 1; and (iii) the change in the OBJ between mixture and non-mixture models was >5.99 (χ2; p<0.05, 2df). In the present analysis, the difference was 19.8.
The mixture model identified the interaction between azithromycin and ivermectin to be due to changes in bioavailability in Subpopulation B. Their mean estimate of bioavailability (F) was 1.37 relative to baseline, whereas F was unchanged for Subpopulation A (0.97). Inspection of noncompartmental data for Subpopulation B were consistent, showing higher Cmax and earlier Tmax values (Cmax A: 54.3 ng.h/mL; B: 67.8 ng.h/mL; Tmax A: 4.1 h; B: 3.4 h). There were no differences in apparent clearance or volume of distribution. However the mechanism for the increase in bioavailability is unclear. Azithromycin, like ivermectin, is a substrate for Pgp, however it has minimal inhibitory effects on this transporter in vitro [20]. Although ivermectin is extensively metabolized by CYP3A4 [16], azithromycin has no inhibitory activity against this enzyme [22]. There are no other plausible metabolic or transporter mechanisms that could explain an interaction, and no clinical covariates were identified that characterized either subpopulation. In addition, mean pharmacokinetic parameters of ivermectin were similar in both subpopulations in the baseline phase (mean AUC A: 1019; B: 805 ng.h/mL; Cmax A: 52; B: 45 ng/mL; Tmax A: 5.3; B: 4.8h).
The model was used to simulate the range of peak ivermectin concentrations that might be encountered if azithromycin and ivermectin were co-administered. These simulated data were then compared with the Cmax data reported in the high-dose ivermectin safety study [7]. The median simulated Cmax data (46.0, 34.1 and 40.3 ng/mL for non-mixture model, mixture models A and B respectively) were approximately 5–7-fold lower than the 261 ng/mL value reported by Guzzo et al [7]. Indeed, the most extreme individual simulated values (201.2, 115.3 and 175.5 ng/mL for non-mixture model, mixture models A and B respectively) were still lower than the mean value reported in the high-dose study [7]. These data give a high level of confidence that peak exposures that are predicted to occur if ivermectin and azithromycin were co-administered would never exceed mean values seen under high dose conditions [7], and which in this study were safe and well tolerated. In the Amsden et al interaction study [5], ivermectin was dosed with food (a high-fat breakfast). Food has been shown to increase the bioavailability of ivermectin over 2-fold [7]. Because dosing of patients in Africa is unlikely to be with high fat meals, extreme peak ivermectin concentrations would be half of those reported in the simulation.
Interestingly, simulations from both the mixture model and the non-mixture model had generally similar predictions of ivermectin exposures (average estimates and variability). Both models confirmed that the maximum concentration achieved in the interaction phase would not exceed 201 ng/mL (Figure 7). In spite of adding two parameters to the non-mixture model; the final parameter estimates for both models were very similar (Tables 1 and 2). The inflation of variability and projections of extreme values for both sets of simulations is a consequence of using 1000 replicates, where the chances of sampling from the very extreme values of random error distributions are more probable. However predicting extreme high values, even if they are very rare, is very useful from a safety perspective, and provide a “worst case” scenario of any extreme high exposures that might be encountered in a clinical setting/trial during co-administration.
There are several important caveats to this analysis. The data collected from the drug interaction study was not intended for population analysis, and a larger data set would have been desirable. The use of a mixture model could be criticized on the basis that random variations in the data could be ascribed post hoc to population differences. Indeed, although the mixture model identified two populations on the basis of different effects on bioavailability, it is unclear mechanistically what this difference might be due to. Finally, modeling and simulation can advise but cannot supplant clinical data. The findings from this study should be confirmed in further clinical or pharmacokinetic studies.
In conclusion, this analysis demonstrates the utility of a population model approach to analyze drug interaction data. The mechanism for the interaction was identified (an increase in bioavailability in one subpopulation). The model was also used to simulate multiple clinical trials, to identify the maximum exposures that might be observed during co-administration, and provides confidence that the peak ivermectin exposures would never exceed mean exposures that have previously been shown to be safe and well tolerated. |
10.1371/journal.pbio.1001150 | Active Control of Acoustic Field-of-View in a Biosonar System | Active-sensing systems abound in nature, but little is known about systematic strategies that are used by these systems to scan the environment. Here, we addressed this question by studying echolocating bats, animals that have the ability to point their biosonar beam to a confined region of space. We trained Egyptian fruit bats to land on a target, under conditions of varying levels of environmental complexity, and measured their echolocation and flight behavior. The bats modulated the intensity of their biosonar emissions, and the spatial region they sampled, in a task-dependant manner. We report here that Egyptian fruit bats selectively change the emission intensity and the angle between the beam axes of sequentially emitted clicks, according to the distance to the target, and depending on the level of environmental complexity. In so doing, they effectively adjusted the spatial sector sampled by a pair of clicks—the “field-of-view.” We suggest that the exact point within the beam that is directed towards an object (e.g., the beam's peak, maximal slope, etc.) is influenced by three competing task demands: detection, localization, and angular scanning—where the third factor is modulated by field-of-view. Our results suggest that lingual echolocation (based on tongue clicks) is in fact much more sophisticated than previously believed. They also reveal a new parameter under active control in animal sonar—the angle between consecutive beams. Our findings suggest that acoustic scanning of space by mammals is highly flexible and modulated much more selectively than previously recognized.
| Most sensory systems have an active component, i.e. driven by an animal's behavior, which contributes directly to its perception. For example, eye movements are important for visual perception, sniffs are crucial for olfactory percepts, and finger movements for touch percepts. A classic example of an active-sensing system is bat echolocation, or biosonar. Echolocating bats actively emit the energy with which they probe their surroundings, and they can control many aspects of sensory acquisition, such as the temporal or spectral resolution of their signals. A key open question in bat echolocation concerns bats' ability to actively change the area scanned by their emitted beam. Here, we used a large microphone array to study the echolocation behavior of Egyptian fruit bats. We found that these bats apply a new strategy to alter the area scanned by their beam; specifically, bats changed their acoustic field-of-view by changing the direction of consecutively emitted beams. Importantly, they did so in an environment-dependent manner, increasing the scanned area more when there were more objects in their surroundings. They also increased their field-of-view when approaching a target. These findings provide the first example for active changes in sensing volume, which occur in response to changes in environmental complexity and target-distance, and they suggest that active sensing of space is more flexible than previously thought.
| The importance of “active sensing,” by which an animal actively interacts with the environment to adaptively control the acquisition of sensory information, is fundamental to perception across sensory modalities [1]–[7]. Echolocating bats emit ultrasonic signals and analyze the returning echoes to perceive their surroundings. Bat echolocation, an active sensory system, enables an acoustic representation of the environment through precise control of outgoing sonar signals. Laryngeal bats control many aspects of their sensory acquisition: they determine the timing of acquisition and the information flow [8]–[11], they control the intensity of the emission as well as its direction [12]–[17], and they control the spectral and temporal resolution of the acquired data [18]–[23]. Another acoustic parameter potentially under active control by echolocating bats is the pattern of the sonar beam. It has been debated whether bats can actively adjust the width of the sonar beam in response to task conditions, but empirical studies have not yet adequately addressed this question. It seems likely that bats would benefit greatly from the ability to control the beam pattern. They could for instance narrow the beam in order to concentrate energy onto a certain object, or they could widen the beam to increase the size of the sector that is being scanned. Studying the bat's active control over the shape and directionality of sonar emissions is technically difficult because reconstruction of the beam pattern requires a large circumferential ultrasonic microphone array in a setting where a free-flying bat engages in sonar tasks. A recent study suggests that laryngeal echolocating bats can change the space covered by their beam through adjustments in their call spectrum [24]. Here, we aimed to examine a very different mechanism by which echolocating bats might control the effective space they scan, namely adjustments in the angle between sequentially emitted sonar clicks.
We studied this question in lingual echolocating bats. Lingual echolocation is exhibited by one family of fruit bats, Rousettus, and has been historically considered to be more rudimentary than laryngeal echolocation [25]. The primary reason behind this notion was that these bats were believed to have very little control over their sonar emissions. In contrast, we recently demonstrated that the lingual echolocator Rousettus aegyptiacus (Egyptian fruit bat) uses a sophisticated strategy for beam-steering: This bat emits sonar clicks in pairs, and it directs the maximum slope of each sonar beam towards the target, rather than directing the center of the beam, thereby optimizing stimulus localization in the horizontal plane [15]. Here, we further tested Egyptian fruit bats' active control over their echolocation-based sensory acquisition. To this end, we tracked the flight trajectories of Egyptian fruit bats in a large room, and recorded their echolocation behavior when performing a landing task under different levels of environmental complexity. We found that lingual echolocation allows much more selective control over sonar signal parameters than previously believed. We discovered that Egyptian fruit bats alter the intensity of their emissions as they approach and lock the sonar beam onto a target, and that emission intensity changes with environmental complexity. Moreover, we found that Egyptian fruit bats apply a novel strategy to change the spatial region, or “field-of-view” that they scan: They increase the angle between the beam axes of sonar click-pairs, to effectively increase spatial scanning. Such a strategy has never been observed before in any bat species, and therefore comprises a new dimension of active control in lingual bat echolocation.
In the first set of experiments—the “one-object experiments”—bats were trained to detect, localize, and land on a 10-cm diameter sphere, similar in size to fruit eaten by this bat species, such as mango. The sphere was the only object in an empty flight room (Figure 1A), and it was randomly moved between trials. Recordings were taken in complete darkness, forcing the bats to rely only on echolocation (see Materials and Methods). The echolocation of Egyptian fruit bats is comprised of pairs of clicks with a short inter-click time interval (∼20 ms) and a longer inter-pair interval (∼90 ms in complete darkness) [26],[27]. The bats direct their sonar beam axes Left-Right→Right-Left, maintaining a certain angle between the sequential clicks of a pair (Figure 1A–B) [15]. When approaching the target, bats significantly increased the inter-click angle by 6.8±0.4 degrees, on average (mean ± s.e.m.; t test of inter-click angle before locking versus after locking, when pooling all data together: p<10−5). This increase in inter-click angle occurred abruptly, coinciding with the time when the bats locked on the landing target, i.e. the time when the average direction of the click-pair coincided with the direction to the target (Figure 1C; see Materials and Methods) [15]. The increase occurred in all individual bats (Figure S1), and on average across all bats the change represented a 15% widening in the inter-click angle (post-locking compared to pre-locking). Population analysis of 236 trials (Figure 1D) confirmed that the increase of the inter-click angle was abrupt; in fact, it could occur within 2 click-pairs, i.e. as fast as 200 ms (Figure 1C–D).
This abrupt increase in inter-click angle may result from the bat's need to increase the field-of-view; or it may represent the animal's attempt to position the maximum slope of its sonar beam onto the target [15]. To further elucidate the possible roles of this abrupt change in inter-click angle, we conducted additional experiments that aimed to challenge the bat's scanning behavior. To this end, we manipulated the spatial complexity (number of objects) that the bat encountered within its field-of-view as it flew towards the landing sphere.
In the next set of experiments, we manipulated the complexity of the environment, and examined how this influenced the Egyptian fruit bat's echolocation behavior. We hypothesized that when introducing a set of objects (obstacles) in the vicinity of the landing-point, which increases the environmental complexity, the bats would alter their scanning behavior to inspect several objects—thus increasing their field-of-view. To test this hypothesis, we studied the bats' behavior in two new setups (Figure 2A): (i) Open room condition: In 56 trials (8–12 trials per bat) we removed the sphere where the bats were trained to land. These trials were randomly introduced in between one-object trials; hence the bats reacted by vigorously searching for the target while flying around the room. We shall refer to this setup as the “no-object” experiment (Figure 2A, left). (ii) Environmentally complex condition: In 54 trials (8–11 per bat) we added two nets that were spread between four poles on both sides of the target, creating a relatively narrow (0.6–1.6 m) corridor for accessing the target (Figure 2A, right). The width of the corridor, its angle relative to the walls of the room, and the position of the landing sphere within the corridor were all randomly varied between trials. This setup mimics natural situations, in which a bat has to negotiate fruitless branches (the nets), before landing on a branch with a fruit (the target). We refer to this setup as the “multiple-object” experiment, because the bats consistently negotiated some or all of the five objects in the room—the single landing sphere (Figure 2A, right, closed gray circle) and the four poles (open circles). In all illustrations, bat's trajectory is depicted by a gray line and the direction of the beam's peak by a black line.
Egyptian fruit bats increased the angle between sequential clicks when environmental complexity increased (Figure 2A, bottom). The angular separation between the beam axes of sonar click pairs in the “no-object” setup was the narrowest; it increased in the one-object setup by 9.2±0.4 degrees (after locking, “L”), and increased even further in the multiple-object setup, widening on average by 12.3±0.6 degrees compared to the “no object” setup (Figure 2A bottom, “multiple object,” after locking). This behavioral pattern was consistent across all the individual bats that we tested (Figure S2). Statistical analysis showed that the increase in inter-click angle was highly significant (Figure 2A, bottom: one-way ANOVA: F>71, p<10−8; post-hoc t tests: p<10−11 for comparing one-object experiments after locking versus no-object experiments; p<10−6 for multiple-object experiments after locking versus one-object after locking). In the multiple-object setup, the bats increased the inter-click angle significantly beyond the point of maximum slope (i.e., the maximum slope of the beam was lateral to the target; t tests: p<10−3 for comparing one-object experiments after locking versus multiple-object experiments after locking). This suggests that, at least in this case, the inter-click angle plays another role in addition to placing the maximum slope on target for optimizing localization. We propose that widening the angle between the beam axes of sonar click pairs serves to modulate the bat's field-of-view. During the last time-bin before landing (Figure 2B, right-most point), the inter-click angle has increased on average by 14.5 degrees, compared to the mean angle in no-object experiments. When doing so, the point in the beam that was pointed to the center of the target was 2.5 degrees medial to the maximum slope. In the multiple-object experiment (Figure 2B), unlike in the one-object setup (Figure 1D), it seemed that the bats did not increase the inter-click angle abruptly (when we used the same locking criterion), but instead began the approach to the landing sphere with a large inter-click angle, and gradually increased even further after the final locking onto the landing target (Figure 2B). However, this gradual change may have been a result of temporal smearing that is specific to the multiple-object condition, and which is due to the difficulty in defining the exact time of “locking” in the multiple-object experiments: Although we defined sonar beam locking with reference to the landing sphere (i.e., when the average of the click pair was directed towards the landing target), the bats often locked onto the net's poles before locking onto the landing target (the 10-cm sphere). This means that they could have been in a “locked” sonar mode (locked onto a pole) when we defined them as un-locked relative to the landing target (see more details in the Discussion). We therefore tested an alternative sonar locking criterion for the multiple-object experiments, defining locking as the moment when the bats entered a corridor between the nets. This criterion revealed a clearer picture of the inter-click angle dynamics in the multiple-object situation (Figure 2C): Well before passing between the nets, the bats used an intermediate inter-click angle (5.8±0.7 degrees wider than no-object), which is between the locked and un-locked one-object situations. When the bats approached closer to the net corridor, they rapidly increased the inter-click angle to nearly its final value; subsequently, after the bats entered the net corridor, another slight increase was observed, which brought the inter-click angle to an average value that was 14.5±2.0 degrees wider than in the no-object experiments. At the plateau, the center of the target was ∼2.5 degrees beyond the maximum slope (t tests: p<10−3 for comparing one-object experiments after locking versus multiple-object experiments after locking). Maintaining such a high inter-click angle could possibly allow the bat to track both the target and the off-axis objects (distal poles) as the animal approaches landing—providing a potential strategy for target landing while avoiding collisions.
Interestingly, when further analyzing data from the one-object experiments, we found that in some trials, especially when the bats flew a long trajectory before landing, the bats sometimes locked their sonar on the landing sphere, then redirected the beam away and later performed the “final” locking when starting the final approach. The “E-L” bar (dark gray) in Figure 2A represents the inter-click angle during these “early locking” instances. It shows that the bats increased the inter-click angle even when they only transiently locked onto the target (during early locking, “E-L”). The widening of the inter-click angle in these instances was not as salient as in the final locking, probably because this beam-angle adjustment occurred for rather short periods of time (only a few click-pairs), and when the bats were rather far from the target (>1.5 m).
In addition to the increase in inter-click angle, we found that Egyptian fruit bats decrease their emission intensity along the approach to landing (Figure 3A–C). We always refer here to peak intensity (see Materials and Methods), but since the duration of the sonar clicks is very constant, this is also highly correlated to the click's total energy. Because the bats in this experiment were free to choose the trajectory of landing, it was not always relevant to analyze the bat's distance to the target: for instance when a bat circles the target, it could be very close to it in terms of distance but very far in terms of time-to-landing (and may in fact be echolocating in a different direction). We therefore examined the intensity versus time-to-locking (Figure 3A–B), as well as intensity versus distance-to-target in trials in which the distance decreased nearly monotonically as the bat approached the landing sphere (Figure 3C). Figure 3C shows six examples in which the bat flew directly to the target, exhibiting a salient reduction in intensity, with a 4–6 dB decrease with halving of the distance-to-target during the final approach (Figure 3C, gray line, close to target). These results are consistent with reports in other bat species [28],[29]. Interestingly, this decrease in intensity began only 80–100 cm before landing—similar to what was observed in laryngeal echolocators [28]. Thus, the intensity dynamics along the approach seem to be shared by clicking and laryngeal bats.
In addition to increasing the inter-click angle, bats also increased the intensity of their clicks with environmental complexity. The intensity increased by 6.5±0.6 dB on average in the one-object experiments compared with the no-object experiments, and further increased by 2.6±0.8 dB on average in the multiple-object experiments—that is, a total intensity increase of 9.1 dB in the multiple-object versus no-object condition (Figure 3D). These modulations of intensity could be used by the bat to maintain fixed signal energy directed towards the region of interest, compensating for changes in signal-to-noise ratio due to a widening field-of-view (see Figure 4, and next section). These differences in intensity were highly significant (one-way ANOVA: F>108, p<10−9; post-hoc t tests: p<10−33 for t test of one-object versus no-object; p<10−16 for multiple-object versus one-object; here we pooled together data from the approach phases before and after locking). Since we used a planar rather than a 3-D microphone array, and could not calculate the absolute emitted intensity, we performed explicit tests to control for the effects of bats' height, the distance from the microphones, and flight pitch (see Materials and Methods). The increase in intensity, together with the increase in inter-click angle, both contribute to an increase in the effective area that is sampled by the bats via a single click-pair (see next section and Discussion).
Our two main findings—that Egyptian fruit bats increase their inter-click angle and also increase the click intensity with increased environmental complexity—suggest that the field-of-view scanned by the bat is under active control and adapted to the environment. These adaptive sonar signal changes served to increase the bat's field-of-view when the environment became more complex (i.e., contained more objects). To examine this notion further, we quantified the field-of-view scanned by the bat, assuming a constant ensonification-intensity level and calculating the change in the angle of the sector covered by the bat's beam. When we used the intensity at the crossing point of the two beams in the one-object setup as reference (Figure 4 dashed lines, normalized intensity 1, see Materials and Methods), we found that the angle of the sector scanned by the bat with a single click-pair increased by a factor of 2.18 in the one-object experiments in comparison to the no-object (from 44 to 96 degrees), and by a factor of 2.73 in the multiple-object experiments in comparison to the no-object setup (from 44 to 120 degrees, see Figure 4C versus 4A). Interestingly, the same intensity (corresponding to a normalized intensity of 1 in Figure 4) is directed towards the crossing point of the two beams (where the object of interest is positioned) in both the multiple- and one-object setups, and it is the peak intensity (directed forwards) in the no-object setup. These modulations might thus reflect the bat's attempt to maintain a fixed energy impinging on the region of interest, compensating for the changes in signal-to-noise ratio due to the changes in field-of-view. The maximum distance (range) scanned by the bats also increased with environmental complexity, because detection range increases as the fourth root of the increase in intensity [30]. Thus, the 3-D “sensory volume” of space [31] that was scanned by the bats has increased at least 3-fold in the multiple-object versus the no-object experiment.
We further examined several additional echolocation parameters in this set of experiments, and the results are summarized here. (i) We did not find any significant change in the beam width of the single clicks in the different environments. (ii) The bats did not significantly change the click repetition-rate in the multiple-object experiments in comparison to the one-object experiments (i.e., the intra-pair interval remained 23 ms on average and inter-pair interval was 93 ms on average). However, in the no-object experiments there was a small but significant decrease in the repetition rate, whereby the inter-pair interval increased by 8 ms (101 ms on average, t test of no-object versus one-object, for intervals >40 ms: p<10−10; Figure S3). (iii) We tested the spectral content of the echolocation clicks (recorded with a wide-band microphone) only in the one-object experiments; hence we cannot exclude changes in the spectra of the clicks. However, spectral changes seem physiologically unlikely, considering the tongue-production mechanism of the brief lingual clicks. Thus, the most salient changes that we observed were changes in inter-click angle, and changes in click intensity. These two parameters changed in opposite directions along the approach path to an object (inter-click angle increased while click intensity decreased during the approach), and both of these parameters increased substantially with environmental complexity.
The research findings presented here suggest that lingual (click-based) echolocation allows more adaptive control than previously reported. Egyptian fruit bats performing a landing task changed both their emission intensity and inter-click angle as they approached a target, in a manner that depended on both the environmental complexity and the behavioral phase. The increase in inter-click angle might serve two different functions: (i) Pointing the maximum-slope to the target: In the one-object setup, the increase in inter-click angle coincided with the moment of locking (Figure 1C–D), thus representing a behavioral phase-transition that could serve the function of directing the maximum slope to the center of the landing sphere, in order to optimize stimulus localization [15]. In comparison, in the no-object setup, the bats aimed most of the energy forward, in the direction of the flight, by decreasing the inter-click angle (Figure 2A, bottom). The narrow inter-click angle before locking, in the one-object situation, is very similar to the angle in the no-object situation, and might thus represent a narrow, forward focused field-of-view that is used before the final approach to the target. (ii) Changing the field-of-view: When shifting from the one-object to the multiple-object setup, the increase in inter-click angle was likely caused by the need to increase the field-of-view. In the multiple-object setup, the bats had to land on a specific target that was placed in the vicinity of other obstacles (e.g. poles, nets). In this situation, the bat's own motion created very large and rapid angular changes in the directions to nearby objects, and hence the bats would need to increase the field-of-view in order to track these objects.
Interestingly, the bats also decreased the emitted intensity while approaching the landing target (within a given level of environmental complexity). Such intensity decrease was not reported in a previous study of Rousettus echolocation [27], probably because they did not record bat signals during landing in that study. In our study, we observed a decrease in click intensity only during the last 80–100 cm before landing (Figure 3C), which suggests that the intensity decrease is initiated only when the bat actually approached the landing sphere. Thus, Rousettus bats increase the field-of-view and concurrently reduce the emitted intensity when approaching landing. A similar behavior is exhibited by approaching laryngeal echolocators [24]: The calls in the terminal group of these bats have more energy in low frequencies, and thus a wider beam, but also lower peak intensity. Lingual echolocators (e.g., Egyptian fruit bats) seem to have developed an alternative way to increase the effective beam width, which does not require them to change the spectral content of their emission. Instead, they change the scanning width by adjusting the angle between the axes of these two beams, and may treat the echoes returning from two consecutive clicks as a single “information unit” [15]. Such a strategy, which is based on adjusting the angular separation between two consecutive sonar emissions within a click-pair, has never been reported in any bat species to date, and it suggests an alternative adaptive mechanism in bat echolocation to sample a wider spatial region. Laryngeal echolocators are also known to steer their beams [13] and could thus also adjust the directional aim of successive sonar calls to control spatial sampling. However, there is no evidence for any laryngeal echolocator that constantly emits pairs of signals, similar to the Egyptian fruit bat; and accordingly, there is no evidence for any laryngeal echolocating bat that regards pairs of signals as their basic “sonar unit.” In addition, Egyptian fruit bats are probably able to achieve such quick changes in beam steering by rapid tongue movements [15], while changes in beam steering in laryngeal echolocators would probably require head movements, and would thus be slower than 20 ms. Thus, the field-of-view control strategy, suggested here for Egyptian fruit bats, might be a unique phenomenon among echolocating animals.
The increase in emission intensity in the different environmental setups may represent an attempt by the bat to maintain fixed energy directed towards the region of interest, thus compensating for changes in signal-to-noise ratio due to changes in field-of-view. Figure 4A–B shows that the region of interest (i.e., the crossing-point of the right and left beams) has approximately the same intensity in the one-object as in the multiple-object setups (horizontal dashed lines). Interestingly, the peak intensity that is being directed towards the direction of interest in the no-object setup is identical to the crossing-point-intensity in the one-object and multiple-object setups (Figure 4C, horizontal dashed line shows normalized intensity 1). This could be interpreted as a principle of “conservation of signal-to-noise” in lingual bat echolocation, and can explain the seemingly paradoxical behavior of decreasing emission intensity when performing a search task (in the no-object setup).
In a previous report [15], we described a trade-off between detection and localization in the Egyptian fruit bat, whereby detection is maximized by pointing the peak of the beam towards an object, while localization is optimized by pointing the maximum-slope towards the object. This tradeoff predicts that the bat will direct its sonar beam towards an object of interest at an angle that rests between the peak and the maximum slope. In our current multiple-object experiment, the bats deviated from this principle by consistently increasing the inter-click angle such that they directed the beam towards the target at points beyond the maximum slope of the beam (Figure 4A, see “+”). This finding is surprising, because it means that target localization cues were now likely diminished. In light of these results, as well as the other results presented in this article, we believe that a new dimension has to be considered, thus introducing a three-way tradeoff between (i) detection, (ii) localization, and (iii) angular scanning (modulated via changes in field-of-view). We suggest that in a complex environment, the need to scan the area around the landing-point, and to increase the field-of-view, is sufficiently important for the bats to reduce localization accuracy. Note that detection was actually not reduced by the increase in the inter-click angle in the more complex environments, because the bats also increased the click intensity, possibly as a compensatory mechanism (Figure 4A–C dashed lines).
What is the functional relevance of the sonar field-of-view? All the previous studies that were conducted on beam steering in laryngeal echolocating bats suggested that, despite their broad emission beams (60–70° width at −3 dB [24],[32]), these bats carefully direct the center of their beam towards the object of interest [13],[14],[24],[32]. Our previous study of sonar beam steering in Egyptian fruit bats showed that these lingual echolocators direct the center of their beam-pair onto the target [15], reminiscent of the individual calls of laryngeal echolocators. The behavior observed in the current study suggests that Egyptian fruit bats collect sensory information also from their acoustic periphery.
In the multiple-objects experiments, the bats exhibited a wide repertoire of behaviors before landing on the target (see details in Materials and Methods). In many cases (∼30% of trials) the Egyptian fruit bats only locked onto one of the poles, or occasionally did not lock on any of the poles while entering the corridor between the nets. We cannot completely exclude the possibility that the bats were relying on spatial memory (see Materials and Methods), but data from these trials imply that the bats can localize an object to some extent without the need to point the center of the beam-pair towards it. Increasing the field-of-view in order to follow objects near the landing target thus makes perfect sense from the bat's point of view.
In summary, our findings reveal two new aspects of adaptive control in lingual bat echolocation, namely the ability to change emission intensity as well as changing the inter-click angle between sequential emissions. The ability of lingual bats to change the inter-click angle reveals a new strategy for bats to actively control the field-of-view that they scan. Adjustment in field-of-view could also theoretically be exploited by laryngeal echolocators through movements of the head, mouth opening, and spectral changes in sonar emissions. The Egyptian fruit bat's directional aim of tongue click pairs demonstrates a new parameter of acoustic control in animal sonar. We suggest that environment-associated changes in emission intensity seem to be related to changes in field-of-view, and can compensate for decreases in signal-to-noise ratio due to changes in field-of-view. Further, our results suggest a three-way trade-off between three goals that a bat has to fulfill with its echolocation in a target-landing task: The detection of an object of interest, its accurate localization, and controlling the field-of-view that is being scanned by the bat. We believe that further studies of sensory trade-offs in echolocating bats will shed new light on bat echolocation—and more generally, on sensory constraints in active-sensing systems.
All experimental procedures were approved by the Institutional Animal Care and Use Committees of the Weizmann Institute of Science and the University of Maryland.
Five adult Egyptian fruit bats (Rousettus aegyptiacus) were trained to detect, localize, and approach a polystyrene sphere (10-cm diameter) that was mounted on a vertical pole positioned inside a large flight-room (6.4×6.4×2.7 m; Figure 1A). The target's size mimics the size of some fruits eaten by these bats in nature, such as mango. To minimize sound reverberations, the walls of the room were covered with acoustic foam and the pole was covered with felt. In order to ensure that the bats were relying solely on echolocation to perform the task, we took the following precautions: (i) To exclude the possibility of using visual cues, the target was painted black and the room was in complete darkness (illuminance <10−4 lux). The experimenter inside the room wore night-vision-goggles with infrared illumination. (ii) To prevent use of olfactory cues, the bats were food-rewarded only after landing on the target. The target was also cleaned with soap and water after every three trials to remove any possible odors that remained on it due to the contact with the bat. (iii) After every trial, the target was randomly re-positioned inside the room, both in the horizontal and in the vertical planes (the pole had a telescopic mechanism that allowed changing the target height). It took the bats ∼4 wk in order to learn the task and once they learned it they always succeeded in landing on the target.
The basic setting included only the landing target (10-cm polystyrene sphere) in the flight room. We also tested two alternative settings: (i) In 56 randomly interspersed trials we removed the landing target from the room, which made the bats eagerly fly in search for the target. We call these experiments the “no-object” experiments. (ii) In 54 trials, we added two nets mounted on 4 poles on both sides of the landing target (Figure 2A top, right). The distance between the nets randomly varied between trials (in the range of 0.6–1.6 m) and so did the position of the landing target and the angle of the nets in relation to the target. The bats learned to correctly land on the target between the nets—within 3–4 trials (which were not counted within these 54 trials); nevertheless, bats still occasionally landed on the poles even after many more trials. They were only rewarded for landing on the original target (sphere). Because these experiments involved five salient objects (1 target+4 poles), they were termed here “multiple-object” experiments.
The bats exhibited a wide behavioral repertoire in the multiple-object experiments: In some trials, they behaved similarly to the behavior described for the laryngeal echolocator, Eptesicus fuscus [14]. In those previous experiments, E. fuscus were trained to fly through a hole in a net, and they typically scanned both sides of the hole (pointing the peak of the beam to each edge of the hole) before flying through it. In the equivalent trials in the current study, Egyptian fruit bats locked the center of their click-pairs on both poles that outlined the opening of the net corridor, and only subsequently they flew through the corridor. In other trials, the Egyptian fruit bats either locked onto one of the poles before landing or did not lock on any of them. Because the bats had the opportunity to fly around the poles and nets before approaching them, and could thus learn their spatial locations, we could not completely exclude their relying on spatial memory. However, it is not likely that this was the only factor facilitating their approach, because the location and layout of the setup was always randomly changed between trials, and the bats did not always scan the setup before approaching the landing target.
The bat's average flight speed was negatively correlated with the environmental complexity (1.2±0.9 m/s in the multiple-object experiment, 1.9±1.2 m/s in the one-object experiment, and 2.4±1.0 m/s in the no-object experiment; mean ± s.d.; p<0.001 for all three t test comparisons, and F>880, p<10−10 in a one-way ANOVA test). This difference in flight speed remained when we analyzed the speeds only for pre-locking or only for post-locking epochs. We believe that the changes in flight speed were a result of the different maneuverability situation, due to the difference in the environmental complexity. In the multiple-object setup, the flight-speed likely decreased also because of the need to slow down in order to allow more time to scan the setup (in the multiple-object experiments, the bats typically slowed their flight before entering the net corridor, or when scanning the poles). Since the bats had the possibility to pre-scan the room, they could potentially adjust their speed to the expected maneuverability conditions, and this is likely why we did not see a change in flight speed between the pre-locked and post-locked situations.
The bats' echolocation behavior was recorded with an array of 20 microphones spaced 1-m from each other around a rectangular supporting frame (5.3×5.2 m), at a height of 90 cm above the floor (Figures 1A and 2A, top: black dots around the circumference of the room show microphone locations) [32]. The signal from each microphone was amplified and fed into a band-pass filter centered around 35 kHz, with a frequency response that matches the frequency content of the Rousettus sonar click (see details in ref. [15]). Next, the signal was fed to an electronic circuit which extracted the envelope of this band-passed signal. The envelope was then low-pass filtered and digitized into a data-acquisition computer. Finally, the maximum value of this signal was translated into a dB scale in which analysis was performed. In order to control for changes in click spectra, in ∼20 trials of the one-object experiment we have recorded the audio using three wideband ultrasonic microphones positioned on the floor (sampled at 250 kHz/channel).
To ensure that we were only using high-quality data, we included only clicks that were clearly above noise level in at least five microphones of the array. In addition, we excluded beam measurements that were either too wide or too narrow relative to the overall distribution of >5,000 beam patterns recorded during >300 trials, because deviant widths led us to suspect a recording artifact due to temporary noise in some of the channels. To this end, we measured the width of the beams [15], and accepted only clicks with: 30°<beam width<120°. This resulted in exclusion of ∼6% of the clicks. In total, we analyzed here 5,144 sonar clicks from 346 behavioral trials in 5 bats (56 no-object trials, 236 one-object trials, and 54 multiple-object trials). We only analyzed clicks that occurred more than 250 ms before landing, because later clicks were emitted when the bat was too close to the target (closer than 15 cm on average), where any angular calculation of direction-to-target would suffer from very high error. This typically corresponded to excluding the last two click-pairs in the trial.
All 20 signals (from 20 microphones) were first segmented to include vocalizations and exclude echoes. Then, the intensity at each microphone was corrected for spherical loss and atmospheric attenuation according to the measured position of the bat and the temperature and humidity in the flight room [32]. The click intensity was then taken as the maximum of these 20 intensity values. In order to calculate the beam direction, we averaged the direction of all microphones that recorded intensities of at least 0.8 of the maximum intensity or higher. This was done after smoothing the raw beam intensities with a 3rd-degree Golay-Savitzky filter [15]. Taking into account the system's noise and our beam estimation method, the error in beam-direction estimate was ∼5.5° (see ref. [15]). The inter-click angle was taken as the difference between two consecutive beam directions within a pair of clicks. The pairs are easy to recognize and can be mathematically defined as two clicks with a time-interval of less than 35 ms between them (Figure S3).
Two high-speed digital video cameras (Photron, set with a frame rate of 125 frames per second), synchronized with the ultrasonic array, were used to record the flight of the bats. The direct-linear-transform algorithm was used to measure the three-dimensional location of the bat and other objects in the room, using the two camera views.
We defined a “locked” click-pair as a pair in which the vector-average direction of its two clicks was <30° relative to the target (see example in Figure 1A; locking time is denoted by arrow). The 30° criterion was chosen since it corresponds to twice the asymptotic standard deviation of all click-pair vector averages, just before landing [15]. This is the same locking criterion as used in our previous study [15]. We tested two additional criteria for locking threshold (20° and 40°, unpublished data), which did not affect the results.
Because our microphone-array was planar, we could not estimate the absolute emission intensity (sound pressure level). In order to be sure that the differences we found in the emitted intensity were not a result of some recording artifact, we tested whether the measured intensity of echolocation clicks is correlated with several flight-trajectory parameters: (i) distance from microphones (r = −0.03; n.s.); (ii) height of flight (r = 0.02; n.s.); (iii) flight pitch (r = 0.06; n.s.). None of these parameters showed any correlation with the emission intensity. We could not control for the head's pitch angle, but an examination of the raw videos did not reveal any tendency of the bats to systematically change head pitch in an environment-dependent manner.
The sensitivity of the array could not have changed between setups because the multiple-object and the no-object experiments were interspersed in time between the one-object experiments.
To control for possible sound-occlusion effects due to the specific layout of objects in the room (e.g., the target may have blocked a specific microphone and thus may have artificially enlarged the measured inter-click angle), we re-ran the entire analysis, taking for the direction of the beam the direction of the single microphone that recorded the peak intensity (rather than weighing over several microphones). This analysis did not affect our findings. It should be noted that such an artifact is not likely for other reasons as well: (i) If the angle increase was a result of an “occlusion artifact,” the angle should have increased gradually (rather than abruptly) in the one-object experiments. (ii) If it were an artifact, we would not have observed a widening of the angle when the bat was far from the target in the pre-locked situations (“E-L” bar in Figure 2A, bottom).
In order to verify that the nets were not blocking sound waves and possibly causing some acoustic artifacts, we estimated the attenuation caused by the nets, by comparing the emission recorded from a test speaker without nets to that recorded through the nets; no difference was found for an impinging angle of 90° (i.e., when emission was perpendicular to the nets).
Because each bat produced its individual typical emission intensity and unique inter-click angle, we always normalized data from each bat separately before averaging across all bats. This means that we first calculated the average (intensity or inter-click angle) in the no-object setup and then calculated the average change relative to this value in the different setups (one-object and multiple-object) or different behavioral phases (unlocked versus locked). We next calculated the average normalized change for all bats in each of the experimental paradigms. Unless stated otherwise, all the data were normalized in comparison to the one-object condition (rather than to no-object condition), because we had almost 5 times more data-trials for the one-object experiments, which provided us with a smooth, robust baseline to compare to.
|
10.1371/journal.pcbi.1006027 | Neuronal gain modulability is determined by dendritic morphology: A computational optogenetic study | The mechanisms by which the gain of the neuronal input-output function may be modulated have been the subject of much investigation. However, little is known of the role of dendrites in neuronal gain control. New optogenetic experimental paradigms based on spatial profiles or patterns of light stimulation offer the prospect of elucidating many aspects of single cell function, including the role of dendrites in gain control. We thus developed a model to investigate how competing excitatory and inhibitory input within the dendritic arbor alters neuronal gain, incorporating kinetic models of opsins into our modeling to ensure it is experimentally testable. To investigate how different topologies of the neuronal dendritic tree affect the neuron’s input-output characteristics we generate branching geometries which replicate morphological features of most common neurons, but keep the number of branches and overall area of dendrites approximately constant. We found a relationship between a neuron’s gain modulability and its dendritic morphology, with neurons with bipolar dendrites with a moderate degree of branching being most receptive to control of the gain of their input-output relationship. The theory was then tested and confirmed on two examples of realistic neurons: 1) layer V pyramidal cells—confirming their role in neural circuits as a regulator of the gain in the circuit in addition to acting as the primary excitatory neurons, and 2) stellate cells. In addition to providing testable predictions and a novel application of dual-opsins, our model suggests that innervation of all dendritic subdomains is required for full gain modulation, revealing the importance of dendritic targeting in the generation of neuronal gain control and the functions that it subserves. Finally, our study also demonstrates that neurophysiological investigations which use direct current injection into the soma and bypass the dendrites may miss some important neuronal functions, such as gain modulation.
| New experimental techniques based on optogenetics allow neuronal activity to be manipulated with a high degree of spatial and temporal precision. This opens up new prospects for testing computational models of neuronal function, including questions such as the role of dendrites in neuronal gain control. However, compartmental models in computational neuroscience have not, until now, incorporated the kinetic models of opsins that are required in order to directly match the predictions of a computational model with observed optogenetic experimental results. Here, we introduce an approach for computational optogenetic modeling to test hypotheses, demonstrating it with application to the role of dendrites in neuronal gain control. We find that gain modulability is indicated by dendritic morphology, with pyramidal cell-like shapes optimally receptive to modulation. All dendritic subdomains are required for gain modulation—partial illumination is insufficient. Due to the simulation framework used, these results are directly testable through optogenetic experiments. Computational optogenetic models thus can be used to improve and refine experimental protocols for direct testing of theories of neural function.
| Neuronal gain modulation occurs when the sensitivity of a neuron to one input is controlled by a second input. Its role in neuronal computation has been the subject of much investigation [1–4], and its dysfunction has been implicated in a range of disorders from attention deficit disorders, through to schizophrenia, autism and epilepsy [5–8]. Neocortical neurons vary in modulability, with gain modulation having been observed in cortical pyramidal cells from layers 2/3, 5 and 6 [9, 10], whereas input-output relationships in some other cell types, such as entorhinal stellate cells, appear to be much less modulable [11]. Despite their role as the principal excitatory neuronal class within the cortex, it is unknown which properties of pyramidal cells are necessary in order to modulate their gain.
Gain modulation is signified by a change in the gradient of the input-output function of a neuron, in comparison to an overall change in excitability, which is instead evident as a lateral shift. There have been several proposed mechanisms for how a neuron alters the relationship between its input and output, including the use of shunting inhibition to shift the input-output curve [12, 13], and varying the rate of background synaptic noise, decreasing the ability of the neuron to detect target input signals [14, 15]. A subsequent theoretical study posits that both mechanisms may be necessary [16], which is supported by experimental evidence from intracellular in vivo recordings [17], indicating that these processes are not mutually exclusive and may instead operate in different regimes. Notably, theoretical studies have used point neuron models, while experimental studies have injected current into the soma. However, as in situ the processing of individual synaptic inputs occurs within the dendrites rather than somatically, this raises another possibility: that gain modulation may involve dendritic processing. The modulation of gain is affected by the balance between excitation and inhibition, and as dendrites act to integrate inputs from throughout their arbors, their capacity for mediating between attenuation and saturation is highly dependent upon the local configuration of dendritic segments and synaptic inputs [18–24]. This suggests the possibility that the morphology of the dendritic tree itself is sufficient for managing attenuation and saturation of inputs, thereby facilitating a neuron’s capacity for gain modulation.
To date, technical limitations in observing and manipulating activity at multiple locations throughout the dendritic arbor have made experimental studies of the dendritic contribution to neuronal gain control infeasible. While recording from single or a small number of dendritic locations is possible [25], this technique is not suited to manipulating activity over multiple locations, mimicking the thousands of inputs a pyramidal cell receives in vivo. However, optogenetics may prove to be a better method for manipulating dendritic activity, as light-activated opsins can be expressed throughout the entire membrane of the neuron—including the dendrites. The existence of both excitatory [26] and inhibitory opsins [27, 28] suggests the possibility of altering the balance of excitatory and inhibitory currents locally in dendrites, to act as a synthetic substitute for the effect of excitatory and inhibitory presynaptic input. This raises the prospect of a viable experimental method with which to investigate the mechanisms of neuronal gain modulation in the whole cell, as opposed to studying somatic effects alone.
Here we demonstrate through a computational model that neuronal gain modulation can be determined by cell morphology, by means of a set of dendritic morphological features which mediate between attenuation and shunting to modulate neuronal output. The local interaction of competing excitatory and inhibitory inputs is sensitive to the placement of dendritic sections. This indicates that gain modulation can be achieved by altering the overall balance of excitation and inhibition that a neuron receives, rather than being dependent on the statistical properties of the synaptic input. As experimental validation of our work would require optogenetics, we tested our hypothesis in detailed, biophysical models of opsin-transfected neurons, using experimentally fitted models of channelrhodopsin-2 (ChR2) and halorhodopsin (NpHR), which when activated produced excitatory and inhibitory photocurrents.
This study proposes a new perspective on the contribution of dendritic morphology to the characteristics of neurons, relating the shape of their dendritic arbors directly to their functional role within the neural circuit. We show that all dendritic subdomains are required for gain modulation to occur, suggesting that distinct innervation through synaptic targeting by discrete presynaptic populations could be an effective mechanism by which a neuron’s output can be quickly gated between high gain and low gain modes. By incorporating kinetic opsin models within our detailed, compartmental neuron modeling approach, we make predictions which are directly testable by optogenetic experiments. In particular, our model leads to the proposal of a new illumination protocol for a more naturalistic method of neuronal photostimulation—in which rather than simply imposing spikes or shutting down neuronal activity entirely, it is possible to increase or decrease the gain of a neuron’s response to existing inputs.
In the results, below, we use the terminology in vitro and in vivo to describe simulations. For clarity, in vitro is here taken to indicate a single point injection (as in typical patch clamp experiments), and in vivo refers to population synaptic-like input activity. In this way, we emphasize the significant difference between whether the driving input as a single point or multiple synaptic sites.
We begin with a modified ball-and-stick model, comprising of a single soma and approximately 126 dendritic sections with only passive properties. We systematically rearranged the dendrites to vary polarity via the number of primary branches np and the branching patterns via the number of sister branches at each bifurcation nb of the resulting arbor (see Methods) such that they were symmetrical around the soma (Fig 1A). The total number of branches is controlled by the number of bifurcation stages (nℓ, this number includes the creation of primary branches). This enabled us to generate 20 distinct dendritic morphologies. As the number of sections remained approximately constant, thus fixing the spatial extent of the dendritic arbor, this allowed us to identify if a neuron’s capacity for gain modulation could be determined by the dendritic arrangement, independently of total dendritic area, and if so, to establish which specific morphological features contribute. Effects of the total dendritic area have been investigated in [29].
The contribution of branching has been investigated previously, and found to attenuate both voltage signals [30] and membrane resistance [20]. To understand the interaction between neuron-wide activation (as provided by the photocurrent) and a single point input (such as current injection or presynaptic input), we evaluated first the steady-state response of the abstract models when photocurrents were induced throughout the entire dendritic arbor, before considering separately the current injection in a single distal branch.
Excitatory and inhibitory opsins (ChR2 and NpHR, respectively) were included throughout the dendritic tree in addition to the soma, and generate excitatory and inhibitory photocurrents when photoactivated. We set the opsin expression (defined by gphoto in Eq (1)) to be inversely proportional to the area of each compartment, and fixed the irradiance to be equal across the entire neuronal surface, thus ensuring the resulting photocurrent induced for each section is constant. Measuring the net photocurrent locally along the length of the dendritic arbor while shifting the ratio between NpHR and ChR2 (xNpHR) whilst keeping all other conductances constant, different dendritic morphologies summate the photocurrents such that the voltage measured at the soma is influenced by branching and polarity (Fig 1C), where morphologies with no branching show a small amplitude for the photocurrent. This relationship between net amplitude and branching was consistent across all arbor shapes we tested (Fig 1B), providing the first indication that different neuron types will sum the photocurrents differently.
Like [30], we injected current at a single point on a distal, terminating branch and measured the membrane voltage across the path between the input site and soma, along with sites at sister branches, which indicated the amount of voltage attenuation that occurred without photocurrents included (Fig 1D, note differing voltage scales). For a fixed amount of current injected on a single terminating branch, the perturbation when measured at the soma followed an identical trend as to that observed for the photocurrents in Fig 1B, due to the symmetry of the dendritic configuration, varying however in magnitude. This suggests that the magnitude of depolarization from photoactivation has to be matched to that obtained from the point input; mismatch will result in the neuron’s output being determined by the dominant term. Thus, if there is a fixed point input, this requires the amount of photoillumination to be matched to the dendritic morphology. We additionally measured the depolarization when both photoillumination and current injection were included, and found that it was a linear sum of the responses we observed separately for both types of input, as expected as there were only passive ion channels in the dendrites.
Whether the input was dendrites-wide or a single point, these dual methods of driving the neuron illustrate their respective effects: that for both methods, depolarization is largest and most effective for branched structures. For sustained whole cell photoactivation, the induced photocurrent acts to raise or lower the effective resting membrane potential, upon which the depolarization from a single (or multiple) distal point can further drive the membrane at the soma to threshold. These results also indicate that the effect of the photoactivation can dominate the neuron’s response if not matched to the relative level of activation induced by current injection at a single point.
To contrast and compare these results with the impact of the dendritic tree when the input is located at the soma, we investigated the four characteristic dendritic morphologies from Fig 1A (see also S1 Fig). This situation has been previously extensively investigated, e.g. Eyal et. al. [29], and was found that the shape and size of dendritic arbors strongly modulate the onset of action potentials by regulating the impedance load attached to the soma. The dendritic tree acts as a current sink in this case, and its impedance affects the soma depolarisation.
We then quantified the transient response by driving the neuron with spiking input, mimicking excitatory postsynaptic potential (EPSP) events included at set of locations at a terminating branch distal to the soma. By changing the rate of presynaptic events and then measuring the neuron’s firing rate we get the background firing rate when not illuminated, before repeating with irradiance (Fig 1E), we were able to measure the gain of the neuron while varying the E:I balance by changing the ratio of ChR2 to NpHR (Fig 1F). For irr = 0.02 mW/mm2, we found that gain modulation was achieved in a subset of dendritic morphologies, marked by an increase in the gradient of the response as opsin activation moved from being dominated by inhibition to excitation (Fig 1F).
To identify whether there was a consistent trend between dendritic configuration and gain modulation, we define a new measure we term the gain modulation index (M), as the relative change in gradient of the response curves (from Fig 1F) when dominated by ChR2 and NpHR respectively, i.e. M = (θChR2 − θNpHR)/θbalanced for the difference in angles for the two responses (the slope for θbalanced is approximately tan(θbalanced) ≈ 1). An M ≃0 indicates that no gain modulation occurred, whereas increasing M indicates an increasing degree of gain modulation. We found that there was a small region for which modulation was substantial, and correlated to dendritic structures that were multipolar with a small degree of branching (Fig 1G, point np = 4, nc = 2; note that a discrete set of measurements is additionally presented as a continuous colourmap for the purpose of better visualisation). Following our earlier indication that photoactivation has to be matched to dendritic structure, we measured the modulation for irradiance values an order of magnitude smaller and greater. When irr = 0.002 mW/mm2 we observed no gain modulation for all dendritic configurations (not shown), as their responses were dominated by the current injection. Increasing the irr = 0.2 mW/mm2 expanded the region of dendritic configurations which displayed gain modulation, which was now most prominent for bipolar morphologies (Fig 1H). To obtain better intuition as to how irradiance affected modulation, we charted M for our four example neurons over four magnitudes and observed a clear trend, with preferred irradiance values for which a neuron will display maximal modulation (Fig 1I). For irr = 2 mW/mm2 some configurations (e.g. (4, 1, 31)) start entering tetanic stimulation for full illumination with ChR2, for which the modulation M starts decreasing.
Following the predictions made by our abstract models, we investigated whether these principles still hold for detailed neuronal models, using a highly detailed Layer 5 pyramidal cell (Fig 2A), previously published in [31]. Its reconstructed morphology is roughly bipolar with moderate branching, which, from the abstract models we tested, demonstrates a strong capacity for gain modulability. However, the model also contained 9 additional ion channel types heterogeneously distributed throughout the soma, apical and basal dendrites. These included multiple variants of Ca2+ and Ca2+-gated channels, which introduced non-linearities as well as significantly longer time constants, which may alter the capacity of a neuron to generate spikes and thus indirectly alter its capacity for gain modulability.
To reproduce experimental tests, we began by driving our L5PC by injecting current at the soma, in a similar manner to a typical in vitro electrophysiological experiment, and compared the firing rates upon illumination against the background firing rate (Fig 2B). This revealed that IF curves were co-located (Fig 2C), indicating no gain modulation. However, this was consistent with findings from our abstract models where we observed that gain modulation was site specific for the driving input. Consequently, we moved the injection site to a distal location on an apical dendrite. This time, we observed clear changes to the gradient of the IF curve as increasing amounts of current were used to drive the cell while varying the E:I balance (Fig 2D).
While this demonstrated that the gain of this pyramidal neuron may be modulated in an in vitro scenario, neurons in situ are instead driven by thousands of excitatory and inhibitory synaptic inputs located throughout their dendritic arbor. Thus we repeated our simulation, but changed the input to mimic PSPs, by identifying 384 sites for excitatory inputs, and 96 sites for inhibitory inputs, throughout the apical and basal dendrites (Fig 2E)-two bottom panels. We observed that gain modulation was still clearly evident (Fig 2F), although the firing rates saturated for input firing rates greater than 20 Hz.
To examine the effect of dendritic morphology, we also investigated gain modulation in stellate cells, which are also present within cortical circuits, but whose morphology is very different from pyramidal cells. We used a Layer II hippocampal stellate cell model previously published by [32], based on reconstructions from [33]. Morphologically, it is multipolar with a small degree of branching (Fig 3A), which places it near to the abstract models for which we observed little to no gain modulation.
Unlike L5PCs, the response to an in vitro input of injecting current at a dendritic location (Fig 3B) revealed that stellate cells do perform divisive gain modulation (Fig 3C). In the case of current injection at the soma the same effect was observed as for the L5PC cell shown in Fig 2C: an approximately co-linear response (S2 Fig).
We then drove the cell by supplying synaptic inputs throughout the dendritic tree to mimic in vivo conditions (Fig 3D), and observed no gain modulation but rather a linear shift as xNpHR was varied (Fig 3E). This suggests that while stellate cells presumably play an important role within the neural circuit, the gain of their input-output functions is unlikely to be modulated in vivo, but are instead likely to be subject to shifts in overall excitability through changes in the amount of excitation or inhibition.
From our results, it is clear that for some neurons, such as pyramidal cells, it is possible to retune their output by applying whole-field photoactivation. However, by measuring the output firing rate, we ignore spike train structural characteristics such as the timing of the spikes. To examine how spike timing was affected by optogenetically altering the balance of excitation to inhibition, we considered a L5PC’s spike train in response to frozen noise input for an in vivo-like scenario and define a period during which we wish to increase or decrease the firing rate while the driving input remains fixed (Fig 4A). High-level illumination, which is commonly used experimentally, dramatically reshapes the spike train as the membrane potential is either completely hyperpolarized or the neuron fires with a high-frequency, regular rate (Fig 4B). Thus while this technically reprogrammes the gain of the cell, it does so by artificially rewriting the output spike times.
Instead, preserving subthreshold dynamics that arise from the hundreds of presynaptic events should allow the spiketrain characteristics to remain naturalistic, retaining rather than overwriting existing information processing functionality. To test this, we used a significantly lower level of illumination and found that the resulting spiketrains are qualitatively similar to the original response (Fig 4C). To what extent can we perturb the neuron through external photoactivation before spiketrain characteristics are destroyed?
To quantify how the intrinsic spike timing of a neuron is altered by increasing levels of optogenetic activation, we measured the interspike interval (ISI) and then calculated the coefficient of variation of the interspike interval sequence (CVISI), and the Fano factor (FF), which describes the variance of the spiketrain normalized by the mean firing rate. We compared the CVISI and FF during the period where the firing rate was altered, as both the strength of illumination and the balance of excitation to inhibition was varied, for different levels of intrinsic activity. We observed that FF (Fig 4D) and CVISI (Fig 4E) could be maintained in the same range as the unperturbed spiketrain, but that this was dependent on level of illumination, suggesting that the artificial drive has to be matched to the level of the input the neuron already receives. The best matched level was for irr = 0.002mW/mm2, which closely matched the intrinsic CVISI and FFISI values of the neuron. Using this irradiance, we observed a smooth transition from the original response as the optogenetic drive moved from NpHR-dominated to ChR2-dominated (Fig 4F).
Our findings for both biophysically detailed and abstract models demonstrate that dendritic morphology greatly contributes to determining a neuron’s capacity for gain modulability. Up until this point, we have only considered scenarios with equal illumination for every dendritic subdomain. Experimentally, however, this is not guaranteed due to unequal expression of opsins throughout the cell membrane as well as uneven light scatter as photons move through tissue. Thus we investigated how gain modulation was affected when dendritic subdomains were unequally photoactivated. Mechanistically, this is relevant for gain modulation as synaptic input to a neuron is not likely to be uniformly distributed throughout the entire dendritic tree, but may instead be organized by presynaptic origin [21, 34]. Could it be that such organization is present to allow the coordinated activation of dendritic subdomains, which is required for modifying the neuron’s output?
We began by examining partial illumination in abstract models, illuminating only one dendritic subdomain (pole) to examine how this altered gain as ChR2:NpHR was varied. We first want to pick the case with the strongest modulation gain during full illumination and we find that by looking at results in Fig 1H): that is a bipolar model np = 2, nb = 6, now illustrated in Fig 5A) and the output frequency is shown in Fig 5B). By illuminating only one pole instead (Fig 5A), we observed that partial illumination abolished gain modulation (Fig 5C). Measuring the voltage along both the illuminated and non-illuminated branches revealed that during partial illumination, the non-illuminated pole/branch acts as a current sink.
In this scenario, only 50% of branches were illuminated: perhaps gain modulation was still possible with an increased but incomplete set of dendritic subdomains? To test this, we need a multipolar abstract neuron (rather then the bipolar shown in Fig 5A) and we chose one with np = 4 (and nb = 2, nl = 5), and successively activated additional subdomains (poles) until all branches were illuminated. We found that M increased as successive poles were illuminated (Fig 5D). As this principle would hold for all dendritic morphologies, our findings demonstrate that partial illumination incapacitates gain modulation and illustrates that coordinated activation between dendritic branches is necessary for full gain modulation.
We then tested partial illumination in our detailed neuron model of a L5PC by targeting the apical dendrites, reflecting a realistic scenario in which light from a superficially located source would be more likely to penetrate the apical rather than basal dendrites (Fig 6A). Similarly to abstract models with two primary branches, we found that partial illumination abolished gain modulation in L5PC when driven by current injection at a site in the apical dendrites (Fig 6B).
A more realistic experimental scenario is one in which the likelihood of photons scattering rises with increasing depth, corresponding to a continuous gradient for the effective irradiance that decreases with distance from the surface (Fig 6C). Furthermore, as opsin activation occurs by illumination using a wavelength that is normally chosen optimally for each opsin (subject to available laser lines), and longer wavelengths proportionally penetrate distances, we examined the penetration gradients for ChR2 and NpHR independently (Fig 6D). Previously, we had only considered full-illumination of both ChR2 and NpHR with no graded illumination, which was equivalent to a gradient value of 0.0 (Fig 6D, top-left corner, purple circles). Now, by fixing xNpHR and irradiance while activating the L5PC without any additional driving stimulus, we could chart how independently varying each gradient impacted on the firing rate.
We observed three trends that are consistent with our earlier observations: (i) increasing the xNpHR factor increased the contribution of the NpHR gradient, which decreased the firing rate; (ii) a ChR2 gradient = 0.0 (signifying full ChR2 illumination) and a NpHR gradient of 1.0 resulted the largest firing rates; (iii) higher irradiance values led to higher firing rates, increasing the range of ChR2 gradients for which the neuron fired. However, we were interested in cases that correspond to the realistic scenario in which longer wavelengths penetrate through tissue further. As the preferential activation wavelengths for ChR2 and NpHR are λ = 475nm and 590nm respectively, this manifests as a bias towards NpHR-dominated regimes. Introducing a small degree of graded illumination reduced the firing rate; the neuron was further silenced by increasing the NpHR gradient from a slight bias (Fig 6E, blue circles) to a significant relative difference between ChR2 and NpHR gradients (Fig 6E, green circles).
The modulation by graded illumination was ubiquitous, although dependent on the irradiance, xNpHR value and scatter gradients for each wavelength. Experimentally, these effects can be easily overcome by prior calibration to compensate for the effects of scattering, but serve to highlight the sensitivity of a neuron to deviations from unequal innervation.
Previous work [12–14, 16] examined what input properties are required to alter the output gain of the neuron. Critically, these studies took a somatocentric viewpoint, concentrating on the output of the neuron for a given input, but bypassing the computation performed by the neuron itself. In this work, we addressed the contribution of the dendrites directly, by considering how their configuration may help or hinder modulation of the neuron’s activity and thus explain why some classes of neurons, but not others, contribute to setting the gain in a neural circuit. We established that the configuration of dendrites can affect a neuron’s capacity for gain modulability, with a centrally placed soma and a moderate amount of branching being most receptive to gain modulation. As this shape closely matched pyramidal cells, this reinforces that their role within neural circuits is to act not only as the primary excitatory neuron but also as a key element in the setting of the gain of the circuit. Thus, in addition to the influence of dendrites on firing patterns [24, 35] and their role in dendritic computation [36, 37], our results demonstrate a new aspect to dendrites that directly relates the morphological properties of an individual neuron to its functional role within a network.
We explored the relation between a neuron’s dendritic morphology and capacity to alter its firing rate by using excitatory and inhibitory photocurrents locally input to each dendritic section. The use of photocurrents, as well as making the study relate more closely to putative optogenetic validation experiments, was intended to mimic the local excitatory and inhibitory currents induced by the numerous presynaptic inputs located throughout the entire dendritic tree, with the notable difference in that while postsynaptic potentials are transient, the photocurrents we induce were typically close to steady-state. Further input was additionally applied that mimicked in vitro or in vivo input. We made no specific assumptions as to the specific type of stimulus representation of the input, such as visual contrast [38], orientation [10] or other stimulus traits; our results hold for the general case in which a driving input at discrete set of location is modulated by neuron-wide distributed drive.
Using this framework allowed us to identify that dendritic branching and the relative location of the soma were the most important morphological characteristics, as dendritic branching allowed balance between saturation and attenuation while a centrally located soma avoided it acting as a current sink. As previously established in [30], these effects become crucial when considering the compounded local input that is applied to each dendritic section (here, the non-driving input i.e., the photocurrents). For L5PC neurons, the stratified output for illumination suggests that if net drive to the dendrite is able to sufficiently cover the entire arbor, then L5PCs will be gain-modulable independent of the location of the driving input, which has been shown to govern the input-output relationship for single inputs [18, 39].
In this respect, our findings suggest that gain modulation should be achievable regardless of the specific input location. However, there is one critical caveat: that the input must be dendritic. The absolute abolition of gain modulation when the driving input was located at the soma in a pyramidal cell reinforces the role of dendrites in processing input and their contribution to modulating gain. It also highlights the difficulties associated with experimentally unraveling neuronal mechanisms which involve dendritic processing. While recording at the soma gives us an exact measure of the cells output, injecting input directly to the soma bypasses the dendrites, rendering their contribution invisible. Instead, techniques such as dendritic patching or extracellular drive are more suitable for this purpose, despite their respective technical challenge or lack of control for the number and locations of synaptic sites. The future development of holographic methods in combination with optogenetics provides potential solution to both limitations, although it is currently limited by the tradeoff between number of distinct sites that can be targeted and the frequency of their stimulation [40].
We note here that we attempted to use only passive conductances in biophysical models, in order to have better comparison with abstract neurons in interpreting the role of dendritic morphology, since their models include some active conductance (although in relatively low concentrations). However, due to their configuration and complexity we were unable to obtain normal neuronal responses i.e resting potential did not equilibrate in the experimentally observed range, due to different parameters required for the L5PC and stellate cells to operate in the passive regime.
Quantifying the effect of partial illumination revealed that gain modulation also requires the participation of all dendritic subdomains. Removing background input from dendritic subdomains resulted in the unactivated arbors becoming a current sink, and reduced the ability of the neuron to modulate gain.
Tracing studies have hinted that within pyramidal cells in sensory areas, there is synaptic targetting with feedforward presynaptic input from the thalamus tends to synapse onto basal dendrites, while input from higher cortical areas instead connects within apical dendrites [34, 41]. This arrangement of separate innervation to distinct dendritic subdomains would very easily allow for the same mechanistic process as we have observed here, whereby both feedforward and feedback connections are required for full gain modulation. Removal of one of these sources, such as the feedback input from higher cortical areas, would quickly act to shunt any background drive to corresponding dendritic subdomain, thus providing a mechanism for rapid switching between full gain modulation and no gain. As we observed that the modulation of gain approximately scales with the fraction of the dendritic subdomains that receive background driving input, the change between full gain and no gain can be most effectively controlled with two dendritic subdomains, as increasing the number of subdomains requires greater coordination between distinct input areas.
An important feature of computational models of neuronal information processing is that they be experimentally testable. Traditionally, for biophysically detailed, compartmental models of neurons, this has involved making predictions that can be confirmed by intracellular recording (whole cell patch clamp or sharp microelectrode) experiments. However, recent years have seen new optogenetic experimental paradigms come to the forefront of neuroscience, which are likely to form the basis of many experimental designs to test principles of neuronal function. Computational modeling of neuronal function should incorporate simulation of experimental predictions made by the model; whereas in the past this was largely electrophysiological, this now includes both electrophysiological and optogenetic predictions. We envisage that computational optogenetic modeling is likely to assist in bridging the gap between computational and experimental studies in areas ranging from neuroscience [42, 43] to cardiac electrophysiology [44]. For this reason, in the current study we incorporated kinetic models of opsin into the biophysically detailed neuron models described here.
Optogenetic illumination protocols in current use can generally be classified as “hard control”, in which the output of a cell is written directly by using high levels of illumination to induce either spiking or hyperpolarization [45, 46]. The problem with such approaches is that they effectively reprogram the output of the neuron, disrupting/eliminating the information processing operation that it is performing on its inputs. We suggest that a more refined method of optogenetic modulation would preserve the cell’s ability for its outputs to be affected by its inputs, but altering the gain of this input/output transformation. Our findings demonstrate the feasibility and support the development of such optogenetic control of individual neuronal gain. In this approach, using whole-field, low-level illumination allows for subthreshold dynamics to dominate, and the neuron remains driven by its presynaptic input, with the gain of its input-output function modulated by activation of a mixture of opsins. In the current work, we demonstrated that a combination of channelrhodopsin and halorhodopsin can provide a suitable opsin mix, with effect dependent upon target cell morphology.
For the general purpose of optogenetic gain control, step-function opsins (SFO) [47] and stabilized step-function opsins (SSFO) [48] may be suitable, as they do not required continuous illumination to be active, and have their suitability for loose control has already been demonstrated when driven by inputs located at the soma [47]. SFOs and SSFOs have already been proposed for use in the study of plasticity and homeostatic mechanisms during development. Our results support their suitability for application to gain modulation in vivo but also predict a restriction to their usage that will be dependent on the class of neurons to be targeted. More generally, our findings suggest that smaller, rather than larger, photocurrent amplitudes are desirable for the purpose of modulating gain.
Unequal or incomplete optical activation of the entire dendritic arbor also has significant implications for experiments that include optogenetics. We used optogenetics specifically as opsins are expressed on the surface membrane, and therefore can generate photocurrents locally within the dendrites. Experimentally, however, opsins may be non-uniformly distributed throughout the neuronal membrane, while optical point sources incompletely illuminate the entire membrane surface area, which is further compounded by scatter effects as light moves through tissue. We quantified the impact of optical scattering by examining how graded illumination alters the gain modulation curves of a L5PC, for the scenario when the scattering was equal but also for the more realistic scenario where it is unequal. For instance, ChR2 is activated at λ = 475nm, while NpHR is preferentially activated at λ = 590nm, which penetrates further through tissue. From our results, approximately equal attenuation for both wavelengths only slightly decreases the firing rate; as this imbalance increases and shifts towards longer wavelengths, we found a substantial decrease in firing rate due to this physical constraint that biases in favor of NpHR. Additionally, although ChR2 and NpHR have different peak absorption wavelengths, they both are activated over a larger range of wavelengths, and consequently there is low-level activation of NpHR at λ = 475nm, further biasing dual activation towards NpHR-dominated regimes. These effects can be experimentally compensated for by calibrating curves as the ratio of ChR2 to NpHR is varied, but require explicit measurement for individual experiments. The issue of precise co-activation of both excitatory and inhibitory opsins can be addressed by development of better dual opsins, as well as of better techniques for controled 3D illumination. Several options already exist, such as holographic spatially patterned illumination technology [49] and individually addressable LED micro-arrays [50].
More generally, identifying the impact of experimental effects is critical for the improvement and refinement of optogenetics. While optogenetics offers new possibilities for precise spatial and temporal targeting of distinct neural populations, practical hurdles such as optimally designing illumination protocols are more difficult to identify through experimental means. Additionally, as new opsin variants with differing kinetics becoming available, the task of identifying which opsin is best suited to match the intrinsic dynamics of a target neuron class becomes increasingly impractical to test. For these aims, the use of computational models of opsins will become increasingly significant [51], from the level of channel kinetics, to the level of a single neuron [52, 53], and beyond to the level of the network.
All simulations were performed in NEURON and Python [54]. The TREES toolbox [55] was used for steady-state analysis of injected current in abstract models, and NeuroTools toolbox [56] was used for spiketrain analysis. Our simulations were calibrated with experimental results for which we performed a rigorous fitting process to the experimental data (ours for ChR2 and ref. [57] for NpHR). Details of the processes can be seen on our Cloud Computing Portal—named Prometheus—which hosts all our computational tools for optogenetics (called PyRhO): http://try.projectpyrho.org/. This is a Python-based Jupyter Notebook that works in a web-browser. Within the user interface, functions for fitting parameters allow experimental data to be loaded from which parameters are extracted for 3, 4 and 6 state models. The tool offers nine different fitting algorithms to chose from, as well as post-fit optimization.
Our model of a co-activated opsin utilizes our previously published 6-state models of channelrhodopsin-2 (ChR2) [43, 52] and halorhodopsin (NpHR) [53]. A 6-state model was chosen for both ChR2 and NpHR, which includes two open states, two closed states and two inactivated states that are coupled together with by rate constants. Only the open state contributes to the generation of the photocurrent iphoto for each opsin type, which is calculated per compartment and is additionally proportional to the area A and maximal conductance for each opsin g ¯ photo in combination with two terms related to the irradiance ϕ and the membrane voltage Vm:
i photo = A g ¯ photo ψ ( t , ϕ ) f ( V m ) (1)
Critically, these models accurately capture the ion concentration kinetics and so allow accurate modelling of subthreshold dynamics, and can be tuned to provide a faithful reproduction of the temporal courses induced by opsin activation.
Throughout this work we refer to the ratio xNpHR as the relative strength of NpHR illumination in reference to a fixed value for ChR2 illumination.
Each of the abstract neuron models we created included a soma and approximately 125 dendritic sections that were arranged with varying degrees of branching. Dendritic morphology was described as the number of primary branches np, the number of sister branches nb and the number of branching stages nℓ. The total number of branches (Ntotal) is given by:
N total = n p · ( ∑ k = 0 n l - 1 n b k ) . (2)
Altogether, we generated 31 different dendritic configurations that, despite their geometrical configuration, had approximately equal surface areas and volumes. Dendritic tapering was excluded to conserve total surface area.
For the biophysical properties, the values for the soma were C = 1pF, diameter = 10μm and length L = 10μm, while each dendritic section had parameters diameter = 0.4μm, C = 2pF and length L = 50μm. All sections and soma had passive membrane properties (g_pass = 0.00005, e_pass = -75mV) while the soma additionally had Hodgkin-Huxley channels, the NEURON built-in hh mechanisms were used, with conductances: gnabar_hh = 0.25, gkbar_hh = 0.1, gl_hh = 0.000166, el_hh = -60mV, ek = -70mV. Each soma and dendritic section had ChR2 and NpHR models inserted, with constant expression throughout the dendrites and soma.
Constant driving input for the steady state response was modeled by injecting constant current in the last segment of a single distal segment, and normalizing the distance from total length to soma. Synaptic locations were chosen randomly from all distal sections. Synapses themselves were modeled using NEURON’s ExpSyn model, with input spiketimes drawn from independent Poisson process. As different dendritic morphologies had different electrotonic distances from synapse location to soma, synaptic weights were chosen where possible such that the output firing rate was approximately equal to the input firing rate, enforcing a loose version of synaptic democracy. For some arbor configurations, the length from soma to distal dendrites was greater than the electrotonic distance and a transient response could not be obtained. The dendrite tapering was not included in our simulation for abstract neurons, but it is included for two realistic biophysical neuron models. Inclusion of tapering for abstract neurons creates a problem of keeping approximately the same total dendritic tree area for various morphologies, which was the more relevant factor with respect to our results.
|
10.1371/journal.pcbi.0030144 | Prediction of Gene Expression in Embryonic Structures of Drosophila melanogaster | Understanding how sets of genes are coordinately regulated in space and time to generate the diversity of cell types that characterise complex metazoans is a major challenge in modern biology. The use of high-throughput approaches, such as large-scale in situ hybridisation and genome-wide expression profiling via DNA microarrays, is beginning to provide insights into the complexities of development. However, in many organisms the collection and annotation of comprehensive in situ localisation data is a difficult and time-consuming task. Here, we present a widely applicable computational approach, integrating developmental time-course microarray data with annotated in situ hybridisation studies, that facilitates the de novo prediction of tissue-specific expression for genes that have no in vivo gene expression localisation data available. Using a classification approach, trained with data from microarray and in situ hybridisation studies of gene expression during Drosophila embryonic development, we made a set of predictions on the tissue-specific expression of Drosophila genes that have not been systematically characterised by in situ hybridisation experiments. The reliability of our predictions is confirmed by literature-derived annotations in FlyBase, by overrepresentation of Gene Ontology biological process annotations, and, in a selected set, by detailed gene-specific studies from the literature. Our novel organism-independent method will be of considerable utility in enriching the annotation of gene function and expression in complex multicellular organisms.
| The task of deciphering the complex transcriptional regulatory networks controlling development is one of the major current challenges for molecular biology. The problem is difficult, if not impossible, to solve without a detailed knowledge of the spatiotemporal dynamics of gene expression. Thus, to understand development, we need to identify and functionally characterize all players in regulatory networks. Data on gene expression dynamics obtained from whole transcriptome microarray experiments, combined with in situ hybridization mRNA localisation patterns for a subset of genes, may provide a route for predicting the localisation of gene expression for those genes for which in situ data has not been generated, as well as suggesting functional information for uncharacterised genes. Here, we report the development of one of the first methods for predicting the localisation of gene expression during Drosophila embryogenesis from microarray data. Pooling the subset of genes in the fly genome with in situ data to form functional units, localised in space and time for relevant developmental processes, facilitates the statement of a classification problem, which we address with machine-learning methods. Our approach promotes a richer annotation of biological function for genes in the absence of costly and time-consuming experimental analysis.
| As a result of a gradual developmental strategy known as epigenesis, an embryo comprising a few cell types is refined to generate a complex organism composed of many precisely organized anatomical structures. Understanding how the genome is dynamically deployed to generate such cellular diversity is a key challenge in modern biology. Spatiotemporal information on gene expression can provide insights into the biological function of gene products, since genes belonging to the same developmental pathway tend to have similar or correlated expression patterns [1–3]. Until recently, research efforts aimed at deciphering the spatiotemporal dynamics of gene expression have been primarily carried out on a small scale, since they were predefined by either a specific gene network or a single gene of interest.
Advances in molecular tagging and imaging techniques, along with high-throughput experimental methods, offer new possibilities for following developmental processes by investigating the spatiotemporal dynamics of gene expression and helping to reveal gene function at a whole genome scale. High-throughput experimental approaches, such as DNA microarrays and mRNA in situ hybridization, have been used to probe gene expression during development in several model organisms [4–8], including the fruit fly Drosophila melanogaster [9]. Microarray studies, using mRNA samples extracted at various stages of embryonic development, provide time courses of expression for individual genes, while in situ hybridization with specific gene probes identifies spatial and tissue-specific expression of particular target genes. Of these, analysis of the former is mostly automatic, while the latter requires manual annotation by experts. Whole-organism microarrays provide semiquantitative information on changes in gene expression levels during the course of development; however, it is difficult to infer spatial information about gene expression since the technique generally uses RNA extracted from whole embryos. Integrative analysis of in situ expression patterns and microarray gene expression data is one possible way to assist in deciphering the roles that genes play during development and to identify sets of genes involved in similar developmental processes. Data on the dynamics of gene expression obtained with whole genome microarray experiments combined with in situ expression patterns may thus provide a starting point for the prediction of gene expression localization and assignment of putative function for those genes for which in situ data has not been yet generated.
In this report, we describe the development of a computational method for predicting the spatial localization of gene expression from microarray data. Our method provides a route for elucidating the roles a gene may play during development by inference from the spatial annotation of its expression [9]. Our approach, based on a machine-learning framework, allows de novo prediction of gene expression localization by training a classifier on a subset of genes with spatial expression patterns annotated by in situ hybridisation experiments. While a variety of clustering algorithms are popular for the analysis of microarray data, their ability to infer gene function or expression localisation is limited. An alternative method of meta-analysis, classification, has been widely used in the context of diagnostics; i.e., separating patients with a disease from the normal population, (see, for example, [10]), or identifying conserved modules in genetic networks [11]. The problem statement in our case resembles the work of Wong et al. [12] and Brown et al. [13], where the problems of predicting gene interactions and of inferring gene function from high-throughput data, respectively, have been formulated as classification tasks.
To develop our approach, we focused on a system—the development of the Drosophila embryo—that has substantial high-quality in situ hybridization data and comprehensive microarray time-course data available. Starting from the anatomical annotation of in situ gene expression patterns obtained by the Berkeley Drosophila Genome Project (BDGP; http://www.fruitfly.org) [9], we assembled genes involved in specific developmental process into groups we define as functional units. We then trained a machine-learning algorithm with microarray data to discriminate between the genes associated with a particular developmental process and the remainder of the genes that are not. The most suitable classifiers, along with a set of functional units producing best prediction results, were selected via a multilevel verification procedure. De novo predictions of tissue-specific expression for genes with only microarray data available were confirmed with literature data and with Gene Ontology (GO):Biological Process annotation. Our method provides a generalized route for generating preliminary functional annotations for genes of unknown function.
To predict gene expression localization, we selected several developmental events taking place at different time intervals during embryogenesis and assembled genes acting in these processes into functional units. We define a functional unit as a set of genes, known to be involved in a particular biological process during a contiguous developmental time interval, expressed in a predefined group of anatomical structures related by developmental lineage. A gene is assigned to the functional unit if, and only if, it is expressed in all anatomical structures in a particular lineage of interest. Thus, a functional unit not only reflects the spatiotemporal dynamics of the expression of a set of genes involved in the biological process under study, but also suggests that a set of genes making up a unit act concordantly in a specified event in organogenesis (see Materials and Methods).
As an example of this approach, we consider assembling genes into a functional unit that reflects the development of the central nervous system (CNS). Formation of the individual tissues of the CNS from their primordial embryonic structures takes place during late embryogenesis, in the 6-h interval spanning stages 9 to 15 of Drosophila development [14]. This is a highly ordered process, which can be presented schematically in three steps [15,16]. The primordium of the CNS is established at stage 9 when neuroblasts, which originate from a precisely defined area of the neurogenic ectoderm, delaminate from the ectoderm into the embryo. Neuroblasts are large cells with stem cell properties whose progeny will populate the differentiated CNS. Neuronal differentiation begins at stage 13, with the formation of neurons and initiation of axonal outgrowth, and shortly after this, at stage 14, the ventral nerve cord condenses [14]. Therefore, functional units corresponding to CNS development are compiled from the genes that are expressed in the ventral neuroderm anlage and the anatomical structures that derive from it: ventral nerve cord primordium and lateral cord. Every gene annotated as being expressed in all three of these anatomical structures in the BDGP in situ database is attributed to the vna2lcord functional unit.
We constructed and examined a total of 15 functional units encompassing three developmental processes occurring in late embryogenesis, one process occurring in mid-embryogenesis, and one process occurring during early development (see Table 1 for the units considered and Materials and Methods for the construction schema). The genes in the functional units are involved in the formation of the procephalic ectoderm primordium, the development of the mesoderm, and the precursors of the embryonic muscle system, embryonic CNS, and, finally, the embryonic digestive system. The functional units are labeled by an abbreviation that reflects the morphological changes in the organism during embryogenesis (i.e., the initial developmental intermediate structure and the terminal differentiated anatomical structure used to assemble a functional unit; Table 1). To associate microarray data with these functional units, we divided a time course of gene expression [9] into three alternating time windows corresponding to early, middle, and late embryogenesis.
Average microarray profiles constructed for selected functional units (see Figures 1 and S1) show changes in the dynamics of microarray gene expression data attributed to different units. The variability in microarray expression signals decreases as the number of anatomical structures a functional unit is assembled with increases. The expression patterns stored in the BDGP database are very rarely identical; however, there is still a noticeable similarity among expression patterns for many genes [9]. Indeed, the anatomical annotations for genes comprising functional units are not homogeneous and display remarkable diversity. Figures S10 to S14 and Tables S14 to S18 show the variability in annotations of expression patterns for genes attributed to the five most complex functional units. Anatomical annotations for functional units encompassing developmental processes that are spaced several stages apart show very little similarity. In contrast, annotations for the two functional units (pep2ebrain and vna2lcord) assembled from genes involved in CNS development during late embryogenesis are similar. Furthermore, although genes are rarely attributed to multiple functional units, these two units are exceptional and share many genes (see Figure S15 and Figure S17A and Table S13).
The developmental processes we selected are major events in Drosophila embryogenesis, and consequently have been well-characterized by many researchers. This is important since, in order to verify predictions of tissue-specific expression and to estimate the performance of the proposed classification scheme, we needed to choose extensively annotated and studied events during organogenesis. In addition, the choice of these processes was partially predefined by the nature of the in situ dataset available from the BDGP (i.e., by the number of genes involved in each process and by the time span of the developmental processes; see Materials and Methods for details). Furthermore, to demonstrate the flexibility of the method, we selected processes from diverse time intervals during Drosophila development, specifically focusing on overlapping developmental processes during late embryogenesis, to determine whether our classification approach is able to separate sets of genes with different tissue-specific annotations but very similar microarray gene expression profiles.
The formulation of our classification scheme uses microarray data as input features and derives class labels as to whether a gene belongs to a functional unit or not (see Figure S2). Because the number of training examples was low, with a few dozen genes in a functional unit being considered, classifier design had to be performed with care to avoid overfitting by classifiers that form complex nonlinear boundaries. We achieved this by extensive cross-validation and bootstrapping to set classifier parameters. Classifiers were evaluated on out-of-sample data (test data) for which BDGP in situ hybridization images exist, and a selected classifier was used in subsequent de novo predictions on genes for which only microarray expression profiles exist. The gene expression profiles utilised for training the classification method were excluded from the test set subsequently used for de novo prediction of gene expression localisation.
The discriminability in the data is higher in functional units with a higher number of anatomical structures in any developmental lineage (i.e., recognition rates for genes in vna2lcord are higher than those for vncp2lcord), reaching a peak performance of 80% sensitivity and specificity rates (see Figure 2 and Table S3). Therefore, the support vector machine (SVM) classifiers obtained with the five most complex functional units, namely mat2pep, aep2egut, pep2ebrain, tma2smusclep, and vna2lcord (Table 1), were used as base classifiers to generate de novo localization predictions from microarray data.
We carried out a series of verification tests on the results of de novo predictions checking against: (1) literature curations from FlyBase [17]; (2) GO:Biological Process annotations from FlyBase; and (3) manual literature examination for a small number of confident predictions.
FlyBase has curated literature data on the spatiotemporal localization of gene expression for approximately 3,912 gene transcripts. Importantly, the FlyBase annotations are independent from the BDGP in situ hybridization annotations. In FlyBase, expression patterns are annotated via manual curation of published papers. Consequently, the annotation of expression patterns is diverse and both less systematic and less specific than those in the BDGP database. For example, the term “lateral cord” is widely used for annotating BDGP data, but it is never used in FlyBase. The Drosophila gross anatomy ontology shows that lateral cord is a “part of” the ventral nerve cord; therefore, the term “ventral nerve cord” is used in our verification analysis. Similarly, if an anatomical structure is used by the BDGP curators but does not have an exact match in FlyBase, we looked either for a higher-level anatomical structure or for a biological process in which the organ is involved. For example, according to anatomy ontology, embryonic brain is a “part of” the procephalic neurogenic region. Therefore, genes for which there is FlyBase curated literature evidence of expression in the procephalic neurogenic region, protocerebrum primordium, and procephalic ectoderm anlage may be used to verify predictions for genes in the pep2ebrain functional unit.
The results of the prediction verifications are summarized in Figure 3A. Two functional units, mat2pep and vna2lcord, which encompass the formation of embryonic central brain precursors in early embryogenesis and development of the embryonic CNS, show the best overall literature verification scores of 74% and 72%, respectively. The lowest score is seen with the tma2smusclep functional unit (57%). The reason for this low score appears to be a significant disagreement in controlled vocabularies for these tissues between FlyBase and the BDGP databases. For the remaining two functional units considered, the enrichment in confirmed de novo predictions exceeds 60%. Since some tissues are more closely studied than others, calculating the fraction of false-positive predictions is difficult because failure to find annotation support for a prediction may simply reflect the fact that a given gene has yet to have its expression characterised.
The number of genes annotated with FlyBase literature data is small for some of the functional units considered (e.g., mat2pep and pep2ebrain); hence, only a small proportion of the de novo predictions can be validated by this route. However, since each of the functional units corresponds to a specific developmental process, we performed additional verifications using GO terms [18] drawn from the “Biological Process” category. A total of 14,332 proteins have GO annotations in FlyBase; of these, 2,794 have a GO:Biological Process term [19]. Fortunately, the majority of the GO:Biological Process terms are annotated by database curators rather than inferred by electronic annotation, and are thus more likely to be accurate [18]. We therefore used the GeneMerge software tool [20], which identifies and ranks GO terms that are statistically overrepresented with respect to a background distribution calculated from all genes with annotated GO:Biological Process terms. The statistical significance cutoff is set at a corrected e-score value of 0.05.
On the whole, the data on GO:Biological Process terms representation confirm the prediction results obtained with our classification method, providing support for predictions that have no FlyBase literature annotation (Table 2). For example, the GO:Biological Process annotation for the mat2pep functional unit assembled with genes controlling early development is enriched for relevant GO terms, such as blastoderm segmentation, formation of embryonic brain, and CNS precursors (Table 2). However, our analysis revealed several functional units, such as aep2egut, for which no overrepresented GO terms were found. FlyBase's QueryBuilder tool (http://flybase.org/cgi-bin/qbgui.fr.html), a user interface that supports powerful searches by offering access to every data field in FlyBase, provides support for the suggestion that there is a deficiency in annotations for the development of the Drosophila digestive tract. According to FlyBase's GO:Biological Process annotation, the number of genes known to be involved in endoderm development is 16, while the number of genes known to control midgut development is only 25. We suggest that a paucity of biological knowledge on the development of the digestive system reflects this in poor GO annotation, even for genes where either in situ hybridization data or literature data are available (i.e., in the groups of genes forming the training set and confirmed de novo prediction set, respectively; Table S1).
As shown in Figure 3B, the GO annotation provides evidence supporting approximately 40% to 70% of the de novo predictions obtained with our method, depending on the functional unit in question. For the remainder, the genes are either not annotated with a GO term, or the predicted expression localisation is not correct. Although microarray profiles for the genes in the selected functional units are separable with our classifier model, many genes are located in the immediate vicinity of the hyperplane, separating positive and negative predictions (Figure S3). A consequence of the geometric structure of the SVM classifiers is that the prediction confidence increases with the distance of the gene from the hyperplane in multidimensional space. Hence, for every functional unit, we selected the highest-scoring 30% (i.e., those predictions furthest from the hyperplane) and reanalysed this set with GeneMerge. As expected (Table 2), for the majority of the functional units, the GO:Biological Process annotation becomes more specific. For example, genes in the pep2ebrain functional unit are enriched with the dendrite morphogenesis GO term, a lower-level GO term compared with brain development. However, for the top predictions in the vna2lcord functional unit, no overrepresented GO terms were found. Manual inspection of the GO annotations for these 125 genes (Table S11) indicates that 24% of the predictions are supported with GO:Biological Process terms related to CNS development. Thus, the GO:Biological Process annotation provides additional evidence to support the validity of our method for predicting spatial localization of gene expression.
De novo predictions of localization of expression for the functional units of interest obtained with SVM classifiers can be found in Tables S8–S12. As with the training sets, only a small number of genes associated with de novo localisation predictions belong to multiple functional units (see Figures S16 and S17B). However, the number of genes that, according to de novo predictions are simultaneously attributed to different functional units in late embryogenesis, increases. A surprisingly large overlap in shared de novo predictions is found for genes in the aep2egut and pep2ebrain functional units (see Table S13). An analysis of the shared genes using GeneMerge revealed significant enrichment for three GO terms: epidermal growth factor receptor signaling pathway, peripheral nervous system development, and plasma membrane. Unfortunately, a paucity of literature or GO annotation data on expression patterns for most of these genes prevents us from coming to an unambiguous conclusion on their role in development. However, we suggest that genes from this set may be largely responsible for cell proliferation, cell migration, cell adhesion, and attachment to the extracellular matrix during the development of both the nervous system and midgut.
Undoubtedly, an in situ hybridisation is the most reliable test for any prediction of expression localization. We therefore searched the literature for published evidence to support our predictions. Thorough analysis of available literature data, electronic resources, and BDGP in situ experiments produced a list of 38 genes for which the expression localization is very likely to be correctly predicted (Tables S4–S7). In some of these cases, the literature evidence in support of the expression prediction confirms the localization of expression in a higher-level or lower-level anatomical structure. For some of the genes in this list, a BDGP in situ experiment exists; however, due to the problems with the annotations, these genes have not been used in our classifier training set. Recovering such genes further supports the utility of our prediction method.
For example, according to our prediction, selenide,water dikinase (SelD) is expressed in anatomical structures that form the digestive system, including the embryonic midgut && anterior midgut primordium && anterior endoderm primordium (aep2egut functional unit). Persson et al. [21] report SelD expression in the endodermal anlagen, midgut primordium, and in the gastric caecum. The latter structure is a “part of” the embryonic midgut, while the first is a precursor of the anterior endoderm primordium. Therefore, our predicted expression localisation is confirmed.
In the case of chiffon (chif), we predict expression in the ventral neuroderm anlage && ventral nerve cord primordium && lateral cord functional unit (vna2lcord). The BDGP flagged their in situ experiment for chif as ubiquitous expression; as a consequence, chif was not part of the training set used for classification. However, the BDGP in situ images are annotated, and clearly show elevated chif expression in the ventral nerve cord primordium, the ventral nerve cord, and the ventral ectoderm anlage. Since the BDGP-controlled vocabulary annotates the lateral cord as a “part of” the ventral nerve cord, our prediction of chif localization is consistent with the in situ data.
Pendulin (Pen) is predicted to belong to the mat2pep functional unit, and, in agreement with this, Torok et al. [22] report maternal expression as well as zygotic expression at blastoderm and in precephalic, cephalic, and ventral neuroectoderms. Similarly, we predict that Minichromosome maintainence 7 (Mcm7) belongs to the mat2pep functional unit; maternal expression is reported by Ohno et al. [23], while the evidence of strong expression in the embryonic central brain and CNS (Feger et al. [24]) implies early ubiquitous expression in precephalic, cephalic, and trunk neuroectoderms.
wingless (wg) expression is predicted in the embryonic central brain and its developmental precursors (pep2ebrain). Although the in situ experiment for wg exists in the BDGP database, the expression in embryonic central brain is not annotated. Therefore, this gene was not used in training of the SVM classifier and can be used to verify prediction results. Baker et al. [25] found wg to be expressed in the procephalic lobe, while Shmidt-Ott et al. [26] detected wg expression in the antennal, labral, and intercalary segments of the embryonic brain, thus supporting the prediction results obtained.
These observations support the utility of our predictions for particular developmental processes. However, we also noticed that some genes are predicted to be members of more than one functional unit. For example, homothorax (hth), spitz (spi), and bangles and beads (bnb) are associated with the mat2pep and pep2ebrain functional units, predictions supported by the published literature (see Tables S4, S6, S8, and S10). These functional units encompass a set of anatomical structures that reflects 16 stages of Drosophila embroyogenesis from fertilization to the differentiated embryonic central brain. This suggests that classification experiments designed to predict expression in a wider set of anatomical structures, reflecting developmental lineage from early developmental intermediates to the terminally differentiated anatomical structure, may also be successful. Furthermore, the ability to classify genes according to functional units overcomes the problems of divisive or agglomerative clustering approaches that force genes into a single cluster.
Finally, we note that in the case of Semaphorin-2a (Sema-2a), published in situ expression data by Kolodkin et al. [27] contradicts the BDGP in situ experiment. Kolodkin et al. report Sema-2a expression beginning at stage 10 and localised primarily in the developing CNS. On the other hand, the BDGP database reports maternally derived expression at the cellular blastoderm but not in the CNS. Our prediction indicates that Sema-2a belongs to the mat2pep functional unit, supporting the BDGP in situ experiment, and implies localisation of expression in precursors of the developing embryonic central brain, a “part of” the CNS. Again, since the BDGP annotation at later stages of development is ubiquitous, this gene was not used in our training set.
By formulating a supervised learning framework, we have been able to predict tissue-specific gene expression in some sets of anatomical structures with high accuracy, achieving 80% sensitivity and specificity rates for the sets of anatomical structures we selected.
Our de novo predictions were verified using curated literature data, GO: Biological Process annotations, and published experimental reports. In the case of genes for which annotated in situ expression patterns are available, we achieve a true positive rate of 60%–70% for most of the functional units considered. In addition, we observed clear enrichments for annotations that are assigned to morphogenetic events arising from the processes governed by the products of genes in the functional unit. Finally, prediction results obtained for genes such as chif, SelD, Pen, Mcm7, wg, and an additional 33 genes were verified with published in situ hybridization experiments. As we described for the case of Sema-2a, our method may also have utility for automatically improving existing annotations in the BDGP or FlyBase databases by detecting potential anomalies or annotation conflicts.
The approach we present in this study allows us to combine qualitative information on gene expression from in situ hybridisation studies or gross anatomy ontologies with semiquantitative descriptions of gene expression dynamics obtained from microarray experiments to identify the tissue localization of gene expression on a large scale. We further note that the classification approach we apply here is superior to other, more frequently used, methods of function prediction from microarray gene expression data, such as cluster analysis or principal component analysis (see Figure S9 and a discussion on results of the benchmarking study on discriminant versus projection methods accompanying this Figure). This is because we are able to use prior knowledge in the form of class labels associated with functional units and infer complex correlations from the microarray measurements [13,28].
The classification approach proposed here is subject to two distinct limitations. First, superior performance results are obtained for genes that are characterized with complex continuous microarray gene expression profiles. Not only are these expression signatures better recognized by our classification method, but also the functional units assembled with such profiles are more homogeneous, facilitating faster classifier training and enhancing the performance of the method. In addition, the predictive power of the method and the range of developmental process for which the localisation of gene expression are possible depend on the time resolution of the microarray experiment (i.e., number of microarray time points available for analysis). Second, many high-throughput in situ experiments are accompanied by high-quality matching microarray datasets; however, this is not always the case. If this analysis were to be repeated for another organism for which spatial expression patterns are documented by in situ photomicrographs only, there may be complications due to data integration issues. For example, synchronization of microarray expression data obtained with different experimental platforms, blending time points, and corresponding expression values from various time courses, etc.
Recognising these limitations, however, our prediction method is versatile and can be applied to any microarray and in situ hybridization data regardless of the organism and biological process under study. The increasing efforts aimed at cataloging the spatiotemporal patterns of gene expression in vertebrates (i.e., Neidlhard et al. [29] and Yoshikawa et al. [30]), where the collection and analysis of in situ data are more difficult than with Drosophila, combined with the increasing quantities of available gene expression data, suggest an immediate application for our classification method. The in situ expression patterns, however, should be exhaustively characterized with a list of features in order to facilitate the classification.
In summary, we believe our tool offers a flexible method for increasing the utility of high-throughput gene expression data. The ability to combine different data modalities for integrative data mining is becoming increasingly important as we seek to translate the information encoded in genome sequence into biological function. By facilitating improved functional annotation of transcription units, our method provides an important way of adding value to expression data without the need for expensive experimental approaches.
Whole-genome Affymetrix microarray developmental time-course data corresponding to 11,904 genes and in situ hybridization data of 2,500 experiments originating from BDGP were obtained from ftp.fruitfly.org. With the in situ Expression Patterns database release used here (Release 2, 5 April 2004), experiments documented as “junk,” “no staining,” “production problem,” “too weak,” “ubiquitous,” and “maternal” were removed, reducing the set to 1,875 experiments corresponding to 1,565 unique microarray gene expression profiles. The in situ experimental data was annotated using a controlled vocabulary for anatomical structures with photomicrographs ordered according to stages of development. The BDGP annotations were collected by visual inspection of the images, resulting in six temporal classes corresponding to groups of developmental stages.
The whole-genome microarray data (ArrayExpress accession E-RUBN-2 [31]) were produced by BDGP. RNA samples were prepared from whole-embryo homogenates and cDNA-hybridized to Affymetrix GeneChip Drosophila Genome Arrays (http://www.affymetrix.com) using standard protocols and equipment. The samples were taken every hour, starting from 30 min after fertilization. The scanned array images were analyzed by Tomancak et al. [9] using the RMA algorithm [32] as implemented in the open source package Bioconductor [33]. In total, there were 36 array scans comprising three independent replicate time series for each gene. The expression levels were averaged among the replicates to obtain final expression values for 12 time points that correspond to 15 developmental stages.
Construction of functional units was not automated due to differentiation of tissues and organs, uncertainties in annotation of in situ data, and technical issues arising with the use of microarrays in developmental biology. Identifying correlation dependencies between microarray gene expression profiles and in situ hybridization staining patterns was not straightforward for numerous reasons (see Tomancak et al. [9] for a detailed discussion). The in situ staining used in BDGP experiments is performed with an enzymatic reaction; therefore, the staining intensity is dependent not only on the strength of the probe, but also on the amount of time the color reaction is developed. Short staining times may result in weak ubiquitous expression that may not be detected in a whole-organism microarray experiment. In addition, the total amount of RNA produced by small anatomical structures may not be sufficient to be detected as an expression level change in a whole-organism microarray. Therefore, a selection of anatomical structures in which microarray experiments are expected to be capable of detecting gene expression changes were obtained by visual inspection of in situ photomicrographs and corresponding microarray profiles.
Each of the above factors contribute to the small number of genes that are known to be involved in a specific developmental process of interest and in which variations in annotation of expression patterns are accompanied by changes in expression dynamics obtained in a microarray experiment. To choose a set of anatomical structures to assemble a functional unit, we selected microarray gene expression profiles that exhibit fluctuations during continuous intervals in the course of development (six to eight consecutive time points on the microarray time course; see Table 1 and Figures 4 and S2) and identified structures that display strong staining on the in situ photomicrograph and therefore are likely to contribute to the expression signal. A functional unit is assembled with a subset of structures linked by morphology and involved in a developmental process of interest.
Functional units were constructed starting from the cluster analysis reported by Tomancak et al. [9], which identified sets of anatomical structures involved in key biological processes (see Figure 5A). To perform a clustering analysis with the in situ data, Tomancak et al. transformed the textual annotation of expression patterns into a binary matrix. Each gene in this matrix is annotated with respect to the presence or absence of expression with the in situ hybridization study. The output of the clustering analysis revealed developmental processes that are widely represented in the in situ dataset along with lists of genes associated with each process. Denoting a group of anatomical structures by St =
,…,
(Figure 5, shown in blue) and the corresponding sets of genes involved in the development of these structures by G =
: (1) from the Drosophila gross anatomy ontology (http://obo.sourceforge.net/browse.html), we obtained the morphogenetic hierarchy of anatomical structures (Figure 5B); (2) by visual inspection of the in situ expression patterns and corresponding microarray gene expression profiles, we selected anatomical structures that were considered capable of producing sufficient amount of mRNA for reliable microarray measurements; (3) anatomical structures linked by “part of” or “develops from” relationships in the gross anatomy ontology were assembled into lists
&&
&&
; and (4) to facilitate reliable training of the learning methods, we restricted ourselves to developmental processes for which there are a reasonable number (at least 30) of genes known to be involved.
Functional units are the sets of genes that satisfy the above criteria.
SVMs are powerful class prediction techniques that have been used in a range of biological applications, including microarray data analysis problems [13,34–37]. Given a functional unit, we designed SVM classifiers in the space of microarray gene expression data (Figure 4B) using the implementation SVMlight [38]. We used polynomial and radial basis function kernels to construct classifiers and optimized kernel width and cost-insensitive margin parameters in a cross-validation loop (Figure 4B) to maximize the F1 measure [39]. This is an appropriate objective function given a high imbalance between the numbers in each class. Data was partitioned at random into 50%–25%–25% for training, validation, and test purposes: the first 50% for optimizing the SVM weight parameters, the second 25% for cross-validation to select kernel width and margin parameters, and the final 25% to draw receiver operating characteristic (ROC) curves quantifying the performance of each classifier.
For many years now, since the pioneering works of Swets [40,41], ROC curves and the area under the ROC curve measure have been used to evaluate the performance of machine learning methods. We averaged the ROC curves of all bootstrap partitions to find a reliable operating point. Following Flach [42], we used vertical averaging of the ROC curves in 20 bins along the false-positive axis and used cubic spline interpolation to obtain an average ROC curve [43–46]. An operating point on this average curve was chosen by the projection method [47]. From among all the classifiers we designed, we selected the classifier closest to this operating point to perform de novo predictions.
With the resulting classifiers, we were able to identify genes that belonged to the defined functional units at a high level of accuracy. Discrimination achieved between genes in a functional unit and the remaining genes [48] is shown in the receiver operating characteristics curves in Figure 2. To confirm that the data distributions require nonlinear class boundaries achieved by SVMs, we also designed linear classifiers by the Fisher linear discriminant analysis (FLDA) method [49] and found significant gains in performance of the SVM classifiers in terms of the areas under the ROC curves. The result of comparison of the areas under the ROC curves measures either for the whole range of specificities or for the selected interval of specificities [43,44] and determines the classifier that discriminates in the given data space in the best possible way. The optimal point of the ROC curve that gives the best parameters of the classifier is calculated according to a standard projection method as described in [47]. The FLDA classifiers generally had lower average performances, but also showed lower variability across different bootstrap partitions of the data (Figures S4–S8).
We have verified de novo predictions generated by the classifier using both literature data curated by FlyBase and functional GO annotations. The verification is constrained by two factors: (1) the comparatively small number of genes with expression patterns annotated by FlyBase curators, and (2) possible ambiguities in GO annotations. Apart from these considerations, there is an inherent asymmetry in the functional annotation of genes. When a gene is reported to be expressed in a certain set of tissues, we may be confident that this is true. However, when such an association is absent, it may simply reflect the fact that experiments have not been conducted or the results have not been reported in the relevant databases. We thus estimate true-positive and false-positive rates for the de novo localisation predictions obtained with our classifier using the operating point of the ROC curves constructed to evaluate performance with the BDGP data (see Figures 2, S4–S8). In this case, both microarray and anatomical annotation data are available. These estimates are reliable given the assumption that the genes annotated in the BDGP in situ hybridization data are a representative sample of the Drosophila genome. The robustness of the classification model has been evaluated in extensive cross-validation and bootstrapping steps, rather than assessed on a specific pair of training and test sets; therefore, we expect the extrapolations to be reliable. The true-positive and false-positive rates are in the range of 70%–85% and 15%–30%, respectively, depending on the functional unit in question.
The Gene Ontology (http://www.geneontology.org) terms from the Biological Process category discussed in this paper are assigned with the following accession numbers: blastoderm segmentation (GO:0007350), endoderm development (GO:0007492), midgut development (GO:0007494), dendrite morphogenesis (GO:0048813), brain development (GO:0007420), epidermal growth factor receptor signaling pathway (GO: 0007173), peripheral nervous system development (GO:0007422). The Gene Ontology accession number for the term “plasma membrane” from the Cellular Component category is GO:005886.
The FlyBase (http://www.flybase.net) accession numbers for the genes discussed in this paper are SelD (FBgn0020615), chif (FBgn0000307), Pen (FBgn0011823), Mcm7 (FBgn0020633), wg (FBgn0004009), hth (FBgn0001235), spi (FBgn0005672), bnb (FBgn0001090), and Sema-2a (FBgn0011260). |
10.1371/journal.pgen.1004252 | Synergistic Interactions between the Molecular and Neuronal Circadian Networks Drive Robust Behavioral Circadian Rhythms in Drosophila melanogaster | Most organisms use 24-hr circadian clocks to keep temporal order and anticipate daily environmental changes. In Drosophila melanogaster CLOCK (CLK) and CYCLE (CYC) initiates the circadian system by promoting rhythmic transcription of hundreds of genes. However, it is still not clear whether high amplitude transcriptional oscillations are essential for circadian timekeeping. In order to address this issue, we generated flies in which the amplitude of CLK-driven transcription can be reduced partially (approx. 60%) or strongly (90%) without affecting the average levels of CLK-target genes. The impaired transcriptional oscillations lead to low amplitude protein oscillations that were not sufficient to drive outputs of peripheral oscillators. However, circadian rhythms in locomotor activity were resistant to partial reduction in transcriptional and protein oscillations. We found that the resilience of the brain oscillator is depending on the neuronal communication among circadian neurons in the brain. Indeed, the capacity of the brain oscillator to overcome low amplitude transcriptional oscillations depends on the action of the neuropeptide PDF and on the pdf-expressing cells having equal or higher amplitude of molecular rhythms than the rest of the circadian neuronal groups in the fly brain. Therefore, our work reveals the importance of high amplitude transcriptional oscillations for cell-autonomous circadian timekeeping. Moreover, we demonstrate that the circadian neuronal network is an essential buffering system that protects against changes in circadian transcription in the brain.
| Circadian clocks allow organisms to predict daily environmental changes. These clocks time the sleep/wake cycles and many other physiological and cellular pathways to 24hs rhythms. The current model states that circadian clocks keep time by the use of biochemical feedback loops. These feedback loops are responsible for the generation of high amplitude oscillations in gene expression. Abolishment of circadian transcriptional oscillations has been shown to abolish circadian function. Previous studies addressing this issue utilize manipulations in which the abolishment of the transcriptional oscillations is very dramatic and involves strong up or down-regulation of circadian genes. In this study we generated fruit flies in which we diminished the amplitude of circadian oscillations in a controlled way. We found that a decrease of more than 50% in the amplitude of circadian oscillations leads to impaired function of circadian physiological outputs in the periphery but does not significantly affect circadian behavior. This suggests that the clock in the brain has a specific compensatory mechanism. Moreover, we found that flies with reduced oscillation and impaired circadian neuronal communication display aberrant circadian rhythms. These finding support the idea of network buffering mechanisms that allows the brain to produce circadian rhythms even with low amplitude molecular oscillations.
| Most organisms use 24-hr circadian clocks to keep temporal order and anticipate daily environmental changes. These clocks are based on self-sustained biochemical oscillators that manifest at the molecular, physiological and behavioral levels [for review see [1], [2]]. Circadian clocks have been proposed to work on cell-autonomous basis and to be generated by interconnected complex transcriptional-posttranslational feedback loops [1].
In Drosophila melanogaster, the master genes Clock (Clk) and cycle (cyc) activate the circadian system by promoting rhythmic transcription of several key genes. Three of these target gene products, PERIOD (PER) [3], TIMELESS (TIM) [4], and CWO [5]–[7] repress CLK-CYC mediated transcription on a daily basis. The CLK-CYC heterodimer also activates the expression of VRI and PDP1, which are responsible for the oscillation of Clk mRNA [8], [9]. Post-transcriptional and post-translational regulation also contributes to circadian timekeeping [10]–[12]. A central role for transcriptional feedback loops has been challenged by the idea that other modes of regulation, like phosphorylation of key clock proteins as PER are more important for circadian timekeeping. However, other work has re-confirmed the importance of transcriptional regulation for timekeeping in Drosophila and mammals [13]–[15].
Oscillations of clock gene products occur in variety of fly tissues [16]. However, discrete circadian pacemaker neurons in the brain are responsible for the generation of locomotor activity rhythms [17]. These brain pacemakers show robust oscillations at the molecular level even after weeks in constant darkness (DD) [18]. Approximately 150 neurons drive circadian locomotor activity rhythms. They have been divided into several subgroups based on their location and expression of the clock genes PER, TIM, and CRY and the neuropeptide PDF [19], [20]. These groups are called the ventral lateral (sLNvs and lLNvs), dorsal lateral (LNds), and dorsal (DN1s, DN2s, and DN3s) neurons. The neuropeptide PDF, which is expressed exclusively in the LNvs, is essential for normal circadian patterns of activity in LD and persistent circadian rhythms in DD [17], [21]–[23]. Recent evidence suggests that PDF synchronizes the brain circadian neurons [18], [24]–[30].
Peripheral clocks are spread throughout the fly body and regulate a plethora of functions that include eclosion, olfaction, detoxification, and immunity [1]. These clocks have strong molecular rhythms in light/dark (LD) conditions. Although these peripheral clock rhythms disappear in DD in most tissues [31], a few peripheral oscillators perform well in DD. This may be due to stronger or non-dampening transcriptional oscillations (e.g., olfaction [32]) or signaling from the brain oscillator (i.e., eclosion rhythms [33]).
Although both types of oscillators are thought to work in a cell-autonomous fashion, the neurons in the brain central oscillator communicate timing information to each other. This communication was proposed to be responsible for synchronized molecular oscillations in individual cells, which leads to robust behavioral rhythms in DD [18], [24]–[27]. Conversely, in mammals, communication between circadian neurons provides robustness to the brain oscillator [2], [34], [35]–[38]. Despite the great advances achieved in the last few years, the relative importance of intra and intercellular contributions to generation of robust circadian behavioral is still not well understood.
A few years ago, we generated flies carrying the UAS-ClkGR transgene [5]. This transgene encodes a fusion protein (CLKGR) between the full Drosophila CLK protein and the ligand binding domain of rat glucocorticoid receptor (GR). This type of fusion is widely used for generating inducible systems, as the presence of the GR ligand binding domain assures cytoplasmic retention, which can be reversed by addition of GR ligands (like the artificial analog Dexamethasone [39]). Indeed, we previously demonstrated that addition of Dexamethasone to fly tissues expressing the CLKGR fusion leads to quick and very strong induction of CLK-driven transcription [5]. Importantly, addition of the inducer has no other effects, as there is no endogenous glucocorticoid-like receptor or ligand in flies.
Here, we generated flies, herein referred to as TIM-CLKGR flies that express this fusion protein in tim-expressing cells (tim-gal4/+; UAS-ClkGR/+). Surprisingly, expression of CLKGR in a wild type background without adding dexamethasone reduced the amplitude of CLK-driven circadian transcriptional oscillations by more than 50%. This resulted in low-amplitude protein oscillation, and impaired activity of peripheral circadian oscillators. TIM-CLKGR flies displayed almost no transcriptional oscillations of peripheral organ genes and had aberrant eclosion rhythms and sleep disturbances. Interestingly, locomotor activity rhythms were only weakly affected in TIM-CLKGR flies, demonstrating that the brain circadian clock is more resilient to changes in transcriptional amplitude than peripheral clocks. The resilience of the central oscillator is dependent on an intact and functioning circadian neuronal network structure. Indeed, flies in which the pdf neuropeptide pathway is impaired (by mutations in pdf or the in the PDF receptor Han) showed very strong behavioral phenotypes upon expression of the CLKGR protein. In agreement with the prominent role of the pdf signaling pathway, we found that the pdf-expressing cells have a key role in buffering the adverse effects of low amplitude circadian oscillations in the brain. In sum, our results provide strong evidence that high amplitude circadian oscillations in combination with intact neuronal structure are key constituents of robust circadian systems.
We generated flies expressing the CLKGR transgene [5] under the control of the tim-gal4 driver in a wild-type Clk background; hence these flies (TIM-CLKGR flies) carry two endogenous wild-type alleles of Clk and the UAS-ClkGR transgene. These flies also contain a tim-luciferase transgene (tim-luc), which allows in vivo monitoring of CLK-CYC mediated transcription [40]. Control whole flies displayed strong transcriptional rhythms, as did isolated wings (Figure 1A, red line). Surprisingly, we failed to detect luciferase oscillations in the absence of dexamethasone in TIM-CLKGR whole flies and isolated wings (Figure 1A, blue line). This suggests that the CLKGR fusion protein interferes with the endogenous molecular clock. In order to rule out the possibility that the lack of oscillations is due to toxic effect of CLKGR protein expression on the survival of the tissue, we evaluated the levels of the luciferase expression after adding dexamethasone to the culture media. Indeed, addition of dexamethasone resulted in a significant increase in the luciferase levels, demonstrating that CLKGR expression impairs the circadian system rather than affects the survival of the tissue (Figure S1).
To determine whether this fusion protein also interfered with the oscillation of endogenous CLK-driven mRNAs, we performed Real-Time RT PCR to assess tim mRNA levels from fly heads of control and TIM-CLKGR flies collected at two time points. Figure 1B shows that expression of CLKGR in tim-expressing cells significantly diminished the oscillation amplitude of tim mRNA in Light∶Dark (LD) conditions (Figure 1B). Interestingly, the effect is mainly restricted to the amplitude of oscillations rather than to the average levels of tim mRNA (see below). The differences in levels observed between tim mRNA levels and the tim-luciferase reporter are likely due to the fact that in the tim-luciferase transgene transcription is driven by only 760 bases of the tim promoter (compared to the 7 Kb genomic fragment necessary to fully recapitulate tim gene expression in time and space, Shaul Mezan personal communication). Therefore, we conclude that CLKGR interferes with endogenous CLK activity, and that this causes low amplitude transcriptional oscillations without much effect in the total levels of overall CLK-driven transcription.
We then analyzed the effect of the CLKGR protein on transcriptional oscillations in constant darkness (DD). We performed RT-PCR on total RNA from control and TIM-CLKGR fly heads collected at different circadian time points (CT). The amplitude of vri mRNA oscillations was severely diminished (less than 50% of controls, 2.6 fold difference across the day instead of 5.5 and 6.1 of the control strains) in TIM-CLKGR flies compared to the control flies (Figure 1C). Interestingly, we found that the main consequence of expression of CLKGR is at the level of amplitude, as overall vri levels are not significantly different between control and TIM-CLKGR flies (Figure S2A). We also performed a similar assessment in flies carrying two copies of the UAS-ClkGR transgene (TIM-CLKGR(X2) flies). Interestingly, we found that vri mRNA oscillations were almost abolished in these flies (Figure 1C). These results suggest that the CLKGR fusion protein acts as a weak dominant negative regulator of transcriptional oscillations.
In order to evaluate genome-wide effects of the expression of CLKGR, we collected fly heads from control and TIM-CLKGR at two different time points in LD (ZT3 and ZT15) and analyzed their transcriptome using oligonucleotides microarrays. In agreement with the RT-PCR analyses, vri mRNA showed significantly decreased amplitudes of oscillation without much effect in the total mRNA levels (Figure 1D and S2B). We observed a similar decrease for per, tim, Pdp1, cwo and the luciferase mRNAs from the tim-luciferase reporter which is also present in this strain (Figure 1D). In agreement with what we determined by RT-PCR for vri (Figure S2A), expression of the CLKGR transgene does not significantly affect the overall levels of the core CLK-transcriptional targets in this dataset (per, tim, Pdp1 and cwo, see Figure S2B). Interestingly, expression of CLKGR significantly decreased the number of probes differentially expressed between the two assayed time points (the number of oscillating genes is less than one third in TIM-CLKGR flies compare to control). Moreover, more than half (17/24) of the mRNAs that still show differential expression between the two timepoints in TIM-CLKGR flies did so with diminished amplitude with respect to control flies. This was the case for both direct as well as indirect (e.g., cry) CLK-transcriptional targets (Figure 1E; Dataset S1). Therefore, we conclude that expression of the CLKGR fusion interferes with the endogenous CLK by an unknown mechanism. In any case this transgene allows us to study the consequences of diminishing CLK-driven transcriptional oscillations by approx. 60% or 90% (by using TIM-CLKGR and TIM-CLKGR(X2) flies, respectively).
One possible explanation for the reduced amplitude of mRNA oscillations in TIM-CLKGR flies is that a fraction of the CLKGR protein localizes to the nucleus and competes for DNA binding with the wild-type CLK. To test this hypothesis we determined the subcellular localization of the fusion protein in TIM-CLKGR flies. We performed a nucleus/cytoplasmic fractionation from fly heads and determined the levels of CLK and CLKGR proteins in each fraction using an anti-CLK antibody. As expected, TIM-CLKGR flies expressed an additional CLK immunoreactive protein of higher molecular weight than the wild-type CLK. The levels of this fusion protein were much higher than the endogenous CLK (Figure 2A). We detected high amounts of the CLKGR fusion protein both in the cytoplasm and in the nuclear fraction (Figure 2B). The presence of CLKGR in the nuclear fraction was not due to cytoplasmic contamination, as tubulin, an exclusive cytoplasmic protein, was only present in the cytoplasmic fraction (Figure 2B).
To determine whether CLKGR could inhibit CLK-mediated transcription, we utilized Drosophila S2 cells. These cells do not express CLK, but express CYC at high levels [41]. To test CLKGR/CLK competition, we transfected S2 cells with a vri-luciferase reporter, a plasmid that expresses CLK at constant levels (pAc-CLK), and a third plasmid in which CLK or CLKGR expression can be induced by addition of copper (pMT-CLK or pMT-CLKGR, respectively). Additional CLK production from the MT-CLK plasmid resulted in a further increase in the levels of the reporter gene expression (Figure 2C, left). However, induction of CLKGR resulted in a dose-dependent reduction of the levels of the reporter expression (Figure 2C, center). This demonstrates that CLKGR can compete with CLK and can partially inhibit CLK-mediated transcription. In Drosophila S2 cells CLKGR is incapable of activating CLK-mediated transcription [42]. We postulated that the fraction of the CLKGR protein pool that translocate to the nucleus forms CLKGR-CYC dimers that are inactive due to steric inhibition of the CLK transactivation domain by the GR domain. However, a transcriptionally inactive CLK could activate transcription if co-expressed with an artificial CYC protein carrying a VP16 transcriptional activation domain (CYCVP16 fusion [14]). As an indication for CLKGR binding to the DNA, we determined whether the CLKGR-CYCVP16 dimer could activate transcription. As previously described [14], CYCVP16 alone was not sufficient to active the luciferase reporter (Figure 2D, left bar). However, CLKGR strongly promoted expression of the reporter in presence of CYCVP16 (Figure 2D, right and middle bars). This demonstrates that CLKGR can translocate to the nucleus and bind to CLK-target sites although it cannot activate transcription per se.
To determine the effect of CLKGR expression on the amplitude of circadian protein oscillations, we determined VRI expression by western blot from TIM-CLKGR and control fly heads collected at six different time points during DD1. TIM-CLKGR flies displayed rhythms in VRI expression, although with lower amplitudes than control flies (Figure 3A and 3B). We also recorded protein oscillation in vivo by utilizing the XLG PER-luciferase fusion transgene [43]. These flies carry an in frame fusion of the PER protein with luciferase which is driven by period promoter regions. Hence luciferase activity correlates with PER protein levels. TIM-CLKGR flies displayed lower amplitude oscillations in PER-Luciferase levels than wild-type controls (Figure S3). Therefore we concluded that defects in circadian transcription in TIM-CLKGR flies resulted also in low amplitude oscillations at the protein level, at least for VRI and the PER-LUC fusion reporter.
In order to determine whether expression of CLKGR also affects the amplitude of protein oscillations in the brain, we determined VRI levels by immunocytochemistry at four different timepoints in the brain of control and TIM-CLKGR flies. We performed the experiment in the first day in constant darkness (DD1). As observed in the western blot assays from whole heads, TIM-CLKGR brains display oscillation in VRI levels in the sLNvs, although of smaller amplitude than control flies (Figure 3C and Figure 3D).
TIM-CLKGR flies offer an opportunity to determine the importance of transcriptional oscillation amplitude on circadian physiology. In order to test the effect on a peripheral circadian clock, we assayed eclosion timing. A specialized structure, the prothoracic gland, is responsible for the generation of eclosion circadian gating [31], [33]. Wild-type flies have strong eclosion circadian gating either in LD or DD conditions (Figure 4A, and Figure S4A and S4B). However, both LD and DD eclosion rhythms are completely absent in TIM-CLKGR flies (Figure 4A and Figure S4A). In order to determine whether these defects arose from dampening of oscillations in the brain or in the prothoracic gland, we generated flies in which CLKGR expression was restricted to the master pacemaker neurons in the fly brain, the pdf-expressing cells, or to the peripheral gland driving eclosion (the prothoracic gland). Whereas expression of CLKGR in the pdf-expressing cells in the fly brain did not affect circadian eclosion rhythms (Figure S4B), expression of CLKGR in the prothoracic gland (by the use of the GAL4 driver Mai60 [33]) was sufficient to suppress daily eclosion rhythms (Figure 4A). Therefore we concluded that the impairment of the eclosion rhythms observed in TIM-CLKGR flies is due to expression of CLKGR in the prothoracic gland.
We then determined whether sleep is also affected in TIM-CLKGR flies. Although sleep and circadian behavior are intrinsically linked, sleep is considered a “peripheral-like” output of the circadian clock, despite the fact that it can be controlled by a subgroup of the pdf-expressing neurons, the large LNv cells [44]–[47]. We analyzed the sleep behaviors of TIM-CLKGR, TIM-CLKGR(X2), and control flies. As shown in Figure 4B, total sleep was not affected by expression of CLKGR transgenes, but we observed a dramatic effect on the number of sleep episodes during the dark phase, which was dependent in the dose of expressed CLKGR protein (Figure 4C). The significant increase in the number of sleep episodes in TIM-CLKGR flies was accompanied with an equally dramatic decrease in the length of each sleep episode (Figure 4D). From these experiments we concluded that high amplitude rhythms in CLK-driven transcription are necessary for circadian eclosion as well as for normal sleep consolidation.
We next evaluated whether TIM-CLKGR flies have defects in locomotor activity rhythms. Control flies display strong behavioral rhythms both in LD and DD conditions (Table 1). Surprisingly, TIM-CLKGR flies also display rhythmic locomotor activity (Table 1). This indicates that the circadian system driving locomotor rhythms performs well even when the amplitudes of CLK-driven transcriptional oscillations are decreased by approximately 50%, as in TIM-CLKGR flies. On the other hand, most TIM-CLKGR(X2) flies show disrupted circadian rhythms (they are either arrhythmic or rhythmic with very low rhythmic power, see Table 1). Therefore we concluded that the circadian brain oscillator is not able to function when further dampening occurs (e.g. in TIM-CLKGR (X2) flies). It is possible that the opposite results obtained in the eclosion and behavioral assays are due to different thresholds of the assays (i.e. due to the different number of days utilized in the analysis). To rule out this possibility, we utilized a similar timeframe and statistical analysis for determining daily changes in control and TIM-CLKGR flies. As with the previous analysis, we found that while control flies display daily rhythms in both eclosion and locomotor activity, TIM-CLKGR flies have significant differences only in the behavioral assay (Figure S4C, D and E).
The different effects provoked by the expression of the CLKGR fusion protein in the central and peripheral oscillators is reminiscent of the molecular and behavioral phenotypes of cryptochrome (cry) mutant flies, which lack the main circadian photoreceptor [48], [49]. These flies display overall low amplitude mRNA oscillations in fly heads, which are due to desynchronization and not to low amplitude oscillations in individual circadian oscillators [50]. Therefore we assessed whether TIM-CLKGR flies have normal circadian photoreception during development and adulthood. Indeed, TIM-CLKGR flies that were raised in LD conditions but transferred to constant darkness before eclosion kept the original phase, demonstrating entrainment capability during development (Figure S5A). Moreover, TIM-CLKGR flies had normal Phase Response Curves (PRC), demonstrating normal CRY function in adult flies (Figure S5B). Therefore the effects observed in TIM-CLKGR flies are not due to inhibition or inactivation of CRY.
One possible explanation for the resistance of the brain circadian oscillator to the expression of CLKGR is that this fusion protein does not have the same molecular effect in the fly brain as it does in peripheral systems. Although we observed lower amplitude protein oscillations in the brains of TIM-CLKGR flies (Figure 3C, D), we decided to further test this possibility. For doing so, we evaluated transcriptional rhythms in cultured fly brains using a tim-luciferae reporter. Control tim-luciferase fly brains displayed strong rhythms in LD (Figure 5A, dotted line). TIM-CLKGR brains displayed non-oscillating levels of the transcriptional reporter similar to the patterns displayed by whole TIM-CLKGR flies or cultured fly wings of flies from this genotype (Figure 5A solid line, compare with Figure 1A). In order to ensure that constant low levels of the luciferase reporter in TIM-CLKGR flies are not due to death of the tissue, we added dexamethasone to the TIM-CLKGR brains and followed CLK-driven transcription using the same tim-luciferase reporter. TIM-CLKGR cultured fly brains treated with dexamethasone displayed significantly increased levels of CLK-driven transcription relative to untreated brains (Figure 5B). These results are in agreement with the VRI immunocytochemistry profiles (Figure 3C,D) and demonstrate that expression of CLKGR leads to dampened transcriptional oscillations in the brain.
Given these results, we decided to more stringently analyze whether TIM-CLKGR flies displayed any circadian behavioral defects. For doing so, we recorded locomotor activity of control and of TIM-CLKGR flies for more extended times in constant darkness (DD). We computed the percentage of rhythmic flies, the period, and the power, as previously described [51] in three time intervals (DD 1 to 5, 6 to 10, and 11 to 15). Control flies displayed very strong rhythms throughout the experiment with more than 90% of flies displaying rhythmic behavior (Figure 5C and S6). Although most TIM-CLKGR flies were rhythmic even after 15 days in DD, we consistently observed that these flies displayed weaker rhythms and more diverse peak phases than control flies (Figure 5C and 5D). Therefore, we concluded that high amplitude transcriptional oscillations are necessary to maintain robust circadian locomotor activity for long periods in absence of environmental cues.
The observation that CLKGR expression results in mild behavioral phenotypes, suggest a brain-specific mechanism operating in TIM-CLKGR to buffer the significant decrease in molecular oscillations. In order to investigate this possibility, we determined the behavior of flies in which the CLKGR fusion protein is expressed in different subsets of the circadian neuronal network. First, we generated flies expressing the CLKGR fusion protein only in the pdf-expressing cells, (referred to as PDF-CLKGR). Expression of CLKGR in the LNvs alone had strong effects on the percentage of rhythmic flies late in DD (Figure 6A and Figure S6). We then combined the tim-gal4 driver with cell-specific GAL80-expression transgenes in order to restrict GAL4 activity to different subsets of tim-expressing cells (specifically the TIM+CRY− and TIM+PDF− cells, as previously done by Stoleru et al [24]). Interestingly, expression of GAL80 either in the PDF or CRY-positive neurons improves the mild behavioral phenotype observed in TIM-CLKGR flies in late DD (Figure S7). These results verified the importance of the pdf-expressing cells for mediating the robustness of circadian behavior (see below and discussion).
To determine whether the ability of the TIM-CLKGR flies to remain rhythmic is mediated by the neuropeptide PDF, we determined whether flies mutant for components of the PDF-signaling pathway are especially sensitive to expression of the CLKGR transgene. For doing so, we generated TIM-CLKGR flies, which also carry null mutations for the neuropeptide PDF (pdf01; TIM-CLKGR-pdf01 flies) or the PDF Receptor (Han mutant flies; HAN-TIM-CLKGR flies). Both pdf01 and Han mutants lose rhythmicity only after several days in DD [18], [21]–[23]. As expected, more than 70% of the pdf01 and the Han mutant flies are rhythmic during the first days in DD (Figure 6B and Figure S8). Interestingly, expression of CLKGR in a pdf or Han mutant backgrounds resulted in a dramatic reduction of the number of rhythmic flies (Figure 6B and Figure S8). Interestingly, the interaction between the CLKGR transgene and the pdf signaling pathway is specific, as we did not observed a similar genetic interaction between the CLKGR transgene and the perL mutation (Figure S9). These results clearly demonstrate that PDF-mediated communication is an essential mechanism mediating the resilience of the brain circadian oscillator to the dampening of circadian transcriptional rhythms provoked by the expression of the CLKGR fusion protein.
In this study, we used the CLKGR fusion protein in Drosophila to determine the relative contribution of high amplitude transcriptional oscillations and neuronal communication for robust circadian behavior. Expression of the CLKGR fusion protein in tim-expressing cells decreased more than 50% the amplitude of circadian transcriptional oscillations. The impaired transcriptional oscillations lead to low amplitude protein oscillations, which were not sufficient to drive outputs of peripheral oscillators like eclosion rhythms. However, circadian locomotor behavior remained rhythmic. This difference was likely due to intercellular interactions between the circadian neurons in the brain that buffer the low amplitude transcriptional oscillations. Despite this compensation, TIM-CLKGR flies display weaker behavior rhythms after many days in constant darkness. We demonstrated that the compensatory mechanism is dependent on the relationship between the amplitudes of molecular oscillations in different neuronal clusters, especially between the pdf-expressing neurons and the rest of the circadian network. Lastly, we showed that the neuropeptide PDF is the key factor contributing to the resilience of the brain oscillator to expression of the CLKGR transgene. Dampening of transcriptional oscillations provoked by CLKGR expression in the context of PDF or PDF receptor mutations resulted in arrhythmicity even very early in constant darkness. In sum, our work revealed the importance of high amplitude transcriptional oscillations in Drosophila and how these oscillations contribute to the robustness of the brain circadian oscillator.
Many dominant negatives CLK proteins have been used in the past [15], [52]–[54]. In all these cases, the effects on transcriptional oscillations are dramatic (almost no amplitude remaining), and mutants have strong circadian behavioral phenotypes. In these mutants, the CLK levels critical for development, cell viability, and normal physiology are also severely reduced. For example, Drosophila ClkJrk and ClkAR mutants, present abnormal development of the circadian neurons, precluding the assessment of whether the circadian defects are mainly due to impaired oscillations or developmental defects [53], [55]. Our manipulation offers two advantages to address this issue: First, we can titrate the amplitude of CLK-driven oscillations by utilizing flies with different number of ClkGR transgenes, and second, our manipulation does not significantly change the overall levels of CLK-driven transcription (see Figure S2A and S2B). We offer strong evidence of the mechanism by which CLKGR partially inhibits CLK-CYC driven transcription. A fraction of the CLKGR fusion protein leaks into to the nucleus and binds to chromatin, inhibiting the action of the endogenous CLK protein (see Figure 2). The inhibitory action of the CLKGR protein can be explained by steric interference of the ligand-binding domain of the glucocorticoid receptor with the CLK activation domain. We were unable to accurately determine how much CLKGR was chromatin-bound. It may be that most is bound and elicits minimal transcriptional activity or that a small fraction is bound and is transcriptionally inactive (or even inhibitory). We were unable to perform the chromatin immunoprecipitation assays that would have addressed the issue due to the low quality of available antibodies against the GR domain. As expression of CLKGR in Drosophila S2 cells did not have any effect on CLK-driven transcription, we favor a competitive binding inhibition scenario. Other than the defective activation, the CLKGR fusion protein seems to respond well to cyclic repression by PER as CLKGR flies still displayed some transcriptional oscillations and similar overall levels of CLK-transcriptional targets as wild-type flies.
Low-amplitude CLK-driven transcription leads to lower amplitude oscillations in all circadian transcription, not just in CLK-direct transcriptional targets (Figure 1E). This demonstrates the centrality of transcriptional control for genome-wide mRNA oscillations. This centrality is highlighted by the strong effect of expression of CLKGR in the physiological output of peripheral oscillators (e.g., eclosion). Although some of the core circadian components like PER and TIM are strongly regulated at the post-translational level, our results suggests that most output genes are regulated at the transcriptional and post-transcriptional levels (rather than post-transnationally). Post-transcriptional regulation also cannot be ruled out, as CLK-driven transcription could affect indirectly the oscillation amplitudes of hundreds of genes by regulation of non-coding RNAs or RNA binding proteins.
In addition to promoting high-amplitude protein oscillations, CLK-driven transcription can serve other functions. For example, for genes with long-lived mRNA and/or protein products, direct CLK control ensures that these genes are expressed in circadian tissues. Moreover, in cases of mRNA and/or proteins with very high turnover rates, CLK-dependent control means that functional levels are reached at least once a day. We speculate that this may be the case for sleep control by the circadian system; CLK may directly or indirectly modulate the levels of dopamine-related arousal signals in the brain (e.g. in the large LNvs), which have been shown to be regulated/influenced by the circadian system: [44], [46], [56].
It has been previously postulated that transcriptional rhythms may not be necessary for accurate circadian timekeeping. Our study definitively demonstrates the necessity for high-amplitude transcriptional oscillations for normal circadian output, especially in peripheral tissues. Although some aspects of circadian behavior can be rescued when TIM and PER are expressed at constant levels [57], per mRNA oscillations in Drosophila and feedback repression in mammals are key for proper circadian control [14], [15]. However, to the best of our knowledge this is the first time transcriptional oscillations have been partially damped in a living organism and their role assessed comprehensively.
Transcriptional oscillations seem to be less important for the brain circadian oscillator. We postulate that in the brain, communication between the circadian neuronal groups can compensate for the dampened transcriptional oscillations. This is not surprising and results obtained in mammals are among the same lines [58]. Mutations in core clock components, which have deleterious effects on transcriptional oscillations in isolated suprachiasmatic nucleus neurons and in peripheral clocks, have mild or no effects on daily locomotor activity patters [38], [59]. This resilience and the general robustness properties of circadian oscillators in the suprachiasmatic nucleus are due to neuron-to-neuron communication [60], [61]. However, in mammals the molecular machinery that drives circadian rhythms in the central and in the peripheral oscillators differs [62], [63], while this does not seem to be the case in flies.
The use of the CLKGR system allowed us to determine until which point the circadian clock can compensate for dampened transcriptional oscillations. For example, the brain oscillator can still function fairly well after reducing the amplitude of oscillations more than 50% (TIM-CLKGR flies) but not after further flattening.
Interestingly, we found that PDF-CLKGR flies display a stronger behavioral phenotype than TIM-CLKGR flies (Figure 6A). We don't think this is due to different levels of expression of the CLKGR proteins in the sLNVs, as pdf-gal4 and tim-gal4 drivers express with similar strength in those cells [25]. Moreover, previous results utilizing this driver in combination with UAS-transgenes that affect circadian period like sgg or CYCVP16 also do not support this possibility [14], [64]. Our interpretation is that the compensatory mechanism operating in TIM-CLKGR flies requires that the molecular oscillations in the pdf-expressing cells be of equal or higher magnitude than in the rest of the circadian neuronal network. We speculate that this is due to the hierarchical nature of the circadian neuronal network (with the sLNvs being at the top of this hierarchy). Hence, in TIM-CLKGR flies, sLNvs can still set the pace of the circuit and the circadian clock in a pdf-pathway dependent way. In PDF-CLKGR flies, low amplitude oscillations in the sLNvs are not enough to drive the rest of the network, likely due to more resistance from the other neuronal groups, which have higher amplitude molecular oscillations than the pdf-expressing cells. This is further supported by the fact that we did not observe any behavioral defect when we expressed the CLKGR fusion only in the CRY+PDF− or the TIM+CRY− cells. The centrality of the sLNvs for the compensatory mechanism is highlighted by the fact that removing PDF signaling eliminates the capacity of TIM-CLKGR flies to keep rhythmic behavioral patterns (Figure 6B and Figure S8).
In sum, our study revealed important differences between the central and peripheral circadian oscillators regarding the dependence on transcriptional oscillations. By dissecting the mechanism mediating the resilience of the brain oscillators, we were able to dissect the contributions of molecular and neuronal network pathways on the generation of robust and coherent behavioral circadian rhythms.
Tim-gal4, pdf-gal4, tim-luc, UAS-ClkGR, per.XLG-luc, pdf-gal80, cry-gal80 and PerL were previously described [14], [17], [43], [53], [65]–[68] Han3369 and P{GawB}Mai60 [23], [33] lines were obtained from Bloomington stock center.
Total RNA was extracted using Trizol reagent (Invitrogen). Probe preparation and hybridization, staining and washing of the Affymetrix high-density arrays were carried out as described in the Expression Analysis Technical Manual (Affymetrix).
We used male flies at all behavior experiment except for the experiment described at figure S8 in which females were used. Flies were monitored using Trikinetics Drosophila Activity Monitors (Waltham, MA, USA). Rhythmic flies were determent by chi square power, using Faas software (http://www.inaf.cnrs-gif.fr/ned/equipe03_eng/faasx.html) [51].
Sleep measurements were performed using Trikinetics Drosophila Activity Monitors (Waltham, MA, USA). In all the cases we recorded the activity of male flies during 5 days in 12 hours light 12 hours dark light regime (LD) on 1 minute intervals. For analyzing the data we utilized the software pySolo [69].
For assaying eclosion ratio at 12 hours light 12 hours dark regime, we placed individual pupas into behavior tubes. New emerged flies were detected by monitoring movements using Trikinetics Drosophila Activity Monitors (Waltham, MA, USA). In order to find eclosion ratios in the first day in constant darkness (DD1), fly populations were entrained in bottles to light regimes of 12 hours light and 12 hours dark for 3 days and then transferred into constant darkness. Adult flies were removed from the bottles at the end of the last light cycle (ZT24) and newly emerged flies were then removed from the bottles and counted every 2 hours.
Flies were entrained for at least 3 days in 12∶12 LD, and then transferred to constant darkness conditions. During DD1 four time points were collected. Whole flies were placed into fixative solution (PBS 4% paraformaldehyde 0.1% triton-x) for 30 minutes in 4°C followed by 2 hours rotation at room temperature. Then flies were transferred to PBS and the brains were dissected, wash 3 times (PBS 0.1% triton-x) and transferred to 30 minutes blocking solution, (PBS 0.1% triton-x 2% horse serum). After 3 more washes, brains were incubated with primary antibody solution overnight, PBS 0.1% triton-x 2% horse serum 1∶3000 G.P anti VRI (gift from Paul Hardin) 1∶1000 Ms anti PDF (gift from Justin Blau). Brains were washed 3 more times and incubated with secondary antibodies solution, PBS 0.1% triton-x 2% horse serum 1∶500 Alexa Fluor 488 Goat anti G.P (invitrogene) 1∶500 Dylight 550 Dnk anti Ms (Abcam), for 1 hour in room temperature. Brains were washed 3 times and mounted in VECTASHIELD mounting medium (VECTOR) on microscope slides. Photos were taken using Eclipse Ti - Nikon confocal microscope in magnitude of ×200. Quantifications were done utilizing NIS-Elements Ar Microscope Imaging Software.
Adult male flies and dissected heads wings and brains were cultured in 12∶12 LD conditions, and luciferase was measured as described previously [70].
Plasmids were described previously: pAc-CYCVP16 [14], MT-CLKGR [71]. vri-luc [72], pAc-clk, Copia Renilla luciferase and tim-luc, [42]. MT-CLK was generated by amplifying the Clk ORF by PCR and ligating it into pMT-V5 (Invitrogen) using the enzymes KpnI and NotI.
Drosophila melanogaster Schneider-2 cells were grown at 25°C in Schneider's Medium with L-Glutamine (Biological Industries, Jerusalem, Israel/Invitrogen, Carlsbad, CA, USA) supplemented with 10% fetal bovine serum (GIBCO) and 1% Antibiotic-Antimycotic, GIBCO. Cells were seeded in a 6 well plate. Transfection was performed at 70–90% confluence according to company recommendations (6 µl of TransIT-2020 Transfection reagent, Mirus and 2 µg of total DNA). In all experiments 75 ng of pCopia-Renilla Luciferase plus 50 ng of the Luciferase firefly reporter were used. For the plasmids MT-CLK, MT-CLKGR and pAc-CYCVP16 100 ng were used.
Twenty-four hours after transfection, cells were treated with CuSO4 in the indicated doses and after 24 hs of induction, cells were lysed and assayed using the Dual Luciferase Assay Kit (Promega) following the manufacturer's instructions. In the transfections with pAc-CYCVP16 cells were collected 48 hs after transfection.
Total RNA was prepared from adult fly heads (30 heads per sample) using Trizol reagent (Sigma) according to the manufacturer's protocol. cDNA derived from this RNA (using iScript Bio-Rad) was utilized as a template for quantitative real-time PCR performed with the C1000 Thermal Cycler Bio-Rad. The PCR mixture contained Taq polymerase (SYBR green Bio- Rad). tim: 5′-CCTTTTCGTACACAGATGCC-3′, 5′ –GGTCCGTCTGGTGATCCCAG-3′ and 5′-GCTGGCCGATTACAGGATAAC-3′, 5′AGTAAAACAGCGGCACACTCA-3′; vri: 5′- GTCTAATTCTCGCTCCCTCT -3′, 5′- GAACTTTCTTTGTTCGTTGG -3′; Rp49: 5′-TACAGGCCCAAGATCGTGAA-3′, 5′-CCATTTGTGCGACAGCTTAG -3′; and RpS18: 5′CCTTCTGCCTGTTGAGGA- -3′ 5′-TGCACCGAGGAGGAGGTC -3′. Cycling parameters were 95°C for 3 min, followed by 40 cycles of 95°C for 10 s, 55°C for 10 s, and 72°C for 30 s. Fluorescence intensities were plotted versus the number of cycles by using an algorithm provided by the manufacturer. mRNA levels were quantified using a calibration curve based upon dilution of concentrated cDNA. mRNA values from heads were normalized to that from ribosomal proteins 49 (Rp49) and RpS18.
Fly heads (20 heads per sample) were collected and homogenized in RIPA lysis buffer (50 mM Tris-HCl at pH 7.4, 150 mM NaCl, 1 mM EDTA, 1% NP-40 0.5% Sodium deoxycholate, and 0.1% sodium dodecyl sulfate (SDS), 1 mM DTT, with protease inhibitor cocktail and phosphatase inhibitors). Head lysates were then centrifuged for 10 minutes and the supernatant was boiled with protein sample buffer (Bio-Rad). Samples were resolved by Criterion XT Bis-Tris gels (Bio-Rad). Antibodies used for Western blotting were as follows: anti-CLK (a kind gift from Paul Hardin), anti-VRI (a kind gift from Paul Hardin), anti Tub (DM1A, SIGMA), anti His3 (Abcam). Quantifications were done utilizing Image J software.
Fly heads were homogenized in a Dounce homogenizer, in the following buffer: 10 mM Hepes pH 7.5, 10 mM KCl, 0.8 M Sucrose, 1 mM EDTA, 0.5 mM DTT, supplemented by protease-inhibitor cocktail (mini complete, Roche) and phosphatase inhibitors. After homogenization, the homogenate was filtered through a column polymer bed support (Bio-Rad unfilled Bio-spin Column 4 minutes 1000 g 4°C) to remove the cuticle. The filtrate was then centrifuged (600 g, 10 minutes 4°C) and the pelleted cell extract were then subjected to nuclear cytoplasmic fractionation. To prepare the cytoplasmic fraction, the cell pellets were re-suspended in cytoplasmic buffer (10 mM Tris HCl pH 8.0, 10 mM KCl, supplemented by protease inhibitor cocktail (mini complete, Roche) and phosphatase inhibitors). Cells were allowed to swell for 15 minutes, and then NP-40 was added to 0.4%, followed by centrifugation (3500 g, 3 minutes 4°C). The supernatant contained the soluble cytoplasmic fraction. The pellets were washed once more with the cytoplasmic buffer before proceeding to nuclear fractionation. For the preparation of the nuclear fraction, the remaining cell pellet was re-suspended in high-salt buffer (50 mM Tris pH 8.0, 5 mM EDTA, 400 mM NaCl, 1% NP-40, 1% Sodium deoxycholate, and 0.025% sodium dodecyl sulfate (SDS), 1 mM DTT supplemented by protease inhibitor cocktail (mini complete, Roche) and phosphatase inhibitors). The nuclear pellet was hard vortex for 30 min at 4°C, and was than centrifuged (1 minute 4°C max speed). The supernatant, which contains the nuclear fraction, was saved. Both fractions were re-suspended in protein sample buffer, heated 5 minutes 95°C.
Flies were entrained to a 12∶12 LD cycle for 4 days. During the fifth dark phase of the cycle, flies groups contains 32 flies were given a 10-min saturating white light pulse (1000 lux) at 13, 15, 17, 19, 21, and 23 h after the last light-on event. A separate control group of 32 flies was not given a light pulse. Flies were then put into DD. The average phase of the locomotor activity peaks after the light pulse was determined and compared with the no-light-pulse control.
In order to test pre adult entrainment flies were grown at 12 hours light: 12 hours dark (LD) light regime till pupa stage. Pupas have been placed into behavior tubes in constant darkness (DD). After eclosion, the locomotor activity of the flies was monitored in DD in order determinate the phase of circadian activity.
|
10.1371/journal.pgen.1004493 | An Intronic microRNA Links Rb/E2F and EGFR Signaling | The importance of microRNAs in the regulation of various aspects of biology and disease is well recognized. However, what remains largely unappreciated is that a significant number of miRNAs are embedded within and are often co-expressed with protein-coding host genes. Such a configuration raises the possibility of a functional interaction between a miRNA and the gene it resides in. This is exemplified by the Drosophila melanogaster dE2f1 gene that harbors two miRNAs, mir-11 and mir-998, within its last intron. miR-11 was demonstrated to limit the proapoptotic function of dE2F1 by repressing cell death genes that are directly regulated by dE2F1, however the biological role of miR-998 was unknown. Here we show that one of the functions of miR-998 is to suppress dE2F1-dependent cell death specifically in rbf mutants by elevating EGFR signaling. Mechanistically, miR-998 operates by repressing dCbl, a negative regulator of EGFR signaling. Significantly, dCbl is a critical target of miR-998 since dCbl phenocopies the effects of miR-998 on dE2f1-dependent apoptosis in rbf mutants. Importantly, this regulation is conserved, as the miR-998 seed family member miR-29 repressed c-Cbl, and enhanced MAPK activity and wound healing in mammalian cells. Therefore, the two intronic miRNAs embedded in the dE2f1 gene limit the apoptotic function of dE2f1, but operate in different contexts and act through distinct mechanisms. These results also illustrate that examining an intronic miRNA in the context of its host's function can be valuable in elucidating the biological function of the miRNA, and provide new information about the regulation of the host gene itself.
| Animal genomes encode hundreds of microRNA genes that impact all areas of biology by limiting the expression of their targets. What remains largely unappreciated is that a significant proportion of microRNA genes are embedded within protein-coding genes, and are often co-expressed with their hosts, which raises the possibility of a functional interaction between them. The mir-998 gene is located within an intron of the gene encoding Drosophila E2F1 transcription factor. E2F1 can induce the expression of cell death genes, and its activity is negatively regulated by the pRB tumour suppressor protein. In certain settings, unrestrained E2F1 activity is sufficient to induce cell death in cells lacking functional pRB. Here, we show that miR-998 limits cell death in Rb-deficient cells by repressing dCbl, a negative regulator of Epidermal Growth Factor Receptor signaling (EGFR). miR-998 also augments EGFR signaling in differentiating photoreceptor cells. Furthermore, we show that the interaction between miR-998 and Cbl is conserved: in human cells, miR-29, a mir-29/998 seed family member, enhances EGFR signaling by targeting c-Cbl. Therefore, by examining the role of an intronic microRNA in the context of its host's function, we identified an important microRNA target and uncovered a biological function of the microRNA.
| MicroRNAs (miRNAs) are short non-coding RNAs that regulate the expression of mRNA targets, thereby modulating biological processes including development, proliferation, metabolism, homeostasis and tumorigenesis. While some miRNAs elicit strong effects, many miRNAs operate more subtly to buffer a system or response to a signal. There is significant redundancy among miRNAs of the same family in regulating their target genes, making it difficult to identify the physiological role of an individual miRNA. The absence of strong loss-of-function phenotypes of a significant proportion of miRNAs has significantly hampered the characterization of their functions in vivo [1]. A number of approaches have been used to reveal miRNA functions, including combining mutations to generate synthetic phenotypes [2]–[5]. What remains largely unappreciated is that approximately 40% of miRNA genes are embedded within, and frequently co-expressed with protein-coding genes [6], [7]. There is a growing number of examples of intronic miRNAs directly impacting the function of the genes in which they reside [8]–[12]. Therefore, investigating a miRNA in the context of its host gene function could potentially provide insight into the biological roles of a large number of miRNAs. The value of such an approach is illustrated by recent studies of the Drosophila melanogaster dE2f1 gene.
The dE2f1 transcription factor, and its mammalian homologs coordinate the expression of genes involved in cell proliferation and cell death. In a variety of systems, E2F is rate-limiting for S phase entry while it triggers apoptosis in specific contexts. The last intron of the Drosophila E2F gene dE2f1 harbors a miRNA, mir-11, which is co-expressed with dE2f1 (Figure 1A and Figure S1). The loss of mir-11 was shown to strongly enhance dE2F1-dependent DNA damage-induced apoptosis even though it was insufficient to cause cell death in unprovoked settings. Therefore, the physiological role of mir-11 was revealed only when examined in the sensitized background of its host gene. This function of miR-11 is explained by its ability to directly regulate the expression of dE2F1-regulated cell death genes, thus highlighting a complex interaction between an intronic miRNA and its host gene [1], [8]. In addition to mir-11, the last intron of the dE2f1 gene contains another miRNA, mir-998. The sequence of mature miR-998 is different than that of miR-11, particularly at the 5′ end in the seed sequence (Figure 1A), which is the primary determinant of miRNA target selection. Therefore the two miRNAs are likely to regulate distinct sets of genes and, consequently, may have different functions. However since miR-998 was only recently identified, nothing was known about its biological function. Here we show that miR-998 limits dE2F-dependent cell death, but it does so in a different context and by a different mechanism than miR-11. While miR-11 repressed components of the core cell death machinery, including rpr and hid, miR-998 limited E2F-dependent cell death by elevating prosurvival signaling downstream of the Epidermal Growth Factor Receptor (EGFR) through regulation of dCbl, a negative regulator of EGFR. Thus, our data reveal a novel layer of intrinsic regulation at the dE2f1 genomic locus involving intronic miRNAs.
Rbf is a negative regulator of dE2F1. The loss of rbf sensitizes cells to dE2F1-dependent apoptosis. As has been previously shown, there is a high level of apoptosis in a band running along the anterior edge of the morphogenetic furrow of rbf mutant eye discs [2]–[5], [13]. This stripe of apoptotic cells can be revealed by staining with the C3 antibody, which specifically recognizes activated caspases (Figure 1B). Importantly, apoptosis is dependent on dE2F1 since it was suppressed in rbf, dE2f1 double mutant animals [6], [7], [13]. To examine the effect of miR-998, we expressed a UAS-mir-998 transgene in the developing eyes of rbf mutant animals using the ey-FLP; Act≫Gal4 (Flip-out) system. Remarkably, no apoptotic cells were found in rbf120a, act>mir-998 eye discs, indicating that miR-998 strongly suppressed E2F-dependent cell death in this context (Figure 1B). Interestingly, unlike miR-998, miR-11 failed to block apoptosis in rbf mutants as the number of C3 positive cells was similar between rbf120a, act>mir-11 and rbf120a eye discs.
Differences in suppression of cell death in rbf mutant cells by miR-11 and miR-998 prompted us to investigate the impact of the two miRNAs on dE2F1-dependent apoptosis in other settings. When dE2F1 expression is driven by the Act88F-Gal4 driver, high levels of apoptosis in the wings of newly eclosed adults give rise to gnarled, blistered wings that have a downward curvature [8]–[12], [14]. This cell death phenotype is strongly rescued by miR-11 (Figure S2A and [8]). However, the wings of Act88F>dE2f1, mir-998 animals were indistinguishable from the wings of Act88F>dE2f1 adults, suggesting that expression of miR-998 was insufficient to suppress apoptosis (Figure S2A). Next, we performed genetic interaction tests in the eye imaginal disc. Ectopic expression of dE2f1 in the posterior compartment of the eye imaginal disc potently induces apoptosis [15]. This apoptosis is strongly suppressed by co-expression of miR-11 (Figure S2B and [8]). In contrast, miR-998 had no effect on E2F1-induced cell death in the posterior compartment, as the level of C3 staining was indistinguishable between GMR>dE2f1/dDP/mir-998 and GMR>dE2f1/dDP eye discs (Figure S2B). In addition to apoptosis, GMR>dE2f1/dDP had been shown to induce unscheduled proliferation that can be visualized by BrdU labeling [15]. However, neither miR-11 nor miR-998 modulated dE2F1-induced proliferation, as the level of E2F1-induced ectopic BrdU incorporation was largely unchanged by co-expression of miR-11, as was previously shown [8], or miR-998 (Figure S2B).
Therefore we concluded that overexpression of miR-998 suppressed dE2F1-dependent apoptosis in rbf mutants but not when dE2F1 was ectopically expressed in the eye or in the wing. In contrast, miR-11 suppressed dE2F1-induced phenotypes but failed to block apoptosis in rbf mutants. Thus, miR-998 and miR-11 both suppressed E2F-dependent cell death, but did so in mutually exclusive contexts.
The results described above using a miR-998 transgene raise the question of whether endogenous miR-998 operates in a similar manner and blocks apoptosis in rbf mutants. Therefore we examined the consequence of the loss of mir-998 in the background of the rbf120a mutation. Since there were no pre-existing mir-998 mutants, we generated a mir-998 null allele (for details see Materials and Methods and Figure S3). It was essential that the mir-998 loss-of-function allele did not disrupt the expression of mir-11 or dE2f1, such that any observed phenotype could be attributed specifically to the function of miR-998. We used one piggyBac element and two P elements inserted near the dE2f1 gene and screened for local transpositions of these transposons into intron 5, which harbors miR-11 and miR-998. Out of 4,254 transposition events a single P element insertion into intron 5 was recovered and then used to screen for imprecise excisions that specifically disrupted mir-998. Among 400 excision events only two imprecise events were isolated and both of them retained a small piece of the P-element. The mir-998exc222 allele contained an 87 bp insertion within the mir-998 hairpin that is expected to disrupt the correct folding and processing of miR-998. Indeed, no mature miR-998 was detected in whole 3rd instar larvae, larval eye imaginal discs, or adult heads from mir-998exc222 mutant animals (Figure 2A). Importantly, the expression of dE2f1 and miR-11 were not affected in the mir-998exc222 mutant animals (Figure 2B). We concluded that mir-998exc222 is a null allele of mir-998.
Clones of mir-998exc222 homozygous mutant cells in rbf mutant eye discs were generated using the ey-Flp/FRT system and apoptotic cells were visualized with the C3 antibody. The mir-998 mutant tissue was marked by the absence of GFP, while tissue that contained a wild-type mir-998 allele expressed GFP. Since the entire disc was mutant for rbf, cell death occurred in the distinctive pattern in the morphogenetic furrow in both mir-998 mutant and mir-998 wild type tissue. However, significantly elevated C3 staining was consistently observed in clones of mir-998exc222 mutant cells compared to adjacent mir-998 wild-type tissue (Figure 2C). In contrast, no apoptosis was detected in the furrow when clones of mir-998 mutant cells were induced in wild type eye discs (Figure 2D). Thus, the loss of mir-998 specifically sensitized rbf mutant cells to apoptosis, while overexpression of miR-998 blocked cell death in rbf mutants.
How does miR-998 suppress apoptosis of rbf mutant cells? The specific pattern of apoptosis in rbf mutant eye discs is due to the coincident transient reduction of EGFR signaling in the morphogenetic furrow that, in turn, lowers prosurvival cues. As a result, the level of unrestrained dE2F1 becomes sufficient to trigger cell death in this region of rbf mutant but not in wild type eye discs [13]. In addition, the cell death in rbf mutants was shown to be dependent on the expression of the pro-apoptotic genes reaper (rpr) and hid. Therefore we examined impact of miR-998 on EGFR signaling and on hid and rpr.
We began by determining whether miR-998 regulates expression of rpr and hid in luciferase sensor assays. The 3′UTRs of rpr and hid were cloned downstream of the constitutively expressed luciferase gene. Increasing amounts of miR-998 were co-transfected with rpr or hid 3′UTR sensors, and luciferase activity was measured. As shown in Figure 3A, expression of miR-998 did not modulate rpr or hid 3′ UTR reporters. To corroborate this result we compiled a list of predicted miR-998 targets and performed Gene Ontology of Biological Processes (GOBP) enrichment analysis. Interestingly, none of the GOBP terms associated with apoptosis were statistically enriched among miR-998 targets (Figure 3B, Table S2). In contrast, as it has been shown previously [8], GOBP terms that relate to the induction and positive regulation of cell death were significantly enriched among miR-11 targets (Figure 3B, Table S2). Furthermore, although miR-11 and miR-998 share 170 common targets, cell death GOBP terms were not enriched among them but were overrepresented among genes that are exclusively miR-11 targets. Thus, the sensor assays and bioinformatics analyses do not support the explanation that suppression of apoptosis in rbf mutants by miR-998 occurs through the direct regulation of cell death genes.
Next, we asked whether EGFR activity is altered in mir-998 mutants. The level of EGFR activity is accurately reflected by di-phosphorylated, activated ERK (dpERK) [16]. During eye development, EGFR signaling is transiently reduced within the morphogenetic furrow, while EGFR activity is high in groups of cells that form the ommatidial preclusters in a column immediately posterior to the morphogenetic furrow, which is revealed by the dpERK antibody. Within a column, clusters are specified sequentially in short intervals beginning at the midline. This gives rise to a gradual rise and fall of dpERK staining within a column (Figure 4A, left panel). To examine the level of EGFR signaling in mir-998 mutants, clones of mir-998exc222 mutant cells were generated in rbf120a mutant eye discs and stained with the dpERK antibody. While the pattern of natural variation of dpERK staining was not altered in mir-998 mutant tissue, the intensity of dpERK staining was reduced within clones of mir-998 mutant cells, as well as in wild type tissue immediately adjacent to the clonal boundary (Figure 4A, see yellow arrowheads). Since the level of dpERK expression reflects the level of EGFR signaling, this indicated that the loss of mir-998 reduces EGFR activity and this may explain the enhancement of apoptosis in rbf mutants.
EGFR signaling is used reiteratively throughout development, including in the recruitment of photoreceptor cells into the ommatidial clusters of the developing larval eye [17]–[19]. To confirm that EGFR signaling is reduced in mir-998 mutants, we performed genetic interaction tests between the mir-998 mutant allele and the dominant gain-of-function Ellipse (Elp) allele of the Egfr gene. The number of ommatidial clusters is significantly reduced in EgfrElp/+ larval eye discs as revealed by staining with an ELAV antibody that marks differentiated photoreceptors (Figure 4B and [18], [20]). Strikingly, the mir-998exc222 mutation strongly suppressed this phenotype: there was a dramatic increase in the number of ommatidial clusters in EgfrElp/+, mir-998exc222 double mutant eye discs compared to EgfrElp/+ single mutant eye discs (Figure 4B). Consistently, the small, rough eye phenotype of EgfrElp/+ adult flies was suppressed by the loss of mir-998 (Figure 4B). To determine whether the effect of mir-998 is cell autonomous, we generated clones of mir-998exc222 mutant cells in heterozygous EgfrElp/+ eye discs. As shown in Figure 4C, the EgfrElp mutant phenotype was partially suppressed in cells that were immediately adjacent to the mir-998exc222 mutant tissue suggesting some non-cell autonomous effects. Therefore, the loss of mir-998 suppresses the EGFR gain-of-function phenotype, which is consistent with the reduction of dpERK activity (Figure 4A) and therefore EGFR signaling in mir-998 mutant tissue.
To gain insight into the molecular basis of the genetic interaction between miR-998 and EGFR signaling, we asked whether the loss of mir-998 results in misexpression of gene(s) that are known to be connected to the EGFR signaling pathway. To identify such genes in an unbiased manner we performed gene expression microarrays using RNA isolated from mir-998exc222 homozygous mutant and wild type eye discs. This analysis led to identification of a set of 382 genes that were differentially expressed (DE) in the mir-998 mutant. The list of DE genes was then mined for genes with terms related to the EGFR signaling pathway by referring to AmiGO and KEGG pathway databases, as well as the literature [21]–[23] (Table S3).
Of 80 genes associated with the EGFR pathway, only one was significantly differentially expressed in the mir-998exc222 microarray: dCbl-S. dCbl negatively regulates signaling from EGFR by binding to the activated, phosphorylated receptor, and inducing its ubiquitination and endocytosis [24]–[27]. According to the microarray data, the expression of dCbl-S was increased 3.6-fold in the absence of mir-998, which made it a good candidate as a target of repression by miR-998, the prevailing mechanism of miRNA function. To confirm the results of the gene expression microarray, dCbl expression was measured in mir-998 homozygous mutant eye discs by quantitative RT-PCR (qRT-PCR). The dCbl gene encodes two dCbl proteins generated from alternatively spliced transcripts which, like their mammalian Cbl homologs, both negatively regulate EGFR signaling in Drosophila [28], [29]. As shown in Figure 5A, both dCbl transcripts were significantly upregulated in mir-998exc222 eye discs (dCbl-S: 4.4-fold, dCbl-L: 2.3-fold). Together with the results of genetic interaction tests described above, this suggested a model where miR-998 represses dCbl, a negative regulator of EGFR signaling, to enhance signaling downstream of the EGF receptor.
This model was tested in three sets of experiments. First, we asked whether dCbl is a miR-998 target. Since dCbl was not identified as a target of miR-998 by bioinformatic prediction, we asked whether the mammalian ortholog, c-Cbl, was a predicted target of the mammalian miRNA members of the mir-29/998 miRNA seed family: mir-29a-c [30] (Figure 5B and Table S1). Analysis revealed that c-Cbl contains five predicted miR-29 target sites including two sites in tandem within a highly conserved region that encodes the tyrosine kinase-binding (TKB) domain. Further analysis of dCbl using the RNA22 prediction algorithm with low stringency parameters [31] revealed a putative target site for miR-998 in dCbl-L (Figure 5C). The functionality of these sites was then tested in luciferase sensor assays. The identified predicted target sites for miR-998 in dCbl were cloned downstream of the Renilla luciferase gene, and were transfected with increasing amounts of a miR-998 expression plasmid. Significantly, miR-998 exhibited dose-dependent repression of luciferase 3′UTR sensor constructs carrying either the double miR-29 target site within the highly conserved TKB domain (site 1), or the miR-998 site near the 3′ end of the dCbl-L coding sequence (site 2) (Figure 5C). Moreover, the repression of site 1 and site 2 was completely blocked when the target sequences were mutated. Thus, miR-998 directly repressed luciferase sensor constructs carrying dCbl target sequences. While our sensor assay results are consistent with the notion that miR-998 directly represses dCbl in eye discs, we acknowledge the possibility that miR-998 represses dCbl indirectly through a different mechanism in vivo.
Since dCbl is a negative regulator of EGFR signaling, its elevated expression in mir-998 mutants may explain the increased apoptosis in the morphogenetic furrow of rbf, mir-998 double mutants. To test this idea, we examined the effect of dCbl overexpression on cell death in rbf mutant eye discs. When dCbl-S or dCbl-L were expressed in rbf120a mutant eye discs using the Flip-out system, a wider band of C3 antibody staining was detected than in the control rbf120a mutant eye disc. Importantly, co-expression of miR-998 blocked the increase in cell death induced by dCbl, which is consistent with the notion that miR-998 can repress dCbl expression in vivo (Figure 5D). In the converse experiment, clones of dCbl mutant cells were generated in rbf120a mutant background and mosaic eye discs were stained with the C3 antibody. Consistent with results of dCbl overexpression described above, the loss of dCbl strongly suppressed apoptosis in rbf mutants as the number of C3 positive cells was dramatically reduced in the dCbl mutant tissue (Figure 5E). Therefore miR-998 represses dCbl, a negative regulator of EGFR signaling that is functionally important for triggering cell death, in the morphogenetic furrow of rbf mutant eye discs. Furthermore, dCbl is a critical target of miR-998 in modulating apoptosis in rbf-deficient cells since overexpression of dCbl in eye discs mimics the mir-998 mutant phenotype, while the dCbl mutant mimics the miR-998 overexpression phenotype.
miR-998 is part of the mir-29 seed family of miRNAs, which is defined by having identical seed sequences ([30] and Figure 5B). The presence of multiple miR-29 target sites in mammalian c-Cbl raises the question of whether miRNA-dependent regulation of dCbl is conserved in mammals. To address this question we generated and expressed three different luciferase sensors carrying single, or paired predicted miR-29 target sites from the human c-Cbl gene, along with increasing amounts of a miR-29a expression plasmid (Figure 6A). While some miRNAs exert strong repression of their targets, others including miR-29 have been shown to elicit more modest effects in sensor assays and in vivo [32]–[36]. Indeed, miR-29a exerted modest but clearly dose-dependent repression of a sensor carrying two sites in tandem in the TKB domain near the 5′ end of the CDS (site 1). miR-29a also repressed a luciferase sensor carrying a predicted miR-29 site in the 3′UTR (site 2), while a sensor carrying tandem sites near the 3′ end of the 3′UTR failed to respond to miR-29 (site 3) (Figure 6A). To address the specificity of repression, we introduced mutations into the mir-29 target sequences in the c-Cbl site 1 and site 2 luciferase sensors and found that miR-29 did not repress these mutant sensors, regardless of the amount of transfected miR-29. Therefore, even though miR-29 modestly repressed sensors containing site 1 and site 2, this repression was specific since the response to miR-29 was dose-dependent, while mutating the sites completely blocked the effect of miR-29. We concluded that miR-29 seed family members can directly target c-Cbl through corresponding paired sites that are present within the highly conserved region in both Drosophila and human genes encoding the TKB domain. In addition, Drosophila and human miR-29 seed family members can regulate Cbl sensors through distinct target sites that are not conserved between the two species (Figure 5C and Figure 6A).
Having established the functionality of putative miR-29 target sites in the c-Cbl sequence using sensor assays, we asked whether the endogenous c-Cbl could be repressed by miR-29. HeLa cells were transfected with a miR-29 expression plasmid or an empty vector and the level of c-Cbl was analyzed by Western blot analysis 48 hr after transfection. As shown in Figure 6B, the expression of endogenous c-Cbl protein was significantly reduced in cells expressing miR-29 compared to the vector control. Importantly, downregulation of c-Cbl was accompanied by an elevated level of di-phosphorylated ERK (dpERK), which reflects an increase in EGFR/MAPK activity, in response to miR-29a expression.
Previous work showed that in the absence of functional Cbl, the sensitivity of cells to signals from the extracellular milieu is enhanced, and cells exhibit increased growth factor-induced motility in wound-healing scratch assays. The cell migration is mediated in part by ERK-MAPK signaling [37]–[39]. Therefore we tested whether reduction of c-Cbl by miR-29a expression alters the rate of wound-healing in scratch assays. Scratch assays were performed on HeLa cells transfected with constructs expressing miR-29a, c-Cbl, c-Cbl and miR-29a, or an empty vector control. The area of the wound was measured 0, 17.5, and 25.5 hours after the scratch was introduced (Figure 6C and 6D). As expected, cells expressing c-Cbl exhibited decreased motility and delayed wound closure compared to the control. In contrast, expression of miR-29a significantly increased the rate of wound healing compared to both cells expressing c-Cbl, and cells transfected with empty vector. miR-29a also increased the rate of wound healing in cells transfected with c-Cbl, which was consistent with the notion that miR-29 directly represses c-Cbl expression and, as a consequence, its function in vivo. 25.5 hours after the scratches were introduced, cells expressing miR-29a had closed all but 24% of the original area of the scratch wound, while the scratch wound of controls cells occupied 39% of the original area, and the scratch in cells expressing c-Cbl was 55% of the original area (Figure 6C and 6D). Co-expression of lower and higher amounts of miR-29 with c-Cbl led to scratches that were 38% and 34% the original area after 25.5 hours, suggesting that miR-29 can suppress the function of its target in vivo.
From these experiments we concluded that c-Cbl expression is modulated by miR-29a in mammalian cells and that miR-29a represses endogenous c-Cbl expression. Importantly, repression of c-Cbl expression is accompanied by increased ERK-MAPK signaling and an elevated rate of growth factor-regulated cell migration.
The potential for complex interactions between intronic miRNAs and their host is illustrated by the Drosophila dE2f1 gene, and the two miRNAs embedded in its last intron: mir-11 and mir-998. We previously showed that mir-11 directly represses a subset of apoptotic genes that are transcriptional targets of the host gene dE2f1, and in doing so, miR-11 limits E2F-dependent cell death induced by DNA damage. Here, we identified a novel layer of regulation in the dE2f1 locus as miR-998 enhances EGFR cell survival signaling, thereby suppressing E2F-dependent cell death in rbf mutant animals. Therefore, the proapoptotic function of dE2F1 is under intrinsic control by two distinct and complementary mechanisms.
Many miRNAs elicit relatively weak effects and therefore their mutant phenotypes are rather subtle. Therefore it is not surprising that both the mir-11 and mir-998 mutant alleles were viable, and exhibited no phenotypes on their own. The lack of a mutant phenotype represents a major hurdle in identifying the physiological function of a miRNA [2], [40]. A number of approaches have been taken to reveal miRNA functions such as generating compound mutants, and analyzing mutant phenotypes in the context of disruptions to core regulatory pathways [3]. In our work we have used a different strategy and investigated miRNAs in the context of the function of their host gene. This novel approach turned out to be highly informative and allowed us to identify the elusive functions of two intronic miRNAs embedded within the dE2f1 gene. Notably, both miRNAs exhibited phenotypes only in E2F-sensitized backgrounds but lacked phenotypes on their own. One implication of our work is that the functions of intronic miRNAs can be linked to their host gene, and where known, the host gene function can be exploited to uncover the physiological roles of embedded miRNAs. This idea is particularly relevant given that approximately 40% of all miRNAs are embedded in protein-coding genes and therefore such approach can be applicable beyond the dE2f1 locus.
In various systems, inactivation of Rb provides a cellular context to investigate E2F-dependent apoptosis. Interestingly, animal models revealed that not every Rb mutant cell is equally sensitive to apoptosis. In the Drosophila rbf mutant eye disc, apoptosis occurs in a highly reproducible pattern that is determined by the level of prosurvival signaling from the EGF receptor [13]. Our data show that the loss of mir-998 enhanced cell death in rbf mutant eye discs but did not alter the overall pattern of apoptosis. Notably, we did not find evidence that miR-998 directly repressed cell death genes. Therefore, miR-998 is unlikely to function by altering the expression of apoptosis genes. Rather, our results support a scenario where miR-998 represses cell death in rbf mutants by enhancing pro-survival EGFR signaling. Using genome-wide approaches we identified dCbl as a highly upregulated gene in mir-998 mutant eye discs. Genetic interaction tests demonstrated that dCbl phenocopies the effect of mir-998 on dE2f1-dependent cell death in rbf mutants. Therefore dCbl behaves as a critical player that mediates the effect of mir-998 on apoptosis. We acknowledge that our data do not rule out the possibility that the effect of miR-998 on dCbl is indirect. For example, miR-998 may function by limiting the level of a positive regulator of dCbl expression. However, miR-998 can directly repress dCbl in sensor assays and this repression occurs in a sequence-dependent manner mediated by three different miR-998 target sites in dCbl. Similarly, a miR-998 seed family homolog, mir-29 repressed mammalian c-Cbl sensors in HeLa cells. Thus, while the mechanism of regulation of Cbl by the mir-998/mir-29 seed family in vivo may involve indirect or direct regulation, we favor a model where miR-998 modulates EGFR signaling by directly regulating dCbl. Further testing of this model would require introducing mutations in the miR-998 binding sites in the endogenous dCbl gene in order to generate dCbl alleles that are insensitive to direct targeting by miR-998.
Cbl is a negative regulator of EGFR signaling that binds the activated receptor and induces its ubiquitination and subsequent endocytosis, after which either further downregulation occurs through receptor degradation, or the receptor is retained in endosomes, or recycled back to the plasma membrane [24]. In the eye disc, the transient decrease in EGFR signaling prior to ommatidial specification occurs in part through the sequestration of the EGFR ligand, Spitz, which limits communication from neighbouring cells [13], [41], [42]. Moreover, expression of an EGFR isoform that cannot transmit ligand-initiated signals also stimulated EGFR-dependent cell death in rbf mutants [13]. Consistent with this mode of regulation, we showed that changes in the levels of dCbl, which limits EGFR signaling from the plasma membrane, modulated E2F-dependent cell death in rbf mutants. We suggest that through repression of dCbl expression, miR-998 supported ligand-dependent EGFR signaling in this context, which limited the proapoptotic activity of dE2F1.
miR-998 belongs to the miR-29 seed family of miRNAs, which also includes miR-285 and miR-998 in Drosophila [30]. While no mutant phenotypes have been reported, a recent investigation of gain-of-function phenotypes showed similar wing defects caused by overexpression of miR-285 and miR-995, although their molecular basis unknown [43]. Similarly, the functions of miR-29 seed family members cel-miR-49 and cel-miR-83 have not been reported. Seed family members in humans include the less abundant mature miRNA generated from the mir-21 oncomir, mir-21* (mir-21-3p), and mir-593* (mir-593-5p), and miR-29a, miR-29b and miR-29c. Unlike for other seed family members, a number of functions and targets of miR-29 have been reported. While it is not clear whether these functions and targets are common to other seed family members, this possibility warrants further investigation.
In humans, two intergenic mir-29 clusters give rise to three different mature miR-29 miRNAs that share an identical seed sequence, but differ slightly in their 3′ sequences. miR-29 was shown to disrupt epithelial polarity, and cooperate with oncogenic Ras in inducing epithelial-to-mesenchymal transition (EMT) and metastasis through the repression of the tristetraprolin nuclease [44]. Here, we identify a previously unknown miR-29 target, c-Cbl, which functions upstream of Ras in many signaling pathways. This novel link raises the question of whether miR-29 modulates receptor tyrosine kinase-induced EMT. Additional evidence for an oncogenic function of miR-29 came from the identification of PTEN as a miR-29 target [45]. Moreover, the oncogenic viral miRNA BLV-miR-B4 shares the same seed sequence as miR-29 and directly regulates miR-29 targets peroxidasin and HBP1 [46]. However, in different contexts, miR-29 has been shown to function as a tumor suppressor [47]–[49], and miR-29c repressed cancer cell proliferation, and limited E2F activity through the indirect activation of Rb [50]. It is not yet clear whether either of the two mir-29 clusters are transcriptionally regulated by E2F, although the expression of miR-29a is induced in the G1 phase of the cell cycle [51], which is coincident with increased E2F activity.
c-Myc bound both mir-29 cluster gene promoters and repressed miR-29 expression, suggesting that miR-29 is directly regulated by c-Myc [52]. Furthermore, expression of miR-29 lead to a decrease in the expression of E2F targets Cyclin A2, MCM2, and PCNA [52]. It is currently unclear whether the expression of E2F and miR-29 are linked directly, or through c-Myc, which is well-known to induce the expression of E2F1. We note that while miR-998 and miR-29 share at least one common target, their mechanisms of interaction with E2F function may differ. Interestingly, overexpression of dMyc in Drosophila embryos lead to both cell death, and decreased expression of miR-998 [53]. It is not known whether miR-998 could block dMyc-induced cell death in Drosophila embryos, or whether this would be accomplished through repression of dCbl, and a consequent increase of EGFR signaling.
EGFR/MAPK and Rb/E2F pathways intersect in the regulation of proliferation and cell death and are frequently disrupted in cancer. Our results reveal a novel connection between Rb/E2F and EGFR signaling: miR-998 interacts with both pathways and integrates their activities to effect an overall cellular response. This link may represent an attractive target to intentionally perturb cellular homeostasis thereby sensitizing cells to therapeutic reagents in the treatment of malignancy.
All fly crosses were done at 25°C. The following stocks were obtained from Bloomington Drosophila Stock Center at Indiana University: GMR-Gal4, w; Dr/Δ2-3 99B, Sb, dE2f1EY05005 (FBal0160590), dE2f17172 and; egfrE1. The following stocks were previously published: UAS-miR-11 (Brennecke et al. 2005), Act88F-Gal4 and Act88F-Gal4, UAS-dE2f1 (from Erick Morris and Teiichi Tanimura), CblF165, FRT 80B, UAS-Cbl-L (A18) and UAS-Cbl-S (A1) (from Trudi Schupbach), rbf1120a, ey-FLP; act5c>CD2>GAL4, UAS-GFP/CyO, GFPAct, and rbf1120a, ey-FLP/FM7, GFPAct;; FRT 82B, GFPUbi, and rbf1120a, ey-FLP/FM7, GFPAct;; GFPUbi, FRT 80B (from Nam Sung Moon). dE2F1Δ1 is a deletion which lacks the genomic region between P element insertion P[XP]E2fd01508 and piggyBac insertion PBac[RB]InRe01952 and generated according to [54]. E2f1d01508 (FBal0183912), and InRd03668 (FBti0055281) were obtained from the Exelixis collection at Harvard Medical School. Generation of mir-998exc222 mutant allele is described in Protocol S1.
For in vivo expression of miR-998, an insert encoding miR-998 was cloned into the pUAST plasmid using standard molecular cloning techniques. This insert contained the de2f1 intron 5′ sequence flanking two mir-1 chimeras: mir1/mCherry shmiR [55] replaced the mir-11 gene, and mir-1/998 replaced the mir-998 gene, which was designed as in Haley et al. (2008) and was synthesized by GenScript USA. Flies carrying UAS-miR-998 were generated by P-element transformation at The Best Gene, Inc.
HeLa cells were seeded in 6-well plates, transfected as described, and cultured until 100% confluent. Straight scratches were made across the cell layer using a 0.2 ml pipette tip. The cells were then gently washed three times with PBS to remove cellular debris and the media was replaced at 0 and 17.5 hours. Photographs of the wound region were taken using a Zeiss AxioObserver A1 microscope and AxioCam IC camera. The wound area was calculated using Image J software.
Sequences were cloned downstream of the Renilla luciferase coding sequence in the psiCheck2 (Promega) plasmid using standard cloning techniques. See Table S6 for sensor sequences.
HeLa cells were cultured in DMEM+10% FBS, and were transfected with the X-treme Gene HP transfection reagent (Roche) according to the manufacturer's protocol. Cells were harvested 24–48 hours post-transfection. pcDNA3/mir-998 generated by PCR amplification of the mir-998 gene and standard molecular cloning techniques. pcDNA3/hsa-mir-29a was generated by insertion of a GeneString (Invitrogen) with the mir-29a gene sequence in pcDNA3. Insert sequences were verified by sequencing analysis. Sequences are in Table S4. Firefly and Renilla luciferase activity were measured using the Dual Luciferase Assay protocol (Promega).
Antibodies: c-Cbl 1∶200 (sc-170), from Santa Cruz; di-phosphorylated p42/p44 ERK 1∶5000 (M8159) from Sigma-Aldrich; mouse anti-tubulin 1∶10000 (cat# T9026) from Sigma-Aldrich; goat anti-rabbit-HRP (#31460) and goat anti-mouse-HRP (#31430) from Thermo Fisher.
Antibodies used were as follows: rabbit anti-C3 (Cleaved Caspase3), lot 26, 1∶75 (Cell Signaling), mouse anti-BrdU 1∶50 (Beckton Dickinson), rat anti-ELAV 1∶50, (Developmental Studies Hybridoma Bank), phosphorylated p42/p44 ERK 1∶200 (M8159) from Sigma-Aldrich, and Cy3-, and Cy5- conjugated anti-mouse, and anti-rabbit secondary antibodies (Jackson Immunoresearch Laboratories). Larval tissues were fixed in 4% formaldehyde in phosphate-buffered saline (PBS) for 30 minutes, permeabilized in 0.3% Triton X-100 in PBS twice for 10 minutes each, blocked in PBS with 0.1% Triton X-100 for 30 minutes at 4°C, and then incubated with antibodies overnight at 4°C in10% normal goat serum, and 0.3% Triton X-100 in PBS. After washing three times for 10 minutes each at room temperature in 0.1% Triton X-100 (in PBS), samples were incubated with appropriate conjugated secondary antibodies for 45 minutes at room temperature in 10% normal goat serum, and 0.3% Triton X-100 (in PBS). After washing with 0.1% Triton X-100 (in PBS), tissues were stored in glycerol+antifade reagents and then mounted on glass slides. To detect S phases, dissected larval eye discs were labeled with BrdU for 2 hrs at room temperature and then fixed overnight in 1.5% formaldehyde, 0.2% Tween 20 in PBS at 4°C. Samples were then digested with DNase (Promega) for 30 minutes at 37°C. Samples were then probed with primary and secondary antibodies as described above. All immunofluorescence was done on a Zeiss Confocal microscope and images were prepared using Adobe Photoshop CS4. All images are confocal single plane images unless otherwise stated as projection images. A minimum of 10 larvae were used for each analysis.
Comprehensive lists of predicted miR-998 targets and miR-11 targets were compiled from TargetScan [56], MinoTar [57], PITA [58], miRanda [59], and RNAhybrid [60]. Target predictions use for hsa-miR-29 were from TargetScan [61].
Total RNA was isolated from 10 adult heads, 10 larvae, or 30–50 eye discs, with TRIzol (Invitrogen). Reverse transcription to measure standard mRNAs was performed using the iScript kit (BioRad) according to manufacturer's specifications. Quantitative PCR was performed with the SYBR Green I Master (Roche) on a Light Cycler 480 (Roche). miR-11 and miR-998 were measured by Taqman assay (Applied Biosystems). Primer sequences are in Table S5.
See Protocol S1.
Microarray gene expression data were analyzed using “Affy” package [62] and differential expression analysis by “Limma” package [63]. Functional and pathway enrichment analysis of differentially expressed genes were counducted using Gitools [64]. See Protocol S1 for detail.
|
10.1371/journal.pcbi.1005990 | Imaging of neural oscillations with embedded inferential and group prevalence statistics | Magnetoencephalography and electroencephalography (MEG, EEG) are essential techniques for studying distributed signal dynamics in the human brain. In particular, the functional role of neural oscillations remains to be clarified. For that reason, imaging methods need to identify distinct brain regions that concurrently generate oscillatory activity, with adequate separation in space and time. Yet, spatial smearing and inhomogeneous signal-to-noise are challenging factors to source reconstruction from external sensor data. The detection of weak sources in the presence of stronger regional activity nearby is a typical complication of MEG/EEG source imaging. We propose a novel, hypothesis-driven source reconstruction approach to address these methodological challenges. The imaging with embedded statistics (iES) method is a subspace scanning technique that constrains the mapping problem to the actual experimental design. A major benefit is that, regardless of signal strength, the contributions from all oscillatory sources, which activity is consistent with the tested hypothesis, are equalized in the statistical maps produced. We present extensive evaluations of iES on group MEG data, for mapping 1) induced oscillations using experimental contrasts, 2) ongoing narrow-band oscillations in the resting-state, 3) co-modulation of brain-wide oscillatory power with a seed region, and 4) co-modulation of oscillatory power with peripheral signals (pupil dilation). Along the way, we demonstrate several advantages of iES over standard source imaging approaches. These include the detection of oscillatory coupling without rejection of zero-phase coupling, and detection of ongoing oscillations in deeper brain regions, where signal-to-noise conditions are unfavorable. We also show that iES provides a separate evaluation of oscillatory synchronization and desynchronization in experimental contrasts, which has important statistical advantages. The flexibility of iES allows it to be adjusted to many experimental questions in systems neuroscience.
| The oscillatory activity of the brain produces a repertoire of signal dynamics that is rich and complex. Noninvasive recording techniques such as scalp magnetoencephalography and electroencephalography (MEG, EEG) are key methods to advance our comprehension of the role played by neural oscillations in brain functions and dysfunctions. Yet, there are methodological challenges in mapping these elusive components of brain activity that have remained unresolved. We introduce a new mapping technique, called imaging with embedded statistics (iES), which alleviates these difficulties. With iES, signal detection is constrained explicitly to the operational hypotheses of the study design. We show, in a variety of experimental contexts, how iES emphasizes the oscillatory components of brain activity, if any, that match the experimental hypotheses, even in deeper brain regions where signal strength is expected to be weak in MEG. Overall, the proposed method is a new imaging tool to respond to a wide range of neuroscience questions concerning the scaffolding of brain dynamics via anatomically-distributed neural oscillations.
| The role of neural oscillations in population codes of brain functions, and the possible mechanisms of inter-regional communication between brain regions are not entirely understood. Source imaging techniques with magnetoencephalography (MEG) or electroencephalography (EEG) are time-resolved, non-invasive tools used to test a great diversity of neurophysiological hypotheses [1]. In principle, MEG/EEG imaging can map multiple regional sources of oscillatory activity from external sensor data. However, spatial smearing and heterogeneous signal strength across brain locations limits the performance of current source imaging methods. Consequently, if nearby brain regions express an effect of interest, the area of stronger magnitude will mask the detection of weaker sources, as illustrated in Fig 1. The detection of multiple oscillatory sources therefore remains challenging to MEG/EEG imaging. This limits the insight about distributed brain dynamics that can be gained from the technique.
MEG/EEG localization of oscillatory generators typically relies on a procedure that is non optimal in terms of signal detection. Source time-series are first reconstructed using imaging or beamforming approaches [2]. Second, inferential statistics based on the experimental hypothesis are tested at each voxel of the source space—e.g., using the ratio of oscillatory power between two experimental conditions. Significant and spatially-distinct regional clusters are then interpreted as distinct sources of oscillations. This approach hinders the detection of weaker or deeper sources in the presence of stronger regional activity. We refer to this methodology as the standard approach.
We introduce a novel methodology to alleviate this problem. The technique performs imaging with embedded inferential and group prevalence statistics (iES) altogether. With iES, the experimental hypothesis is not deferred to the stage of statistical inference on the estimated source values. Rather, it explicitly constrains the solution to the hypothesis tested. In essence, iES reduces the (spatial) dimensions of the data, to detect and equalize the contribution of source components that are consistent with the tested hypothesis. The iES methodological apparatus is based on generalized eigen decompositions [see e.g. 3], nonparametric statistics [4, 5] and subspace scanning [MUSIC, see 6]. The paper is organized as follows: Results: We first give a high-level description of the iES approach, starting with the basic principles and then illustrating a full group analysis with an MEG dataset. We then illustrate the advantages of iES using experimental data and simulations, and show that iES 1) has key statistical advantages, yielding improved detection sensitivity, 2) can be used in conjunction with the standard approach, for complementary estimation of source strengths, 3) improves the detection of functionally connected regions, and that 4) iES can implement a wide range of experimental hypotheses. Discussion: We then put iES in the context of previous work and discuss limitations. Materials and Methods: Finally, we provide all the experimental details and the full mathematical formulation of the iES approach.
We describe the basic principles of iES and illustrate the steps involved using a MEG data example. The method per se is detailed in Materials and Methods.
The iES source maps highlight sources whose signals are consistent with a directed hypothesis across a group of subjects. When two experimental conditions are contrasted, this implies that two distinct source maps can be produced: for instance in the previous case example, one map corresponding to increased oscillatory power in one condition over the other; the other map corresponding to decreased oscillatory power. The benefit resulting from this is that mutual interference in the detection and statistical evaluation of the two sets of sources is avoided.
We demonstrate these methodological assets using the same experimental MEG data as above. We analyzed task-induced oscillations in the beta band (13-30 Hz), with the hypothesis that they were strongly suppressed during attention-demanding tasks in the occipital visual cortex [9, 11]. We also wished to test whether other brain regions would reveal a selective increase in beta power during stimulus presentation. This contrast thus serves to illustrate how a strong power effect (decreased beta power) can challenge the detection of weaker opposite responses (increased beta power) with the standard approach, but not with iES.
Fig 4 shows results for the hypothesis of increased beta band power during stimulus presentation. The data from an example subject (Panel a) contained one spatial component consistent with that hypothesis. At the group level, only eight subjects out of 17 showed the effect of interest. Here, we shall emphasize the importance of the notion of effect prevalence, since the majority null hypothesis could not be rejected (Fig 4b). However, the prevalence null hypothesis can be rejected up to γ0 = 0.22, which indicates there is a subgroup of the population from which our subjects were drawn, which shows the hypothesized effect. To better illustrate the significance of this notion, let us first assume the effect is not present in the population. With a probability of 0.95, one may still observe out of chance an effect in up to 3 out of the 17 subjects. The prevalence test therefore indicates that the observed data are unlikely under the assumption that prevalence is 22% or less (at a false positive rate of p < .05). Thus we pursued further the analysis of the subgroup of 8 participants (see Fig 4b), bearing in mind that the results may not generalize to the majority of the population. The validity of such a decision depends on whether the scientific question is pertinent to generic vs. restricted effects among participants. For instance, it can be particularly valuable for identifying effects that are more specific of a sub-type of participants in terms of behaviour or clinical condition.
Fig 4c and 4d shows typical signal traces in a subject from the subgroup presenting stronger beta and alpha oscillations building up during stimulus presentation. The sharp waveforms and the combined alpha/beta spectral pattern were typical of the somatosensory mu rhythm [12, 13]. The effect was localized to right postcentral regions, as shown in the example subject and the group subcorr maps (Fig 4e and 4f). This result replicates previous observations of lateralized beta oscillations during an attention-demanding task [14].
To compare these findings with those from the standard approach, we obtained source maps of log-power ratios using minimum-norm imaging kernels. We used the MNE implementation of Brainstorm, with default parameters [15, see also descriptions in Materials and Methods]. The resulting maps were statistically thresholded following the same permutation procedure based on the maximum statistic. Note that with this procedure, distinct maps of positive and negative effects cannot be produced. For comparison purposes, we used the data from the subgroup (n = 8) that showed the desired effect of higher beta power during stimulus presentation.
Fig 5a shows the complete beta band iES results (i.e. increases and decreases). In addition to the increased stimulus-induced beta power over right postcentral regions, we observed beta suppression in the visual cortex. The prevalence assessment revealed that this latter effect was observed in the entire group, and thus may generalize to the majority of the population. In the minimum-norm map (Panel b), the suppression of beta oscillations in the visual cortex was also readily observed, with similar spatial extension. However the increased, stimulus-induced beta oscillations over the right central regions were absent from the minimum-norm map produced from the 8 subjects presenting the effect in iES. The non-thresholded maps are shown in Supplementary Material, and confirm that a positive peak was indeed present in the minimum-norm maps, but was not deemed statistically significant. The reason for the observed discrepancy between methods can be understood from the permutation and data histograms (Fig 5, right column): By definition, the permutation histograms of the log-power ratios were mirror images for the evaluation of positive and negative effects respectively. This was the case because we drew exhaustive permutations from the data from the 8 subjects (28 = 256). Thus every unique permutation of labels had a corresponding opposite permutation. The consequence was that the variance and spread of the resulting distribution were determined by the strongest effect in magnitude—here the negative effect of beta suppression. The histogram of the observed data indicated that the right tail of the histogram indeed did not reach the statistical threshold. The iES allowed to test two directed hypotheses separately. Hence the permutation distributions were distinct and adapted to each respective hypothesis, revealing the positive effect in the iES statistical source map that was absent in the standard approach.
We detail in Methods that iES requires the estimation of cross-spectral or covariance matrices, and their decomposition in the generalized eigenvalue framework. This means that, in addition to the subcorr statistical maps produced, a corresponding map of the standard approach can be obtained by applying a minimum-norm imaging kernel to those matrices, which allows plotting the value of the quality function f at each location of the source grid. Fig 5 shows an example of this approach to obtain a map of log-power ratios. We emphasize that the combined use of subcorr and minimum-norm source maps enabled by the proposed method provides complementary information with respect to the experimental hypothesis of interest.
We demonstrate such benefit using the same visual-attention MEG data, to detect the origins of narrow-band oscillations (Fig 2a). The corresponding quality function fnarrow quantifies the ratio of signal power in a frequency range of interest with respect to the total power of the broadband signal. Such a quality function highlights signals with a peaky spectral profile [16], which is of specific interest when studying stimulus-independent ongoing oscillations. We used the data of the ongoing visual stimulus period ([1, 3] s after stimulus onset) to investigate the anatomical origins of three frequency bands of interest: theta (4-8 Hz), alpha (8-13 Hz) and beta (13-30 Hz). The reference broadband signal against which to contrast possible effects in the narrow frequency bands of interest was taken between 2 and 100 Hz.
We compared the subcorr statistical maps with the minimum-norm maps of fnarrow (Fig 6). The log-transform of the ratios was not applied because negative effects were of no interest to the question, thus a symmetric measure was not required. A threshold 0 < f n a r r o w * < 1 for selecting relevant signal subspace patterns was computed with the bootstrap procedure described in Materials and Methods. In the alpha and beta bands the results were similar between our approach and standard imaging. Commonly observed brain regions as strong sources of these ongoing rhythms were found [see e.g., 17]. Alpha activity was prominent in medial occipital-parietal regions; beta activity was stronger over bilateral sensory-motor regions. Alpha band oscillations were also found prominently over the right postcentral region, which parallels the finding of enhanced alpha and beta power during the stimulus period in the same brain area, as shown in the previous section.
We found differences between iES and minimum-norm maps in the theta band. The subcorr statistical map revealed involvement of the medial temporal lobes (MTL) bilaterally, and of medial frontal/anterior cingulate regions. Theta oscillations in MTL, including the hippocampus and parahippocampal regions, have been extensively described [18]. Due to their relative depth and therefore lower MEG signal-to-noise ratios (SNR), they have been considered more challenging to detect [19, 20, 21]. The MNE power-ratio maps though showed a lateralized distribution of theta activity in the right MTL. We argue that both results are not mutually exclusive: they indicate that both the left and right MTL were consistent sources of theta oscillations in the tested group. However, the effect strength in the right MTL was higher in the average power ratios of theta. Such insight could not be gained with either approach taken separately and required the direct comparison of the iES and MNE statistical maps.
Fig 7 shows simulation results to illustrate and underline further the difference in sensitivity between the iES and standard approach. For each simulation run (300 iterations) we generated five minutes of data. Source time-series with a 1/f spectral profile were generated for 68 source locations distributed evenly across the brain according to the Desikan-Killiany atlas from Freesurfer [22]. For two of these locations (precentral left and right), we selectively amplified power in the frequency of interest (8-13 Hz) to obtain a specified ratio fnarrow between narrow-band and broadband power (1-100 Hz). Whereas an fnarrow of 0.6 was targeted for source 1, the targeted fnarrow for source 2 was varied between 0.2 and 0.6. After generating MEG data from this simulation setup, we applied iES (with a f n a r r o w * threshold of 0.22) and the standard approach to detect narrow-band oscillations in the frequency band of interest and computed a metric that quantified the probability of detecting both sources of narrow-band oscillations. Fig 7d shows that the two methods differ systematically: the sensitivity of the standard approach scales with the differences in fnarrow between the two sources, whereas iES’ sensitivity is not influenced by uneven source activity and detects sources above the chosen threshold with a constant probability. This encourages using the different sensitivity profiles of the two methods in conjunction, to obtain complementary information as shown in the data example above. We show the results of a related simulation in S4 Fig where source 2 was a deep source (left parahippocampal) and source 1 a cortical source (left precentral). In this challenging scenario iES outperforms MNE in detecting both sources.
Because of spatial smearing, the study of functional connectivity is a challenging problem for MEG and EEG source imaging (see Fig 1). Since the seed region is maximally correlated with itself and neighbouring regions, with correlated time series due to field spread of the MEG/EEG inverse operator, functional connectivity maps tend to be biased towards artificially inflated values of connectivity measures. This issue is discussed in [23] and generally addressed with methods that discard all contributions of zero phase-lag time series, either by orthogonalizing signals [23] or via measures of the imaginary part of coherence [24]. However, zero-lag coherence between distant regions is plausible theoretically [25] and was observed physiologically [26]. We demonstrate the relevance of iES to address this issue, by studying amplitude correlations in the alpha band (8-13 Hz) with respect to an anatomical seed placed in the sensorimotor cortex. The tested hypothesis was to reveal amplitude correlations with homologous contralateral brain regions [23, 27, 28].
Fig 8 shows results from resting-state data obtained during the same recording session as the visual stimulus experiment. a) shows example time series of co-occurring oscillatory bursts, which form the basis of amplitude correlations between two distant brain regions. The time series were extracted from bilateral central regions. We show occurrences of 270°/90° phase differences during alpha bursts—note that the phase estimation of MEG source signals has a 180° ambiguity due to arbitrary conventions on source direction [29]—and of 180°/0° phase differences, which would be discarded by other methods [23, 24]. We argue that the zero-lag correlations shown here are not spurious, as evidenced by their differences in waveform and amplitude dynamics. This data example provides a proof of principle that studying zero-lag connectivity using MEG is achievable. We next proceeded to map significant inter-regional amplitude correlations in the presence of field spread.
We extracted the source time series yref from the left central sulcus location that was the closest to the activation peak (MNI coordinates [-39, -27, 55] mm) corresponding to the search term ‘finger’ in the Neurosynth meta-analysis tool [30]. We defined the iES quality function fampcorr(s, yref) as the correlation between the amplitude of other source time series s and yref. The outcome to this optimization process is the third use case of iES and is illustrated in Fig 2a. The optimization is done using a solution described in [31] and involved splitting the data in epochs of one second length. We set a correlation threshold of r > 0.4 for spatial components to be included in the signal subspace for the subcorr analysis.
Fig 8b and 8c shows—in a single subject example—that iES was able to reveal the contralateral anatomically-homologous region as the primary distant connected region with the reference brain location. The conventional minimum-norm based map of correlation values was dominated by spurious crosstalk correlation surrounding the seed region. The performance of iES is explained by the equalized contribution of spatial components that are consistent with the embedded hypothesis (r > 0.4). To further limit the contribution of the seed region to the data, it is possible to project the signal subspace and forward fields away from the spatial forward field of the seed region, as illustrated in Fig 8c. The group analysis further reveals that connectivity maps were dominated by crosstalk effects from the seed reference signal, both in the minimum-norm based maps and in the raw subcorr map (Fig 8d and 8e). Projecting away the seed’s contribution before computing every subject’s maps was necessary to confirm the hypothesized contralateral coupling. Note that with iES and in contrast with other approaches [23], the temporal dynamics of the seed region are not projected away from the data; only the spatial topography of the seed region is subtracted from the sensor data. Thus iES does not exclude the detection of physiological zero-lag coupling.
We provide a simulated example in S2 and S3 Figs that illustrates this point. If there is a source that is correlated in amplitude with the seed and the underlying oscillations have a zero phase difference, the two sources will be captured by only one subspace pattern. In this case the coupled region will be picked up by the iES procedure only after the signal subspace and forward fields are projected away from the contribution of the seed topography.
As summarized in Fig 2a, iES can be used for a greater variety of experimental designs: whenever a reference signal yref defined 1) on a trial-by-trial basis or 2) as a continuous signal is considered, fampcorr is used to obtain subcorr maps of sources, whose source dynamics correlate with yref. We illustrate such case in Fig 9, using simultaneous MEG and pupil diameter recordings.
We first formed the iES hypothesis based on recent demonstrations in mice [32] that continuous pupil diameter fluctuations correlated with alpha power at rest. Measures of pupil diameter were extracted from continuous video eye-tracking recordings by fitting an ellipse to the pupil on a frame-by-frame basis. We demonstrate the case of trial-by-trial correlations by analyzing pupil diameter changes prior to visual stimulus presentation from data presented in Fig 3. The signal subspace was defined with spatial components who were deemed significant below a p-value of 0.05 computed by a shuffling procedure across trials. The iES maps indicated brain regions in the occipital cortex, consistent with the upcoming onset of a visual stimulus.
We emphasize that a specific strength of the iES approach is its versatility: it can be extended to a great variety of experimental designs and research hypotheses, since the experimental question is formulated as an optimization problem. We derive in Methods the mathematical formulation for iES coherence with a reference signal, as an additional experimental use case. The experimental hypotheses discussed here all have corresponding quality functions that can be solved analytically. An identical framework can be used for hypotheses that require numerical optimization of the corresponding spatial filters. We foresee that the introduction of the iES approach will establish a generic framework for an increasing number of experimental contexts related to a growing diversity of research questions.
We developed the imaging with embedded statistics (iES) method to produce image maps of the sources of neural oscillations from MEG/EEG sensor data. In this article, we showed with ground-truth simulations and experimental data, that iES identifies source patterns that are challenging to standard approaches, especially when masked by field-spread from other sources in the volume. Specifically: 1) iES generates separate maps for event-related increases and decreases in oscillatory power, which facilitates statistical inference; 2) iES can be used in pair with standard approaches such as wMNE, with iES identifying source regions and wMNE extracting their respective amplitudes. We also showed that 3) detection of functionally connected sources presents an extreme case of source strength imbalance that can be solved using iES, without rejecting the possibility of zero-phase delay coupling between regions. And finally: 4) iES can be flexibly extended to a greater variety of experimental designs, by defining quality functions suitable with the neuroscience hypothesis to be tested with the collected data.
The iES method builds on previous methodological work. In particular, we acknowledge inspiration from previous subspace scanning approaches adapted to MEG/EEG, such as the multiple signal classification (MUSIC) [6] and RAP-MUSIC [8] methods. These previous approaches produced subspace correlation maps, which were subsequently pruned to sets of discrete equivalent dipoles. These methods have inspired an abundant literature, with multiple extensions covering different experimental questions [16, 31] and the estimation of time series interdependencies in functional connectivity [33, 34, 35, 36]. We wish to emphasize however that MUSIC-type methods do not provide a clear path to determine the dimensionality of the signal subspace for each subject. With iES, we propose novel solutions using nonparametric statistics for each use case described. Previous suggestions for component selection [37, 38] are applicable mostly for the original MUSIC solution, which performed principal component analysis of the event-related fields. It is also unclear how previous methods could indicate how many elementary dipole sources could be adjusted to the data, e.g., how many iterations of RAP-MUSIC were required. With iES, subspace correlation maps are used to perform statistical inference at the group level, thus circumventing the necessity of registering heterogeneous discrete elementary dipole models across participants. The goal of iES thus is to produce a distributed and statistically thresholded map. Depending on the data, experimental context and hypothesis embedded in iES, such maps can reveal one spatial peak—e.g., in the visual cortex in the example illustrated in Fig 3 or more source regions, as for example in Fig 6, which highlights sources of beta-band oscillatory activity over bilateral sensorimotor regions. The iES methodology also explicitly addresses the recurrent issue of heterogeneity in the expression of an effect of interest across individuals in a tested group of participants. We propose to use an innovative approach in the field, based on prevalence statistics. We acknowledge that prevalence testing was used in multivariate decoding analyses [10], but the strategy used in iES is quite distinct.
In sum, iES has broad, practical value to research using M/EEG imaging, where group-level inferences are very common. In addition, iES features single-subject analysis and group-based evaluation through the concept of effect prevalence, which can lead to new insights on subject stratification. We consider this feature has great potential value in identifying subgroups of participants e.g., responding differently to a specific paradigm manipulation in experimental psychology and cognitive science, or a therapeutic intervention in clinical trials. A further interesting property of iES is that we separate two steps of the analysis, namely 1) evaluating the presence of a hypothesized effect in single subjects (as well as its prevalence across the group), e.g. increased oscillatory power in a frequency band of interest during stimulus presentation; and 2) the spatial consistency of this effect across the group/subgroup. Standard MNE imaging would not be able to detect an effect that is present in all the subjects of a given study, but is not spatially consistent across the group. In contrast, with iES we would detect that the effect is present in all subjects, even if the group spatial maps do not show peaks that exceed the statistical threshold. Again, we see this as a strong asset of iES, since it could prompt a more in-depth analysis of spatial patterns in individuals. This, in turn, could lead to the discovery of subgroups that differ in the spatial patterns they produce in a given experimental paradigm.
Limitations of iES in its present form are essentially in the definition of hyper-parameters in an ad-hoc manner. This is not specific of iES, as even the standard approaches do not provide, in practice, hyper-parameter estimation from the data or based on strong theoretical background. For instance, the estimated covariance matrices in the first step of iES are regularized using a fixed regularization parameter (see Fig 10 for the two variants used in this paper). Recently, automatic methods [39] for selecting those regularization parameters have been proposed, via maximization of the likelihood of unseen data (under the cross-validation principle). We acknowledge that these methods can be easily used for hyper-parameter selection within the iES framework presented. Another iES hyper-parameter is the p-value threshold used when selecting the spatial patterns to define the signal subspace. We used one of the common thresholds (e.g., 0.001 in the bootstrap procedure for the results in Fig 6), and this certainly lacks theoretical foundations. To improve the rationale on this selection, a more established theoretical framework on the physiological mechanisms explaining the relative power ratios observed between the typical frequency bands of electrophysiology would be required. Some recent work is going in that direction and could inform the selection of iES hyper-parameters, for example by modelling power spectra as a mixture of 1/f spectral noise, as well as narrowband oscillations [see e.g., the discussion in 40, 41]. This open scientific question is beyond the scope of the methodological advances we propose, and models embedded in iES can be improved as advances are made.
There is a fast growing literature on supervised learning methods for M/EEG that test specific questions regarding encoding or decoding of e.g., stimulus features [42, 43]. For the most part, these latter are framed as regression or classification problems and, as purely data-driven methods, they don’t have a model of how a specific spatial pattern was obtained. Our contribution with iES is strongly related to these aspects: iES reframes an experimental question into an optimization problem, which can be seen as a supervised-learning objective to identify spatial patterns of interest from the data. We also propose with iES to interpret these patterns using the physical forward model of MEG/EEG, which yields group-level maps in source space. In principle, iES can be used in studies using supervised learning to find physiological sources of a discriminative pattern, provided the learning procedure is aimed at optimizing linear channel weights. This would not be possible if non-linear methods are used, e.g. with support vector machines with radial basis functions. Finally, we described iES with use cases that have quality functions yielding closed-form solutions. Future work should reveal how iES can be used with non-convex quality functions (i.e. presenting multiple local minima) solved with the typical apparatus of numerical optimization.
All MEG data analysis steps were performed with Brainstorm [15], with the novel approaches described in this paper implemented as a Brainstorm plug-in written in MATLAB (available through: https://github.com/pwdonh/ies_toolbox).
The iES method described in this paper is based on subspace scanning, which processes the entire spatio-temporal MEG data matrix X, instead of reconstructing neural activity independently at each time point for the whole source space [see e.g. MUSIC, 6]. The method features two steps, as shown in Fig 2: 1) extraction of the relevant spatial patterns from the data (signal subspace identification), and 2) scanning of the source space for contributions that explain the identified spatial patterns (subspace scanning step per se).
|
10.1371/journal.pcbi.1000473 | Microarray Comparative Genomic Hybridisation Analysis Incorporating Genomic Organisation, and Application to Enterobacterial Plant Pathogens | Microarray comparative genomic hybridisation (aCGH) provides an estimate of the relative abundance of genomic DNA (gDNA) taken from comparator and reference organisms by hybridisation to a microarray containing probes that represent sequences from the reference organism. The experimental method is used in a number of biological applications, including the detection of human chromosomal aberrations, and in comparative genomic analysis of bacterial strains, but optimisation of the analysis is desirable in each problem domain.
We present a method for analysis of bacterial aCGH data that encodes spatial information from the reference genome in a hidden Markov model. This technique is the first such method to be validated in comparisons of sequenced bacteria that diverge at the strain and at the genus level: Pectobacterium atrosepticum SCRI1043 (Pba1043) and Dickeya dadantii 3937 (Dda3937); and Lactococcus lactis subsp. lactis IL1403 and L. lactis subsp. cremoris MG1363. In all cases our method is found to outperform common and widely used aCGH analysis methods that do not incorporate spatial information. This analysis is applied to comparisons between commercially important plant pathogenic soft-rotting enterobacteria (SRE) Pba1043, P. atrosepticum SCRI1039, P. carotovorum 193, and Dda3937.
Our analysis indicates that it should not be assumed that hybridisation strength is a reliable proxy for sequence identity in aCGH experiments, and robustly extends the applicability of aCGH to bacterial comparisons at the genus level. Our results in the SRE further provide evidence for a dynamic, plastic ‘accessory’ genome, revealing major genomic islands encoding gene products that provide insight into, and may play a direct role in determining, variation amongst the SRE in terms of their environmental survival, host range and aetiology, such as phytotoxin synthesis, multidrug resistance, and nitrogen fixation.
| We describe the first use of a method for the analysis of bacterial microarray comparative genomic hybridisation (aCGH) that includes information about the spatial organisation of genes in the reference bacterium. We demonstrate that using this information improves predictive performance over standard bacterial aCGH methods in discriminating between genes from the reference organism that either do or do not have putative orthologues in the comparator organism. Our approach produces good results on more distantly related bacteria than can successfully be analysed by the standard methods. We apply our analysis to comparisons between four commercially-significant plant pathogenic bacteria, and identify several regions of the genome that are likely to contribute to their ability to cause disease, and to proliferate in the environment, generating hypotheses for future experiments.
| Microarray comparative genomic hybridisation (aCGH) provides an estimate of the relative abundance of genomic DNA (gDNA) taken from comparator and reference organisms by hybridisation to a microarray containing probes that represent sequences from the reference organism. This method has been used in a number of biological applications, including the detection of human chromosomal aberrations [1],[2]; comparisons of bacterial human pathogens [3]–[10]; bacterial plant pathogens [11],[12]; industrially-important bacteria [13]; and comparative transcriptomics of Xenopus laevis [14].
Numerous algorithms and software packages have been applied to the analysis of this aCGH data in prokaryotes. The majority of these partition reference organism sequences into two mutually exclusive classes: sequences that are ‘present’ and sequences that are ‘absent or divergent’ in the comparator organism [e.g. 5],[12],[15],[16]. Observed hybridisation data are, in each case, assumed to be reliable proxies for these classes.
In this manuscript we describe and apply an improved method for analysis of aCGH data from bacterial genome comparisons. This method incorporates spatial information about CDS location on the reference genome in a hidden Markov model (HMM). This spatial information is expected to capture pertinent biological and evolutionary information, such as operon structure, and regions of lateral gene transfer. Our approach differs from previously proposed, and widely-used, methods applied to bacterial aCGH, such as GACK and MPP, that consider hybridisation intensities of each reference probe as measurements that are independent of their genomic location [15],[16], and is thus more similar to methods such as ArrayLeaRNA [17], which incorporates predicted operon structure into interpretations of microarray expression data, for a restricted set of organisms. We compare the relative performance of our method to commonly used bacterial aCGH analysis algorithms and software.
We demonstrate that several assumptions of common bacterial aCGH analysis methods concerning the relationship between observed hybridisation scores and ratios and the presence or absence of a reference CDS in the comparator organism do not always hold strongly, and that this is particularly the case for more distantly-related organisms. Our data in particular do not support a distinction between ‘present’ and ‘absent or divergent’ classes of sequence, but rather between those sequences in the reference organism that do, and those that do not, have putative orthologues in the comparator genome. We find that the HMM is a better predictor of reference sequences that do not have a putative orthologue in the comparator organism than the other methods tested.
Spatial organisation of sequences on the reference genome has previously been incorporated into methods applied to aCGH analyses of copy number variation in human genomes. This has been represented using HMM [18] and segmentation methods [19]. Simple smoothing methods have also been used to identify breakpoints in this data [20]. However, the problem domain of human copy number aCGH (detecting copy number variation in a known genome sequence) differs from the problem domain of bacterial comparative genomic aCGH (identifying the presence or absence of putative orthologues of known genes in a genome of unknown sequence). To the best of our knowledge, this study describes the first application of a method incorporating such spatial information to aCGH for comparative genomics of unsequenced bacteria, and the first demonstration of the applicability of the technique as a whole across bacterial genera.
Pectobacterium atrosepticum (Pba), Pectobacterium carotovorum (Pcc), and Dickeya spp. are plant pathogenic soft-rotting enterobacteria (SRE) that share a common ancestor. Despite their many similarities, these commercially significant pathogens differ in their host range, geographical distribution, aetiology and environmental persistence [21]. The molecular origins of these differences are not well understood, but this ecological flexibility is likely indicative of a dynamic, plastic genome with ‘core’ and ‘accessory’ components. There are currently two publicly available annotated genomes for these organisms: Pba strain SCRI1043 (Pba1043) [22], and Dickeya dadantii strain 3937 (Dda3937; https://asap.ahabs.wisc.edu/; v6b). The availability of these sequences has rapidly advanced our understanding of these organisms, but broader comparisons are expected to deliver greater insight into the evolution and function of the SRE.
The major common virulence factors of the SRE are plant cell wall degrading enzymes (PCWDE) that degrade the plant cell wall to release nutrients in a so-called ‘brute force’ attack [23]–[25]. Other virulence factors include virulence-associated secretion systems, siderophores, cell-surface polysaccharides and agglutinins [26],[27].
By contrast, bacterial plant pathogens such as Pseudomonas spp. are associated with a biotrophic ‘stealth’ interaction with the host. These ‘stealth’ pathogens employ mechanisms such as the type III secretion system (T3SS) to translocate effectors into host cells. The effectors modulate the host plant's biochemical responses, implementing a wide array of strategies to circumvent host immunity [28]–[31]. However, Pba1043, Dda3937, and other SREs also encode a functioning T3SS and other gene products associated with this ‘stealth’ interaction, indicating a more complex relationship with their hosts than simple ‘brute-force’ necrotrophy [22], [32]–[36]. Key factors with a confirmed role in virulence include type IV and type VI secretion systems, and the phytotoxin coronafacic acid (CFA), which is synthesised by the cfa gene cluster [22],[37]. Other factors associated with persistence in, and adaptation to, the wider environment have been identified, such as genes associated with opine uptake, biofilm formation, antibiotic production, and nitrogen fixation [22].
In many bacteria, such genes associated with pathogenicity, and other phenotypically-distinguishing characters, are frequently associated with islands of horizontal gene transfer (HGT). This gene complement is often variable between strains and species, and is sometimes termed the ‘accessory genome’, in order to distinguish it from the ‘core genome’ that provides functionality presumed to be essential to all related organisms [5], [9], [22], [38]–[40]. We expect that observed differences between the gene complements of SRE will reflect differences in their phenotypes, and adaptations to their distinct environments, and that these differences will be preferentially located in islands of genes in their genomes. We use aCGH and apply our analysis method to identify genomic islands in Pba1043 that do not have putative orthologues in the unsequenced Pba strain SCRI1039 (Pba1039) and Pcc strain SCRI193 (Pcc193), and in the sequenced Dda3937. In this study, coding sequences (CDS) from Pba1043 that are predicted by aCGH to be absent or divergent in Pba1039, Pcc193 or Dda3937 are of interest because they may potentially contribute to Pba1043-specific phenotypes, including host interactions. Pairwise comparisons between Pba1043 and these three organisms span a range of evolutionary distances since their most recent common ancestor with Pba1043, and represent variation at strain, species and genus levels.
Our results for the SRE support a hypothesis that the genomes of SRE continue to be modified by the acquisition of genomic islands, and the model of an ‘accessory genome’ of niche-specific functionality that is composed, at least in part, of horizontally-acquired genomic islands. We identify major differences in the CDS carried within the accessory genomes of SRE and, while these recapitulate previous observations of major genomic islands made using alternative approaches [22],[38], we also find a number of unexpected differences that provide insight into, and may play a direct role in determining, variation amongst the SRE in terms of their environmental survival, host range and aetiology.
Annotated genome sequences were obtained from GenBank for Pba1043 (accession: NC_004547), Lactococcus lactis subsp. lactis Il1403 (accession: NC_002662), and L. lactis subsp. cremoris MG1363 (accession: NC_009004). Equivalent data for Dda3937 was obtained from ASAP (https://asap.ahabs.wisc.edu/; v6b). CDS annotations from these sources were not modified for this study.
Putative orthologues of bacterial CDS were identified using reciprocal best hit (RBH) analyses. RBH were identified by using each annotated CDS from the reference genome as the query in a sequence search against the comparator genome, and vice versa. A RBH was called when the best match to a query sequence had the query sequence as its own best match in the reciprocal comparison [see also 22],[38]. For protein comparisons, FASTA 3.4t25 was used, and BLASTN 2.2.11 was used for nucleotide comparisons. Reciprocal best hits were interpreted as putative orthologues, and converted to Boolean ‘present’ and ‘absent’ states for model development and training. We would usually employ a threshold for RBH of a minimum of 30% identity over a minimum of 80% of the sequence length for protein comparisons. However, for this analysis we relaxed both criteria completely, and considered the best hit in each direction without such a filter. The division of CDS into ‘present’/‘absent’ classes on the basis of RBH without these thresholds corresponds to a strict classifier for allocating CDS to the ‘absent’ class.
Under the usual circumstances in which we perform these comparative analyses, we wish to exclude weak reciprocal matches from the ‘present’ set in order to avoid inappropriate attribute transfer or assignment. In those cases, we would implement this filter to minimise misallocation of CDS to the ‘present/putative orthologue’ class.
However, in the case of this aCGH analysis, as we note that probes to reference organism sequences that have little or no sequence identity to the comparator may still give very high hybridisation strengths/ratios, we wish preferentially to avoid misallocation of CDS to the ‘absent’ class. Therefore, we aim in effect to give each reference sequence every possible opportunity to be classified as ‘present’ as a putative orthologue in the comparator on the basis of RBH. Any remaining reference CDS that are classified as ‘absent’ - even though no restrictions are made on the basis of sequence identity or match overlap – have no reciprocal similarity by BLAST to any sequence in the comparator.
Genomic DNA was extracted from bacterial cell cultures (∼1010 cells) using the QIAGEN Genomic-tip 100/G (Qiagen) as recommended and labelling was performed using modified Bioprime DNA Labelling System (Invitrogen). Briefly, 2 µg gDNA in 21 µl was added to 20 µl random primer reaction buffer mix and denatured at boiling for 5 min prior to cooling on ice. To this, 5 µl modified 10× dNTP mix (1.2 mM each of dATP, dGTP, dTTP; 0.6 mM dCTP; 10 mM Tris pH 8.0; 1 mM EDTA), 3 µl of either Cy3 or Cy5 dCTP (1 mM) and 1 µl Klenow enzyme was added and incubated for 16 h at 37°C. Labelled samples for each array were combined (if applicable) and unincorporated dyes removed using Qiaquick PCR Purification Kit (Qiagen) as recommended, eluting twice with 1×50 µl sterile water. Hybridisations and washing were performed as recommended (Agilent Protocol v5.5). Genomic DNA from Dda3937 was hybridised to a Pba1043-specific microarray (ArrayExpress: E-TABM-600; manufactured by Agilent, AMADID 012663) carrying 5219 unique probes that represent 4450/4472 annotated CDS from the Pba1043 genome [37],[41]. Hybridisations were carried out in the presence of Pba1043 reference gDNA, Pcc193 reference gDNA and in the absence of a reference sample, and all hybridisations were replicated three times. Scanning was performed with an Agilent G2505B scanner using default settings and data extracted using Agilent FE (AGFE) software v9.5.3.
Raw hybridisation data was imported using MatLab (http://www.mathworks.com) from AGFE format output (Pba1043 array), and from GEO (Lactococcus comparison data, entries: GSM229601, GSM229602, GSM229603, GSM229604) [13]. GEO entries 229602 and 229604 were found to have the labels for channel 1 and 2 inverted, and this was corrected in processing. Raw hybridisation data was corrected for background signal, log-transformed in base 2, then quantile-normalised. Median values were calculated for replicate probes on each array, and then between replicate arrays. Normalised hybridisation scores were associated with a RBH result for each CDS.
Two-dimensional Gaussian mixture models were fitted in MatLab to the paired hybridisation and RBH data using the gmdistribution.fit function. The optimal number of fitted Gaussians was estimated by the Bayesian Information Criterion (BIC), considering a maximum of ten Gaussians.
Threshold models were implemented such that each CDS with a normalised array hybridisation score (or ratio) that fell below the threshold was classified as ‘absent’; those with a normalised hybridisation score above that value were classified as ‘present’. These Boolean states were used for validation of threshold models, and for training of HMMs. Threshold scores were taken at 100 evenly-spaced values between the lowest and highest observed values of hybridisation score (or ratio) for data exploration, and at all observed normalised threshold values (exhaustively to explore all partitions of the data) for rigorous comparisons with alternative models.
First-order hidden Markov models were trained using MatLab's hmmestimate function, given the Boolean ‘present’/‘absent’ states derived from reciprocal best hit analysis ordered naturally along the reference genome as a ground truth, and Boolean ‘present’/‘absent’ states derived from the threshold models as observed emission states. The derived models represent the presence or absence of a putative orthologue in the comparator sequence as hidden states, in conjunction with the observed hybridisation score being above or below the corresponding normalised hybridisation score threshold, as the emitted states. The resulting models were used to obtain predicted hidden states from hybridisation data using the Viterbi algorithm implemented in MatLab's hmmviterbi function, where input data were again ordered naturally according to probe location on the reference genome.
HMMs used in this study were trained separately on the RBH and hybridisation data for two comparisons: Pba1043 and Dda3937; and Lactococcus lactis subspecies lactis IL1403 and cremoris MG1363 [13].
The packages GACK and MPP were obtained from their homepages (http://falkow.stanford.edu/whatwedo/software/software.html; http://cbr.jic.ac.uk/dicks/software/mpp/index.html), and used as recommended in their documentation [15],[16]. Array hybridisation data was converted to the appropriate input format in each case using Python scripts. GACK binary and trinary data used in this study was obtained at %EPP cutoffs of 0%, 50%, and 100%, with all other settings at default values. MPP data used in this study was obtained with default settings.
All predictive models were validated against the ground truth of reciprocal best hit results for Pba1043 vs Dda3937 or the two Lactococcus strains, as appropriate. All model output was obtained as Boolean ‘present’/‘absent’ states, and validation statistics were obtained for consistency tests using MatLab's classperf function (http://www.mathworks.com/access/helpdesk/help/toolbox/bioinfo/ref/classperf.html). Predictions of the absence of a reference CDS in the comparator organism were taken to be ‘positive’ for statistical classification purposes. The optimal HMM and threshold models identified by the validation process were used for subsequent predictions on Pba1039 and Pcc193 hybridisation data.
The software package alien_hunter was downloaded from http://www.sanger.ac.uk/Software/analysis/alien_hunter/ and used to identify regions of divergent genome composition, with recommended settings. This application implements an interpolated variable order motif method derived from the base composition of the chromosome to detect regions of nucleotide bias, and a second-order HMM for change-point detection [42].
Empirical statistical testing of the association of predicted genomic islands with regions of divergent genome composition as predicted by alien_hunter, and with regions of manually-annotated HGT, was carried out using the following procedure, implemented in a Python script.
The locations of genomic islands predicted by HMM, by alien_hunter, and detailed in the NC_004547 annotation were obtained. These were each considered to represent independent, non-overlapping genomic regions. The location of each of the alien_hunter and NC_004547 regions was shuffled one thousand times, to produce two sets of non-overlapping arrangements of each, representing a random distribution of the predicted islands. A count of the number of HMM-predicted genomic islands that overlapped with each of the shuffled sets was taken, as a measure of the expected number of overlaps that would be obtained if the islands were randomly placed on the genome. The observed count overlap count of the HMM predictions with the alien_hunter and annotated islands was tested for significance using a Z-statistic.
A similar procedure was followed for determining whether individual genes were located preferentially within predicted islands. In this case, the gene locations were taken as static, and genomic island predictions shuffled as non-overlapping regions 1000 times. A Z-statistic was again used to calculate significance of the count of genes observed to be coincident with predicted genomic islands.
The genomes of Pba1043 and Dda3937 have been sequenced and annotated [22] (https://asap.ahabs.wisc.edu/; v6b). CDS were defined to be common to both bacteria if a putative orthologue to a Pba1043 CDS could be found in the Dda3937 annotation. This was determined for each CDS at the amino acid level by reciprocal best FASTA protein match, and at the nucleotide level by reciprocal best BLASTN match [22]. Each reciprocal best hit (RBH) result was considered to be a putative orthologue (hereafter used interchangeably with ‘orthologue’) and, as a direct and exhaustive sequence comparison, to be the best estimate of the presence or absence of Pba1043 CDS in Dda3937 available for method validation. The results were used as both reference and training data for aCGH analysis algorithms, in a consistency test approach similar to that used in [4].
Of 4450 Pba1043 CDS represented by probes on the microarray, 451 were found to have RBH to Dda3937 at both nucleotide and amino acid sequence levels. In addition, 2369/4450 Pba1043 CDS made RBH at the amino acid level only, and 7/4450 CDS only at the nucleotide level. For Pba1043 1623/4450 CDS therefore have no putative orthologue in Dda3937, and it may be considered that approximately one third of the Pba1043 genome is not common with Dda3937 (Figure S1). Very few Pba1043 CDS were found to be orthologous at the nucleotide, but not the protein level (a pattern suggestive of positive selection); however, many were orthologous at the protein, but not at the nucleotide, level (suggestive of neutral drift). The ‘core’ of CDS with both protein and nucleotide-level orthologues was found to comprise only around 10% of the Pba1043 genome.
Genomic DNA from Dda3937 was hybridised to a Pba1043-specific microarray in the presence, independently, of Pba1043 reference gDNA and Pcc193 reference gDNA, and also in the absence of a reference sample. Three overlapping populations of raw hybridisation strengths were observed in each experiment (Figure 1). This pattern was similar to that observed in similar experiments [12], and comprised: a strongly-binding population of Pba1043 probes that bind to Dda3937 gDNA with hybridisation strength comparable to their binding to Pba1043 gDNA; a weakly-binding population of probes with lower hybridisation strength to Dda3937 than to Pba1043 gDNA; and a population with either no detectable, or very weak, hybridisation to Dda3937 gDNA (Figure 1).
This observation does not support the assumption commonly made in aCGH analysis methods that there are two populations of probes in a typical experiment: ‘present’ and ‘absent or divergent’ [e.g. 3],[8],[11],[15],[16]. Notably, there is no a priori indication that any of the three observed populations in Figure 1 comprise ‘present’, ‘absent’ or ‘divergent’ sequences.
A linear, or at least monotonic, relationship between the observed hybridisation score and the sequence identity of CDS in the reference and comparator organisms has previously been proposed or observed for aCGH experiments [13],[43],[44]. We did not observe such a relationship. Our data indicated a complex relationship between sequence identity and probe hybridisation affinity (or log ratio), from which three major populations of probes could readily be distinguished (Figure 2). Those probes representing Pba1043 sequences that made RBH with greater than 30% amino acid sequence identity in Dda3937 were considered here to be putative orthologues and therefore may only be classed as either ‘present’ or ‘divergent’, according to the scheme commonly used in aCGH analyses [e.g. 5],[12],[15],[16]. The ‘absent’ sequence set in that scheme corresponds to Pba1043 CDS with no putative orthologue in the comparator organism.
In the Pba1043:Dda3937 comparison, probes matching orthologous sequences could be resolved into two distinct populations on the basis of hybridisation strength using Gaussian mixture models, but not on the basis of their sequence identities (Figure 2A). In particular, the distribution of putative orthologues was bimodal with respect to hybridisation score or ratio, but was unimodal with respect to sequence identity. Sequence divergence was measured in terms of sequence identity, and it was not possible to distinguish between ‘present’ and ‘divergent’ orthologues using hybridisation data. The commonly-used ‘absent or divergent’ classification is the union of the sets of ‘absent’ and ‘divergent’ sequences; our data does not support this distinction between ‘present’ and ‘absent or divergent’ probe sets.
Those probes corresponding to Pba1043 CDS that were not found to have putative orthologues in Dda3937 (i.e. that are ‘absent’) were observed to have hybridisation ratios that ranged from no measurable hybridisation to very strong hybridisation, and to take values on the full range of hybridisation ratios spanned by both ‘present’ and ‘divergent’ CDS. The distribution of hybridisation ratios for probes representing putative orthologues overlapped to a great extent that of probes corresponding to CDS with no orthologue (Figure 2B).
Similar results were obtained for nucleotide sequence comparisons, and for raw hybridisation scores (Figure S2). As can be seen from Figure S2, the observed relationship between sequence identity and hybridisation affinity is qualitatively almost identical whether obtained using hybridisation intensity (univariate) data, or hybridisation ratio (bivariate) data. The complexity of this relationship is therefore not due to the use of a log-ratio summary of the hybridisation signal.
Taken together, these results indicated that a distinction might reasonably be drawn between ‘putatively orthologous’ and ‘putatively non-orthologous’ CDS on the basis of aCGH, but not between ‘present’ and ‘absent or divergent’ CDS.
Analytical models for aCGH based on a single threshold that partitions CDS into ‘present’ and ‘absent or divergent’ classes have previously been shown to perform acceptably well under some circumstances [e.g. 5], [44]–[46]. However, while the data obtained in this study did not support that particular interpretation of the partitioning of sequences, a threshold approach may still distinguish successfully between reference CDS that do and do not have a putative orthologue in the comparator organism.
The influence of horizontal gene transfer has been in many cases to introduce islands of genes whose collective function distinguishes the recipient organism from its close relatives, as part of the ‘accessory’ genome [5],[9],[22],[38],[39],[47]. One notable influence of HGT on the reference genome is to confer collocation of transferred genes in that genome; such transferred genes may additionally be expected not to have an orthologue in a given comparator genome. In particular, it would be expected that, where a reference genome CDS has been acquired by HGT of a genomic island, it and its neighbours are less likely to have an orthologue in a comparator genome than another CDS randomly selected from the reference genome. Similarly, prokaryotic genes are frequently collected into operons, collocated groups of sequences that often work towards a common function. Loss of function may thus entail loss of a collocated set of genes. We implemented a HMM that exploits this anticipated collocation of sequences on the reference genome, particularly if they have no orthologue in the comparator, in the expectation that taking into account this spatial bias would improve predictive performance in the presence of data noise, and in marginal cases that are difficult to resolve with only a single threshold-based predictor. Such cases might include genes with an unexpected level of redundancy in the comparator organism, such as those with variable copy number due to representation on plasmids [5].
Threshold and HMM models (as defined in Materials and Methods) were constructed for all hybridisation scores and ratios observed in each array experiment, exhaustively enumerating all such models that could be constructed from the data. All possible outcomes of each method were thus obtained, facilitating general claims concerning their performance on this data. In each experiment, a threshold model could be obtained that performed acceptably well when distinguishing between Pba1043 CDS that do and do not have putative orthologues in Dda3937. However, the optimally performing HMM outperformed the optimally performing threshold model on measures of correct prediction rate and specificity, in consistency tests for all such experiments (Table 1).
It was observed that HMMs and threshold models constructed from experiments involving reference gDNA performed significantly better than those constructed from experiments where no reference gDNA was used. Also, models built using log hybridisation ratios performed better than those derived from single-channel raw hybridisation scores (Table 1). Using log-transformed ratio data, the threshold and HMM predictors predicted that similar total numbers of CDS from Pba1043 did not have a putative orthologue in Dda3937 (HMM: 1179; threshold: 1191; in silico analysis: 1630) but differed in their classification of 372 (approximately 30%) of these CDS. The predictions made by the two approaches differ qualitatively, rather than quantitatively (Figure 3). The HMM predictions appear to form larger contiguous islands of CDS on the genome, while the threshold method predicts a greater number of ‘orphan’ CDS with no orthologue whose immediate neighbours are predicted to have orthologues, and splits several large islands (confirmed as single islands by in silico sequence comparison) into several smaller fragments.
Additionally, the behaviour of each model is seen to differ as the hybridisation ratio threshold varies from the minimum to maximum observed value. Both models predict a mixture of CDS with and without orthologues in the comparator at low hybridisation ratios, but at high ratios the threshold model predicts that all Pba1043 CDS are without an orthologue in Dda3937. At high hybridisation ratio thresholds, the HMM assigns the majority state for the data to all CDS (Figure S3). Also, ‘blocks’ of contiguous sequences with no orthologue in the comparator persist to higher hybridisation ratios, using the HMM approach.
Two packages for analysis of bacterial aCGH data are GACK (perhaps the most widely-used such application) and MPP, amongst a wide range of proposed alternative aCGH analysis algorithms [10], [15], [16], [48]–[52]. Nearly all of these methods make the assumption that array probes partition into ‘present’ and ‘absent or divergent’ classes, and that these classes are unimodal. It was seen that this assumption is not met in the Pba1043:Dda3937 comparison but, as for the threshold-based classification, it is likely that these applications are able to segregate CDS from Pba1043 that do have orthologues in Dda3937 from those that do not.
We applied GACK and MPP to the same log hybridisation ratio data for the Pba1043:Dda3937 comparison that was most informative for both the HMM and threshold methods above. In GACK it is possible to modify the required stringency of the prediction by varying a parameter representing “estimated probability of presence” (EPP). This may be set at values ranging from 0% - indicating an expectation of statistical ‘certainty’ that CDS predicted to have no orthologue in the comparator organism truly have no such orthologue - to 100% - indicating an expectation of statistical ‘certainty’ that CDS predicted to have an orthologue in Dda3937 truly do have such an orthologue. GACK was applied with EPP values of 0%, 50% and 100%, in binary prediction mode. With these settings, GACK predicted that 84 (0% EPP), 344 (50% EPP) or 595 (100% EPP) CDS from Pba1043 have no putative orthologue in Dda3937 (Figure S4).
MPP with default settings predicted that no Pba1043 CDS were without a putative orthologue in Dda3937 – the majority state - and thereby achieved a correct prediction rate of 0.65. Although GACK obtained a correct prediction rate of 0.75 at 100% EPP, its sensitivity was very low and, unlike the threshold and HMM methods, neither GACK nor MPP identified a substantial proportion of the 1630 Pba1043 CDS that do not have an orthologue in Dda3937. Validation statistics for these analyses are shown in Table 2, and indicate that the HMM outperformed both GACK and MPP on the Pba1043:Dda3937 comparison in terms of sensitivity and total number of correct predictions, although GACK obtained better positive predictive rates at the expense of much reduced sensitivity.
It is possible that the less impressive performance of GACK and MPP observed for the Pba1043:Dda3937 comparison was due to the relatively large evolutionary distance between these organisms, or to the particular array configuration used in these experiments (see Discussion). Most aCGH studies have hitherto focused on variation at the subspecies level, and this is the domain on which GACK and MPP have previously been and, it was assumed, were intended to be, applied [5],[9],[13],[15],[16]. In order to compare the performance of the HMM to GACK and MPP on a comparison of sequenced bacteria with a more recent common ancestor, data for aCGH between Lactococcus lactis subspecies lactis IL1403 and cremoris MG1363 [13], employing an alternative array platform, was obtained from the GEO public repository. The HMM approach again outperformed both GACK and MPP in terms of sensitivity, correct positive rate, and positive predictive rate on this comparison data (Table 2). Although GACK more closely approximated the number of non-orthologous sequences in its predictions, its false positive rate was found to be rather high.
Pcc193 and Pba1039 gDNA was hybridised to the Pba1043-specific microarray, in separate experiments, using Pba1043 gDNA as the reference in each. The distribution of log hybridisation ratios was found to be approximately unimodal in both cases, reflecting the relatively close evolutionary relationship between these organisms (data not shown).
MPP, with default settings, was unable to fit curves to the hybridisation data from the Pba1043:Pcc193 experiment, and so its performance was not further assessed. The HMM trained on Pba1043:Dda3937 comparison data predicted that 440 Pba1043 CDS have no orthologue in Pcc193. GACK predicted that between 1187 (EPP: 0%) and 1846 (EPP: 100%) Pba1043 CDS have no orthologue in Pcc193. As noted earlier, in silico sequence comparisons indicated that 1643 Pba1043 CDS have no orthologue in Dda3937, whose most recent common ancestor with Pba1043 is more ancient than that of Pba1043 and Pcc193. It would therefore be expected that more Pba1043 CDS would have orthologues in Pcc193, than in Dda3937. This implies that the GACK prediction for the Pcc193 comparison at 100% EPP is an overprediction. There was also a large discrepancy between the prediction count from HMM and the most conservative GACK prediction at 0% EPP, in that GACK predicted nearly three times as many CDS to be without an orthologue in Pcc193 than did the HMM.
The Pba1043:Pba1039 comparison experiment was a comparison between reference and comparator organisms at the strain level. The HMM built from the Pba1043:Dda3937 comparison data may be inappropriate for analysis of more closely-related organisms, and so the second HMM, trained separately on the Lactococcus comparison data, was also tested. MPP predicted that 299 Pba1043 CDS have no orthologue in Pba1039, and GACK predicted between 335 (EPP: 0%) and 637 (EPP: 100%) such CDS. The HMM built on the more divergent Pba1039:Dda3937 comparison predicted 198, and the HMM built on the more recently-diverged Lactococcus comparison predicted 255 such CDS. The variation in prediction totals between the HMMs built on the two distinct comparisons is not as great as the variation between the HMM predictions and those made by GACK and MPP, and the predictions made by the HMMs are each in close agreement, implying that the HMM approach is reasonably robust to training set variation, independent of the organism on which it was trained (Figure S5).
While no genome sequences were publicly available at the time of submission to validate these particular predictions, some trends may be inferred from this data. GACK appeared to predict a greater number of CDS to be absent than did the HMM. This behaviour, which potentially results in an increase in sensitivity at the expense of specificity, has previously been reported by other groups [e.g. 5]. Qualitatively, both GACK and MPP predicted a greater proportion of ‘orphan’ CDS, while the HMM favoured prediction of islands of CDS with no orthologue in the comparator (Figure S5). This may be a more biologically appropriate prediction mode.
We observed apparent overprediction, combined with reduced sensitivity and diminished correct positive prediction rates for the GACK and MPP methods, in comparison to the HMM approach. We also found that variation in results between HMMs built on alternative training sets is minor. Thus we proceeded to consider the biological implications of aCGH results obtained for the Pectobacterium and Dickeya species investigated, using only results obtained using the HMM analysis model built from the Pba1043:Dda3937 comparison.
HMM analysis predicted 165 islands (1179 CDS) from Pba1043 to have no orthologues in Dda3937, 60 islands (440 CDS) to have no orthologues in Pcc193, and 17 islands (198 CDS) to have no orthologues in Pba1039. This method also identified 16 islands (169 CDS) that were unique to Pba1043 only, and a further 40 islands (231 CDS) to be present only in Pba1043 and Pba1039. The count of genomic islands and CDS with no orthologue in the comparator diminished as the evolutionary distance from the last common ancestor of Pba1043 to the comparator decreased. These islands are illustrated in Figure 4 and Figure S6, and described in detail in Tables S1, S2, S3, S4 and S5.
We considered those CDS that are present in Pba1043 but that do not have orthologues in the most recently diverged organism in this study: Pba1039, to reflect either recent acquisitions in Pba1043 or recent losses in Pba1039. These CDS are a putative Pba1043-specific ‘accessory’ genome, and mostly comprise hypothetical proteins and phage-related sequences, located in 17 islands on the Pba1043 genome (Table S1; islands prefixed Pba1039I).
Fifty-six islands of Pba1043 CDS were predicted to be present only in Pba1043, or to be common to both Pba strains, but not to have orthologues in either Dda3937 or Pcc193. These are likely to represent genes encoding functions that biochemically distinguish Pba from its near evolutionary relatives. Such sequences included CDS encoding coronafacic acid synthesis (cfa), phenazine antibiotic synthesis (ehp), and various multidrug resistance genes. Several of these CDS, in particular those for the synthesis of coronafacic acids (CFA) have also previously been shown experimentally to contribute to virulence in Pba1043 [22]. These CDS were predicted to be components of the putative Pba-specific ‘accessory’ genome, and some examples are summarised in Table 3. A substantial minority of these CDS were annotated only as hypothetical proteins in their public sequence database submissions (Table S2).
One-hundred and sixty-eight islands of Pba1043 CDS were predicted to have orthologues in both Pcc193 and Pba1039 but not in Dda3937, and thus represent a putative Pectobacterium-specific accessory genome. These islands are expected to include genes encoding functions that distinguish pectobacteria from Dickeya spp., and were found to contain CDS encoding PCWDE (pel and peh), a syringomycin-like NRPS (syr), siderophore biosynthesis (pvc) and octopine transport (occ).
Seventeen putative horizontally acquired islands (HAI1-HAI17) were identified in manual curation of the Pba1043 genome on the basis of evidence such as divergent base composition and the presence of flanking insertion sequences [22]. Of these, all but HAI1 coincided with at least one island identified by aCGH, and most include genes with putative or demonstrated roles in pathogenesis and niche adaptation [38] (Table 4).
Two of these islands, HAI1 (capsular polysaccharide biosynthesis) and HAI15 (type I secretion) were predicted to be entirely or substantially conserved in all organisms examined in this study. If the shared presence of each of these two islands is the result of horizontal gene transfer, then the most parsimonious inference is that acquisition occurred in a common ancestor of all three species, rather than as independent transfer events in each organism.
Two HAIs were predicted to have substantial orthologues only within the pectobacteria: the portion of HAI2 that is homologous with the SPI-7 pathogenicity island (PAI) flanking the coronafacic acid synthesis genes (the cfa genes themselves have no orthologues in Pcc193), and HAI6, which encodes a syringomycin-like NRPS. A parsimonious explanation for this distribution might be that these islands were acquired after the divergence of Dickeya and Pectobacterium spp. but before the divergence of Pcc and Pba species; alternatively, there may have been loss of these islands in the Dickeya lineage. However, the PAI itself has been observed in several unrelated bacterial genomes, and found to contain multiple alternative functional ‘payloads’ in those cases [53],[54]. As the PAI genes, but not their cfa ‘cargo’ were predicted to be present in Pcc193, it may be that there has been independent acquisition of this sequence in Pba and Pcc, carrying alternative payloads in each case. This may be determined by sequencing of that region in Pcc193.
Similarly, five HAIs (HAI3, HAI5, HAI10, HAI11 and HAI12) were found only in the two Pba strains either substantially, or in their entirety (Table 4; Table S2). These are expected to have been acquired after the divergence of Pba from Pcc. Amongst the gene functions carried by these HAIs are lipopolysaccharide biosynthesis (rfb) and phenazine antibiotic synthesis (ehp).
A further five HAIs (HAI4, HAI7, HAI9, HAI13, and HAI17) appeared to be substantially or entirely unique to Pba1043, but these almost exclusively encode for phage-related sequences, and hypothetical proteins. These were presumably recently acquired, subsequent to the divergence of strain SCRI1043 from strain SCRI1039.
HAI14, which putatively encodes nitrogen fixation function, is anomalous in that it was predicted to have a substantial number of orthologues in both Pba strains, and in Dda3937, but to have far fewer orthologues in Pcc193. The most parsimonious explanation for this distribution is that the common ancestor of Dickeya and Pectobacterium possessed this capability for nitrogen fixation, and that this has been progressively lost in the Pcc193 lineage. Alternatively, nitrogen-fixing ability may have been acquired independently in both Dickeya and Pba lineages.
The software package alien_hunter [42] was used to identify regions of putative HGT in the Pba1043 chromosome. An empirical statistical method was used to determine whether there was a significant association between Pba1043 CDS without predicted orthologues in each comparator species and regions of putative HGT as predicted by alien_hunter.
In total, alien_hunter identified regions of putative HGT overlapping 731 CDS in Pba1043. These included 118/173 CDS that were predicted by aCGH to be specific to Pba1043 (Z-score 5.54; P<0.0001); 254/400 CDS predicted to be specific to Pba strains (Z-score: 9.08, P<0.0001); 256/440 Pba1043 CDS predicted to have no orthologue in Pcc193 (Z-score: 8.47, P<0.0001); and 463/1179 CDS predicted to be have no orthologue in Dda3937 (Z-score: 9.03, P<0.0001). This indicates a significant tendency for Pba1043 CDS that are predicted to have no orthologue in one or more comparator organisms to be located within the regions of divergent base composition predicted by alien_hunter. This is consistent with the hypothesis that the composition of the ‘accessory’ genome of Pba1043 is greatly influenced by horizontal gene transfer.
A similar statistically significant association of predicted islands of CDS in Pba1043 predicted to have no orthologue in at least one comparator organism was observed with predicted regions of putative HGT identified by alien_hunter. In total, 11/16 (Z-score: 3.86, P<0.0001) Pba1043-specific islands; 32/56 (Z-score: 6.29, P<0.0001) Pba-specific islands; 32/60 (Z-score: 5.72, P<0.0001) islands predicted to have no orthologue in Pcc193; and 50/165 (Z-score: 2.39, P<0.01) islands predicted to have no orthologue in Dda3937 were found to overlap with the regions of putative HGT identified by alien_hunter.
It is particularly notable that nearly three-quarters of all Pba1043-specific islands also overlapped at least one region of divergent base composition predicted by alien_hunter. This is consistent with the proposal that these islands have been acquired through lateral gene transfer subsequent to divergence of Pba1043 and Pba1039 from their most recent common ancestor, suggesting a dynamic genome plasticity that persists and distinguishes between Pba strains [22].
Each CDS in the Pba1043 chromosome was classified according to whether a putative protein orthologue was found in completely sequenced plant-associated bacteria (PAB), or in completely sequenced animal-pathogenic enterobacteria (APE) [38]. Those CDS that have at least one such orthologue in PAB, but none in APE were considered potentially to encode a biochemical function that is useful to a plant-associated lifestyle, but likely not to an animal-associated lifestyle. A similar inference may be drawn for CDS for which the Pba1043 sequence shares significantly greater identity with its most similar PAB orthologue than it does with the APE orthologue. As Pba1043 shares a more recent common ancestor with APE such as Yersinia spp. and E. coli strains, such a distribution of orthologous sequences may also imply acquisition by HGT.
An empirical statistical test was performed to determine whether genomic islands in Pba1043 identified by aCGH were enhanced for such CDS. A significant enhancement was seen for 6/56 Pba-specific islands, 6/60 islands with no orthologue in Pcc193, and 9/165 islands with no orthologue in Dda3937 (all tests Z-score>3.0; P<0.001; Table S6). These islands may therefore represent functions that are not only likely to have been acquired by lateral gene transfer, but may also be specific to a plant-associated, and not a generalist or animal-associated, lifestyle. Islands identified in this way include PbaI7, which contains genes that encode for coronafacic acid synthesis, and also a number of hypothetical proteins (see Discussion).
This partitioning of sequences between ‘core’ and ‘accessory’ regions of the bacterial genome, such that variable regions are enhanced for strain- or niche-specific functions has also been observed for other pathogenic bacteria, including P. syringae [9],[39], and appears to be a common strategy for the evolution of these organisms.
Microarray comparative genomic hybridisation (aCGH) is a valuable technique for rapidly, and relatively inexpensively, obtaining comparative genomic data for bacterial strains in a high-throughput manner. However, aCGH has inherent limitations that restrict the applicability of the method, and the information that can be obtained. Foremost is that an aCGH experiment is only able to identify which reference probe sequences do or do not hybridise well to gDNA from a comparator organism. In particular, aCGH is unable positively to identify sequences that are present in the comparator gDNA but that are absent from the reference or otherwise unrepresented in the probe set. Thus aCGH is unable to reflect sequences that are unique to the comparator organism. This may be overcome to some degree by the use of arrays that contain probes not only to the reference organism, but also to other related organisms, as proposed in [5]. Here, the wider the scope of the probes beyond the reference organism alone, the greater is the theoretical coverage of sequences that may be present in the comparator, but not in the reference organism. However, sequences that are unique to the comparator still cannot be disclosed by this approach unless they are present on the array.
It is commonly assumed that aCGH cannot distinguish between sequences that are absent in the comparator gDNA, and those that are merely sufficiently divergent that they cannot hybridise to the array probe set [5],[8, etc]. However, to some degree these classifications are indistinguishable, as the statement that a sequence is ‘absent’ in a comparator can be equivalent to the statement that there no significant sequence similarity. The use of ‘divergent’ as a classifier is ambiguous and potentially misleading in these circumstances. It is also commonly assumed that the assessment of ‘absence or divergence’ reflects overall sequence similarity, and that a relationship between hybridisation and sequence similarity holds for intermediate levels of sequence identity, such that intermediate hybridisation strengths reflect an intermediate degree of sequence identity [13]. An important observation made in [44] was that, even for closely-related sequenced strains of Camplylobacter jejuni, the log ratio of each probe was not sufficient to make a positive prediction of percentage sequence identity. We confirm and extend this observation for SRE with Agilent arrays.
It is often intuitively expected that microarray probes will hybridise to comparator gDNA with a reduced signal, where the comparator sequence is not identical with its homologue in the reference. In all aCGH experiments hybridisation strength is a measurement taken at the reference probe and not across the full length of the sequence from either organism, unless the probe covers the full length of the sequence. Where sequence identity is not homogeneous across the full length of the sequence, or there is similarity between the probe and a non-homologous sequence, this expectation may break down. A comparable break down may occur if there is the possibility of a confounding interaction between hybridising reference and comparator gDNA to a probe. Circumstances in which sequence divergence at the probe hybridisation site is not representative of the overall divergence across the sequence are highly likely to occur, and even under the most favourable circumstances it is only possible to refer to the apparent absence or divergence of sequences in the comparator organism.
Most published approaches to interpretation of aCGH data assume that probes which hybridise strongly to comparator gDNA represent sequences that are present in the comparator organism, while those probes that do not hybridise well represent sequences that may be either absent or divergent in the comparator [e.g.13],[15],[16]. By careful analysis of aCGH data for bacteria with complete genome annotations, we have established that this reasoning, while intuitively plausible, may lead to erroneous conclusions. Our data support only a distinction between those sequences that are, and those that are not, orthologous in the comparator organism. In particular they do not support a distinction between putatively orthologous sequences in terms of their degree of sequence identity, using aCGH hybridisation data. That is, the two sets of putatively orthologous sequences that would be classified as ‘present’ or merely ‘divergent’ could not be distinguished by us in terms of their hybridisation scores or ratios, and therefore the two classes of ‘present’ and ‘divergent or absent’ sequences could also not be distinguished.
We recognise that the array platform itself may be a significant factor in the interpretation of hybridisation data. Our microarray spots were designed with probes of 60 nt in length, one for each CDS, and the L. lactis array data we studied was derived from arrays spotted with amplicons of variable length from 80–800 bp [13]. Previous aCGH studies have employed a number of alternative array constructions, including gene-length cDNA probes, cDNA probes of partial genes, but longer than 60 nt; Affymetrix arrays with multiple short (25 nt) probes per spot; and Agilent arrays with 60 nt probes [10],[12],[14],[55]. Our data demonstrate that conclusions about the relationship between sequence identity and array hybridisation drawn using a particular array technology do not necessarily hold for alternative technologies. Measurement and validation of this relationship is essential for correct interpretation of aCGH data, and should be performed for each array platform.
Hybridisation binding strength or ratio data may also be interpreted in terms of a thermodynamic model of probe binding to the comparator organism gDNA, as an alternative to our interpretation in terms of percentage sequence identity. This is a useful technique when applied to resequencing of strains that are very closely related to the reference, as deviations in hybridisation strength may be accommodated within the thermodynamic model, and sequence differences inferred from observed binding affinities, in terms of that model; it may thus be a better approximator to hybridisation strength than is sequence identity. However we do not use it here as our aim is to infer putative orthology, defined in terms of sequence identity, from hybridisation data. The appropriate measure of putative orthology in this case is sequence identity, and not inferred sequence composition based on a model of the thermodynamic properties of probe binding. An interpretation of measured hybridisation in terms of sequence identity, validated on known sequence data, is therefore the most direct and appropriate approach for this study. Also, a typical bacterial aCGH experiment may involve a comparator organism that displays considerably greater divergence than that which would normally be considered for resequencing or other circumstances in which a thermodynamic model would usually be applied. For example, in our genus-level method validation only 807/10280 (less than 10%) of Pba array probes make a best match (with BLASTN) to the Dda genome that covers the probe to within 5 bp of its length. This significant divergence is likely to induce significant uncertainty, and therefore additional error, in the relationship between base composition as inferred from a thermodynamic model, and the subsequent assignment of putative orthology.
Array CGH has previously been applied, in the main, to closely related organisms; in bacteria, this has usually involved comparisons at the intra- or inter-species level [e.g.5],[6],[7],[9],[13],[45],[46],[48],[56]. In principle, as hybridisation affinity is expected to be influenced by sequence identity, and not by schemes of systematic classification, it should be possible to extend the technique with some success to comparisons between organisms with a more ancient last common ancestor. In particular, DNA-DNA hybridisation studies of Pba1043, Pcc193 and Dda3937, 16S rRNA analysis and phylogenetic considerations (data not shown) indicate genome-wide sequence similarity that justifies the use of aCGH to compare the genome complements of these organisms. In this study, we successfully applied our analysis method to comparison data for Pba1043 and Dda3937: bacteria that differed at the genus level.
It has been noted by other groups that a high degree of sequence divergence between prokaryotes may obstruct aCGH approaches, on the grounds that no strong assumption may be made concerning the distribution of hybridisation ratios for a Lowess normalisation step. Extension of aCGH to more distant comparisons has previously been attempted by modification of the normalisation method used on the array data, such as supervised Lowess (S-LOWESS) [5],[13]. However, we note that Lowess and many other array normalisation methods employ a null hypothesis which assumes that, for a significant proportion of probes, the hybridisation strengths of reference and comparator sequences are random variables drawn from the same distribution. This is a reasonable assumption when applied to isogenic data, such as bacterial mutants, these normalisation operations preserve differences in transcriptional expression while reducing systematic error, as the applied correction of normalisation is valid for the great majority of probes. It is not such a reasonable assumption for aCGH.
Normalisation methods such as Lowess may be useful for aCGH, on the condition that the reference and comparator diverged sufficiently recently, as the proportion of probes that do not conform to the underlying assumptions is likely to be small. This restricts the applicability of aCGH when using these normalisation approaches. However, in cases where the reference and comparator organisms do not share such a recent common ancestor, as for the Pba1043:Dda3937 comparison in which a strict majority of CDS do not have identifiable nucleotide RBH between the organisms (Figure S1), the underlying assumptions of Lowess normalisation fail for the majority of probes.
Subset modifications of Lowess have proven effective on within-species strain comparisons, but require the prior identification of conserved genes, and the assumption that the derived correction is applicable even to the majority of divergent sequences [5],[13]. Therefore in this study we used the nonparametric normalisation method of quantile normalisation (QN) to correct for systematic errors. QN requires no prior assumptions concerning the relatedness of reference and comparator sequences, and specifically makes no assumptions relating to sequence conservation. QN asserts only that the distribution of probe strengths is comparable across replicate arrays, which was established for our data in Figure 1.
The results of our consistency test validation indicate that measures of prediction quality for the interspecies Pba1043:Dda3937 comparison, though lower than that for the interstrain Lactococcus comparison, remain acceptable.
Table 2 demonstrates that the HMM analysis method described in this paper outperforms GACK, MPP and threshold methods in identifying correctly those CDS in Pba1043 that do not have orthologues in Dda3937, and also those CDS in L. lactis MG1363 that do not have orthologues in IL1403. The consistency test of performance on the Pba1043:Dda3937 comparison suggests that GACK has a tendency to overpredict the number of reference sequences that have no orthologue in the comparator, which supports previous observations made using this method [5].
The HMM approach applied herein makes one straightforward improvement to the naïve threshold cutoff classification in that it incorporates information about the state of neighbouring CDS on the genome. Spatial data has previously been incorporated into methods applied to human copy number variation aCGH [18],[19],[20], in which the reference and comparator sequences may be assumed, accounting for noise, to be either so similar as to be near-identical, variant in signal by whole-number ratios, or absent altogether. Bacterial comparative genomic aCGH data also represents sequences that may be nearly identical, occur as copy-number variations, or absent altogether. However, the observed degree of sequence variation in bacterial comparative genomics is very high, and bacterial comparative data also has the potential to include tens to thousands of sequences that may be orthologous or paralogous, and to vary in terms of sequence identity at 50% of their sequence or more. We have demonstrated that the relationship between sequence identity and array hybridisation is complex in this system, and while it was expected that using information about spatial organisation would improve predictive performance, as it has done in the human copy number aCGH problem domain, the magnitude of this improvement was not readily predictable.
The HMM applied here is first-order, and so is the simplest such adaptation that could be applied. Further refinements of the methodology may deliver enhanced predictive performance. Although the resulting improvement in performance over a naïve threshold metric is not as striking as the improvement in relation to GACK and MPP results, it is a consistently better predictor and demonstrates that the incorporation of spatial information about hybridisation scores improves predictive performance. The qualitatively different predictions of the threshold and HMM methods suggest that these approaches identify intersecting subsets of true positives, and that an ensemble approach may be a worthwhile progression of the method.
It is possible that a HMM with bivariate outcomes, or training data (representing cy3 and cy5 intensities) might improve predictive ability of the model. Other methods of identifying an optimal path through the HMM than the Viterbi algorithm are also available. However, our results demonstrate that a HMM with univariate outcome, and using the Viterbi algorithm, performs better than accepted and widely-used approaches on bivariate signal data, and is sufficient to demonstrate that the incorporation of spatial genomic information improves aCGH prediction on bacterial genome sequences.
It might also be interesting to predict membership of one of the populations (high, medium or low hybridisation putative orthologues) observed in Figure 1 for each probe. However, our interest in this study was potential improvements in the prediction of putative orthologues in comparator sequences using spatial information in a HMM, and not optimisation of the predictive HMM.
The improvement over threshold-based prediction seen with the HMM suggests a detectable biological signal from the collocation of sequences in the reference that do not have an orthologue in the comparator, and supports the hypothesis that the ‘accessory’ genome is acquired in large part through genomic islands, rather than individual genes. Accordingly, the number of islands of Pba1043 CDS with no orthologue in the comparator organism was seen to increase with evolutionary distance to a common ancestor. At all evolutionary distances, the predicted islands showed a statistically significant association with regions of putative horizontal gene transfer, whether identified by manual annotation or by the alien_hunter software. Further statistically significant results were observed for the association of these islands with sequences that were putatively orthologous to sequences in plant-associated bacteria, but not in animal-pathogenic enterobacteria. Taken together, this evidence is strongly suggestive of the acquisition of functions specific to the niche of these plant-associated enterobacteria by horizontal transfer.
The identification of islands in Pba1043 that do not have orthologues in Pba1039 is evidence that this process of lateral gene transfer continues in the SRE. Most of the CDS in these islands appeared to be phage-related or to encode hypothetical proteins (Table 4). These may not themselves be critical to the phenotypic differences between strains, but nevertheless indicate a dynamic genome with the potential for acquisition of novel function. However, comparisons at the species and genus level reveal major differences in gene content that may reflect differences in the abilities of each organism to persist in the environment (particularly on plants), and to cause disease on susceptible host plants.
Perhaps the most notable of the accessory islands is that which includes the cfa genes encoding for coronafacic acid (CFA) synthesis. The ability to synthesise this compound appears to be limited to Pba, amongst the SRE. In the plant pathogen Pseudomonas syringae, CFA is coupled to coronamic acid (CMA), to produce the phytotoxin coronatine, which promotes disease through manipulation of plant defences [57]. CFA has also been demonstrated to be required for virulence in Pba1043, providing the first evidence for the involvement of phytotoxins in soft rot pathogenesis, although its precise role has yet to be determined [22],[36]. In Pba1043 the cfa gene cluster is carried on a pathogenicity island highly similar to PAIs found in Pcc193 and other bacterial pathogens. In other pathogens these lack the cfa cluster, but in its place carry other genes with a range of functions, some of which are known to contribute to disease development, such as SPI-7 in Salmonella enterica serovar Typhi that carries the Vi expolysaccharide cluster [53],[54].
Other putative phytotoxic PKS and NRPS, such as the PKS ECA2694 and a syringomycin synthesis-like NRPS are observed to be components of the ‘accessory’ genomes of Pba and Pectobacterium, respectively. Syringomycin is produced by strains of Pseudomonas syringae, and is a virulence factor responsible for pore formation and nutrient leakage through the host cell membrane [57]. The structure of the compound produced by the putatively Pectobacterium-specific NRPS, and any role it may play in virulence are as yet unknown.
The type III secretion system (T3SS) and its translocated effectors promote virulence by the manipulation of host plant defence responses [29]. The T3SS structural apparatus encoded by the hrp/hrc gene cluster appears to be conserved in all four SRE tested. Our HMM predicts that a number of neighbouring effector and helper proteins (e.g. dspEF and hrpW), and agglutinins (hecAB) do not have orthologues in Dda3937 (island DdaI80). However, sequence comparisons between the genomes of Pba1043 and Dda3937 indicate that this result is a false positive, and the Pba1043 CDS do in fact have orthologues in Dda3937. Such false positives may be caused or exacerbated by a tendency to design microarray probes to divergent regions of the reference gene. The distribution of these effectors may be strain-dependent, as the EC16 strain of Dickeya appears not to possess dspE or hrpW in the region flanking the hrp cluster [58],[59].
Some putatively Pectobacterium-specific islands carry genes encoding PCWDE that are not present in Dda3937, such as pectate lyase (pel3) and polygalacturonase (pehA). These differences in PCWDE complement may reflect corresponding differences in environmental niche and/or host range [25]. Other islands carry siderophores similar to pyoverdine and aerobactin, which appear to be restricted to Pectobacterium spp. amongst the SRE tested. Neither of these siderophores yet have a demonstrated association with virulence in pectobacteria, but the siderophores chrysobactin and achromobactin, which are not produced by Pba1043, are known to be involved with virulence in Dda3937 [60],[61].
Several Pba-specific and Pectobacterium-specific islands contain CDS encoding functions that are known or appear to be associated with persistence in the environment, and particularly on plant roots. These functions include phenazine antibiotic production (ehp), multidrug resistance (emr, opr, mex, nfx), and octopine uptake (occ; see Table 3). Phenazine has been shown to target other microorganisms in competition for limited nutrient resources in the rhizosphere [62],[63]. Phenazine does not appear to be produced by Pcc193 or Dda3937, but other antibiotics, such as carbapenem, which is produced by some Pcc strains, may provide equivalent function [64]. The multidrug resistance proteins unique to Pba may provide protection against these compounds. Octopines are tumour-derived compounds produced by the plant using genes transferred during infection by Agrobacterium tumefaciens. The resulting opines, including octopine, are used as a source of nutrition by the bacterium. The octopine uptake CDS, which we find to be putatively Pectobacterium-specific, may reflect an ability to piggyback on Agrobacterium infection and tumour formation [65]. This is an intriguing possibility when taken in context with the observation of genes associated with nitrogen fixation in Pba.
The region encoding putative nitrogen fixation function in Pba1043 is unusual in that orthologues to key genes in this island are found in Dda3937, and predicted to be present in Pba1039, but not in Pcc193. This would appear to imply either that Pcc193 has lost the capacity to fix nitrogen, or that at least one independent acquisition event has occurred to confer this ability. Nitrogen fixation is critical to the nitrogen cycle, and to the soil and rhizosphere environments, in converting atmospheric nitrogen into ammonium compounds that can then be converted by other microorganisms into compounds that may be used by plants. The apparent absence of this capability in Pcc193 suggests that the ability to fix nitrogen is not essential for successful pathogenesis. However, the ability to fix nitrogen may promote establishment and persistence in the environment.
Our aCGH study indicated that HAIs in Pba1043 that were previously identified through manual curation [22] were found to be variably present in the SRE strains investigated (Table 4). A recent study that used 454 sequencing to compare the genome of Pba1043 to the pectobacteria PccWPP14 and P. braziliensis 1692 (Pbr1692) also observed variation in the presence of these HAIs amongst the pectobacteria [66]. HAI2, which in our analysis was found to be present in Pba1043 and Pcc193, was found to be entirely absent from PccWPP14 and Pbr1692, suggesting that its occurrence is sporadic among the pectobacteria. Similarly HAI17, which we predicted to be specific to Pba1043, was found to be present in both PccWPP14 and Pbr1692. These observations confirm the broad theme of our conclusions, that these organisms, though related, have a dynamic, plastic genome composition that results in large functional changes even at the strain level.
A recent study indicates that high-throughput sequencing (HTS; serial analysis of gene expression: SAGE) provides advantages over microarray technology for gene expression analysis, though the false discovery rate appeared to be greater for SAGE [67]. There are several additional disadvantages of aCGH that are overcome by modern HTS methods such as 454 or Solexa/Illumina sequencing. Aside from the question of whether probe hybridisation state is a reliable proxy of sequence identity, or even putative orthology - questions that can be answered directly by sequencing - unlike HTS, aCGH cannot disclose or describe novel sequences in the comparator organism [68],[69]. The cost of sequencing a bacterial genome by these methods is falling rapidly at the time of writing, and there may come a point where the cost of completely sequencing a comparator genome is less than that of carrying out the comparable aCGH experiment. Even before that point is reached, the additional information that HTS provides may be such that it justifies the additional cost of the technique. This area is still moving rapidly, but it has been argued that, for some approaches such as chromatin immunoprecipitation, microarray and HTS experiments complement each other, and the same may be true for aCGH [70].
|
10.1371/journal.pbio.1000618 | Chemotaxis: A Feedback-Based Computational Model Robustly Predicts
Multiple Aspects of Real Cell Behaviour | The mechanism of eukaryotic chemotaxis remains unclear despite intensive study.
The most frequently described mechanism acts through attractants causing actin
polymerization, in turn leading to pseudopod formation and cell movement. We
recently proposed an alternative mechanism, supported by several lines of data,
in which pseudopods are made by a self-generated cycle. If chemoattractants are
present, they modulate the cycle rather than directly causing actin
polymerization. The aim of this work is to test the explanatory and predictive
powers of such pseudopod-based models to predict the complex behaviour of cells
in chemotaxis. We have now tested the effectiveness of this mechanism using a
computational model of cell movement and chemotaxis based on pseudopod
autocatalysis. The model reproduces a surprisingly wide range of existing data
about cell movement and chemotaxis. It simulates cell polarization and
persistence without stimuli and selection of accurate pseudopods when
chemoattractant gradients are present. It predicts both bias of pseudopod
position in low chemoattractant gradients and—unexpectedly—lateral
pseudopod initiation in high gradients. To test the predictive ability of the
model, we looked for untested and novel predictions. One prediction from the
model is that the angle between successive pseudopods at the front of the cell
will increase in proportion to the difference between the cell's direction
and the direction of the gradient. We measured the angles between pseudopods in
chemotaxing Dictyostelium cells under different conditions and found the results
agreed with the model extremely well. Our model and data together suggest that
in rapidly moving cells like Dictyostelium and neutrophils an intrinsic
pseudopod cycle lies at the heart of cell motility. This implies that the
mechanism behind chemotaxis relies on modification of intrinsic pseudopod
behaviour, more than generation of new pseudopods or actin polymerization by
chemoattractants.
| The efficiency, sensitivity, and huge dynamic range of eukaryotic cell chemotaxis
have proven very hard to explain. Cells respond to shallow gradients of
chemotactic molecules with directed movement, but the mechanisms remain elusive.
Most current models predict that cells have an internal “compass”
produced by processing the extracellular signal into an intracellular mechanism
that points the cell towards the gradient and steers it in that direction. In
this article, we present evidence that this internal compass does not exist;
instead, the cell orients itself simply by making use of its
pseudopods—the dynamic finger-like projections on the surface of the cell.
We approached the question by making a computational model of the movement of a
cell without a compass. In this model, the cell moves in a convincingly natural
way simply by using its pseudopods, which respond to positive- and
negative-feedback loops. The concentration of the chemoattractant molecule
modulates the amount of positive feedback. Apart from this, no signal processing
is necessary. This simple model reproduces many observations about normal
chemotaxis. It also accurately predicts the angle at which new pseudopods split
off from old ones, which had not been previously measured. The computational
model thus demonstrates that pseudopod-based mechanisms are powerful enough to
explain chemotaxis.
| Eukaryotic chemotaxis—cell migration towards a source of attractants—is
both biologically important and theoretically interesting, so it has been widely
studied. Recently, a majority of authors have considered that chemotaxis is driven
by a “compass” [1]. The exact meaning of the compass varies. When originally
defined [2], it
implied that there is a simple “compass needle” inside the cell, which
is a localised signal that represents the direction of the chemoattractant gradient
(Figure 1A). If the
hypothetical compass needle points in a different direction from the cell's
current direction, it causes new pseudopods to be made towards attractant sources,
thus steering the cell. More recent compass-based models consider the noisy
environment in which chemoattractants are sensed, allowing the compass to bias
(rather than specify) the positions of new pseudopods. A number of relatives of
compass models (including the LEGI models from the Iglesias and Devreotes groups
[3], the balanced
inactivation model of Levine [4], and inositide-based models such as Narang [5]) share one
property—they focus on information processing, at the level of receptor
occupancy and immediately below, giving the cell a simplified and amplified internal
message that determines the position and direction of future pseudopods. In these
models the cytoskeleton mostly plays a blue-collar role, responding to the
instructions from the internal compass.
However, we [6]–[8] and others [9]–[11] have found that simple
generation of new pseudopods cannot explain observed cell steering and that, unless
gradients are very steep, new pseudopods are usually more strongly controlled by
internal dynamics than by chemoattractants. We have therefore proposed a
“pseudopod-centred” mechanism (Figure 1A) [12], in which there is no
requirement for a compass or other internal messenger representing direction.
Rather, each cell's direction is entirely represented by the pseudopods
themselves. We have demonstrated that new pseudopods are mainly generated by
bifurcation and evolution of existing ones [6],[7]. In a variety of cell types,
close to 90% of new pseudopods are generated when existing pseudopods split
to form two daughters. This severely limits the place and time at which pseudopods
can emerge. We find that directional migration is accomplished by biasing the cycle
of pseudopod generation and retraction, at any of several steps, rather than simply
at the level of new pseudopod initiation. These include selecting the best of
multiple pseudopods generated by random splitting [6] and biasing the position at
which new pseudopods emerge [7]. Several other lines of data support this mechanism. For
example, new pseudopods on average steer the cell away from the
attractant—which disagrees with compass models in which the aggregate effect
of new pseudopods is to steer the cell towards the source.
An alternative, groundbreaking way of addressing the same issues uses a “local
coupling” model (Figure
1A) [13]. Here the leading edge is restricted to a proportion of
the cell and grows by small increments. As with our pseudopod-centred model,
chemoattractants bias an internal process and there is no need for signal
processing. However, this model has two disadvantages. It is limited to cells like
neutrophils with broad, stable leading edges that turn without generating or
retracting pseudopods, and thus does not deal well with cells like Dictyostelium or
macrophages. Similarly, the process that restricts the pseudopod size and prevents
actin polymerization at the sides is central to the model, whereas in many cell
types actin may polymerize at any part of the cell [14]. In this work we therefore
addressed the pseudopod-centred model as a potential broad or universal model for
chemotaxis.
We tested the predictive abilities of pseudopod-centred mechanisms using a
conceptually simple computational model, based on coupled feedback loops (Figure 2A). Feedback is
fundamental to chemotaxis [15] and underpins both compass- and pseudopod-centred
mechanisms but used in different ways. In compass models, feedback is typically
invoked during signal processing, to amplify and simplify the noisy and complex
information from receptors [16]–[18]. We have predicted that such signal processing is not
essential [12],[19]. Rather, in our model feedback loops are used to
define the pseudopods themselves. Positive feedback allows pseudopods to
maintain themselves and to grow, while negative feedback fulfils two
roles—firstly, it restricts the growth of existing pseudopods and the
initiation of new ones at other parts of the cell, and secondly, it makes
pseudopods dynamic, allowing cells to change shape and direction as occurs in
amoeboid movement. To model chemotaxis, we therefore adapted an established
system (Figure 1B) [20] based on
a single pseudopod activator regulated by three feedback loops, one positive and
two negative (Figure 1B). In
the Meinhardt article [20], the cell does not move—the components of the
feedback loop were localised, in a dynamically evolving pattern, within a static
cell perimeter. To allow our simulated cell to move (Figure 1C), we have used an evolving surface
finite element method [21], in which each point of the perimeter moves outwards
normal to the edge of the cell [22], at a rate
proportional to the local activator level. Protrusion is counteracted by a
curvature-based contraction, in which the edge is retracted so that the cell
tends towards a constant area. The different parts of the cell retract in
proportion to their steepness of curvature; this effectively simulates cortical
tension, which retracts highly curved areas and thus causes the cell to tend
towards a circle. A level set method was used to evolve the cell perimeter.
These methods are based on an Eulerian description of a level set function,
where the location of the zero level set identifies the cell perimeter [23]. This
framework confers many well-known computational advantages, including use of
fixed Cartesian meshes and straightforward implementation of high resolution
numerical schemes [24]. Full details of the computational methods are given
in a separate publication [25].
Importantly, movement of the leading edge greatly changes the evolution of the
activator levels, because areas where the level of pseudopod activator is high
tend to expand, diluting the activator. Evolution of the edge therefore mimics
the local inhibitor, in a way that might make possible a future model with only
two further feedback loops, one positive and one negative.
The centre of our model is a pseudopod activator whose level correlates with the
rate of movement of the leading edge. One biologically appropriate equivalent is
actin nucleation driven by the Arp2/3 complex, which is a central driver of
actin-based movement. However, the components of the model are not intended to
directly represent defined molecular species. This is for two reasons. Firstly,
the regulation of the actin cytoskeleton is not understood in the quantitative
detail needed to generate a defined model. Key components have not been defined
or cannot be measured (for example, the affinity of activated Rac for the
SCAR/WAVE complex), and multiple factors such as VASP may modulate the rate of
actin-based protrusion. Secondly, actin-based motility is frequently regulated
by multiple parallel components, so removing individual pathways such as
SCAR/WAVE, Rac, or PI 3-kinase does not block migration, despite the clear
importance of each of these pathways. Molecule-based models have been successful
and informative about individual pathways and the roles of single proteins [26]–[29], but the
dynamic morphology of chemotactic cells has proven too complex for such an
approach. Our approach is more similar to those successfully used by the Wang,
Theriot, and Mogilner labs based initially on cell shape [30] and mechanics [31]. While the
activator is directly related to the level of actin nucleation or
polymerization, we envisage the local inhibitor corresponding to depletion of
required substrates (for example, Arp2/3 complex, activation-competent
SCAR/WAVE) and the global inhibitor corresponding to physical processes such as
mechanical tension. The positive feedback loop driving pseudopod growth could
act at multiple levels, including through Rac [32], SCAR/WAVE [10], or actin
itself [33];
all three have been described and probably act concurrently in real cells,
though we envisage the first two as being more influential. Again, however, the
aim of this model is to test the predictive power of pseudopod-centred models,
rather than the roles of particular pathways.
The results from this simulation (Movie S1 and Figure 3A–C) make several clear points.
Firstly, cells polarize into a front and a rear without needing additional
internal signals (Figure
3A). This polarization is seen as an essential part of efficient
migration and chemotaxis [17]. Secondly, the simulated cells' migration is
persistent—they maintain their direction over several pseudopod cycles
(Movie
S1 and Figure
3B). Persistence has also been measured in most migrating cells and is
thought to be important for chemotaxis [34],[35]. Thirdly, new pseudopods
are mostly made by bifurcation of the leading edge (Figure 3C; compare with Dictyostelium cell in
Figure 3D). Bifurcation
(“pseudopod splitting”) was initially described by Andrew [6] and has
since been observed in multiple types of migratory cells [36], including mouse embryonic
fibroblasts, human dendritic cells, and cultured neurites. In the measured cell
types, the proportion of pseudopods generated by splitting is usually around
90% [6].
Analysis of the positions of pseudopods as they evolve over time also gives a
wavelike pattern (Figure
3E), like that measured in real unstimulated cells [9]. Furthermore, the paths
taken by individual cells are remarkably similar between the simulation and real
cells (Figure 3F), and both
display characteristics of a persistent random walk [37]. The simple model based on
Meinhardt [20] therefore successfully describes a typical
unstimulated cell.
To generate a pseudopod-centred model of the response to chemoattractant, we
departed from the Meinhardt model [20], which relies on hidden
signal processing to provide a fully localised signal (see Methods). In our system, the magnitude of the positive
feedback is directly correlated with the local chemoattractant receptor
occupancy (Figure 2B,
equation 4), with additional elements corresponding to noisy signal perception
and activator feedback. This provides a key difference between our model (and
pseudopod-centred models in general) and most work in the chemotaxis field. In
our model, neither actin polymerization nor pseudopod generation is caused by
extracellular signals. Rather, the signals are only able to modulate the rates
of internal processes. In shallow gradients, internal processes overwhelmingly
dominate.
When this connection to external signalling is added and a moderate
chemoattractant gradient (from 5.3 nM to 6.5 nM across the cell) is applied, the
simulated cell moves very similarly to a real Dictyostelium in a similar
gradient (Movie
S1; Figures
4A,B). This close resemblance to real cells is surprising, given a number
of disagreements with generally accepted points.
Firstly, as previously stated there is no direct connection between the external
signal and protrusion, pseudopod generation, or actin polymerization. The
receptor occupancy only modulates the positive feedback that maintains the
leading pseudopod. Secondly, there is no signal processing—each point on
the cell's surface is modulated by the local attractant concentration,
without reference to points elsewhere in the cell.
Thirdly, receptor adaptation is not required for effective chemotaxis up a static
attractant gradient, even at fairly high receptor occupancy, as long as there is
a significant difference in the proportion of occupied receptors across the
cell. Parts of the edge that lack pseudopods do not gain them when the overall
occupancy increases, because positive feedback of the activator is negligible
when activator levels are near zero. For adaptation to be dispensable
contradicts most current opinion but is supported by several articles, including
those showing non-adaptation of movement to high stimuli [38] and lack of
adaptation at the G-protein level as measured by FRET [39]. Adaptation at some
levels occurs biologically and is required for conditions such as chemotaxis
towards sources of biological waves. It was nonetheless surprising that the
model would support simple chemotaxis up a linear gradient without
adaptation.
The basic motile behaviour of the cell is not fundamentally changed by the
chemoattractant. As observed in real cells chemotaxing in moderate gradients
[6] but in
disagreement with many compass-based explanations for chemotaxis, the rate of
pseudopod generation and orientation of new pseudopods are only slightly changed
by the chemoattractant.
As the steepness of the attractant gradient applied to the model increases, the
accuracy of chemotaxis increases, exactly as seen in real cells (Figure 4C). More surprisingly,
however, with steep gradients the model undergoes a qualitative change that
precisely resembles real cells. While in low gradients cells nearly always turn
from the front, by biasing the behaviour of leading pseudopods, in high
gradients they frequently generate a new pseudopod directly towards the
attractant source [40]. The model replicates this behaviour (Movie S2;
Figure 4D), which was
unexpected because we had believed it to be driven by an alternative mechanism.
This suggests that pseudopod-centred models can account for the mainstream data
supporting signal-induced pseudopods, given steep enough gradients.
When chemotactic cells are presented with a sudden, global change in attractant
levels, they respond in a well-defined way. First actin polymerizes all around
the cell perimeter, then the cell rounds up as the new F-actin is
depolymerized—the “cringe” response [41]—which is followed by
repolarization, formation of new pseudopods, and a second peak of actin
polymerization (Movie S3). Because this behaviour is consistent and tractable it has
been widely used as an assay for chemotactic signal transduction, and the second
peak in F-actin in particular has been attributed to a downstream response to PI
3-kinase activation [41]. When the computational model was subjected to a
similar sudden increase in receptor occupancy, with the addition of an
exponential decay function representing adaptation, the perimeter behaved in a
similar fashion to the experimental observations (Movie S4;
Figure 4E). Activator
levels—corresponding to polymerization of actin filaments—rose
rapidly, and modelled cells rounded up, followed shortly afterwards by a drop in
activator levels as the inhibitors responded. The time for cells to recover is
defined by the rate of adaptation, not by the feedback loops. Strikingly,
however, a second complex activator peak occurred that strongly resembles the
second experimentally observed F-actin peak (Figure 4F).
This provides an alternative mechanistic explanation for the generation of
multiple F-actin peaks. Instead of two pathways with different signal
propagation times, as previously predicted [41], the multiple peaks in the
model are caused by damped oscillation of a single pathway following a sudden
displacement. In this explanation, mutants that mostly lack a second F-actin
peak [42] do
so because of inefficient positive feedback at the pseudopod level, not separate
signalling pathways with different dynamics.
Two further observations support the appropriateness of the pseudopod-centred
computational model. Firstly, movement and chemotaxis are relatively robust. The
parameters we use (Table S1) are mainly taken directly from
Meinhardt [20] and did not need optimization to produce biologically
plausible behaviour. Two-fold changes in most of the parameters make only minor,
quantitative differences to the behaviour of the simulated cells (Figure S1);
indeed many of the single changes shown appear to make chemotaxis more efficient
than in our standard conditions. Interestingly, the parameters that were most
sensitive to alteration concerned the production of the local inhibitor; changes
in the production or decay rates bc and
rc resulted in either slower migration or
repeated movements that are inconsistent with random migration. Raising the
diffusion coefficient of the activator (Da) caused
similar problems with repetition, but these could be compensated by
corresponding rises in the diffusion coefficient of the local inhibitor
(Dc). Secondly, the model handles noise very
effectively. Even when the contribution of noise is far greater than the signal
from a shallow gradient, chemotaxis is efficient; in shallow gradients,
chemotaxis is most efficient over a substantial background of noise (Figure S2).
Robustness and tolerance to noise are central to chemotaxis in real cells [43].
In compass models of chemotaxis, cells first identify the direction of the
attractant gradient, then generate new pseudopods if the cell's direction
needs correcting [2]. However recent work suggests that pseudopods do not
steer cells this way in shallow gradients. Instead at least two mechanisms act
concurrently, both based on the tendency of new pseudopods to be made by
bifurcation of existing ones. In the first, new pseudopods are made without
requiring external guidance, but cells preferentially retain ones that point in
the correct direction [6]. In the second, new pseudopods are generated in
stereotypical directions by bifurcation, but their orientation is biased by the
direction of the gradient, leading to accurate steering after a number of slight
turns [7].
Both mechanisms act concurrently in real cells, though either would be
sufficient for chemotaxis alone. We therefore examined the steering of simulated
cells to determine whether each of these mechanisms was used.
As discussed previously, our computational model generates new protrusions, by
bifurcating existing ones, and retracts others [6]. For the analysis of
pseudopod selection, we simulated migration in a moderate gradient (initially
5.3–6.6 nM across a cell; shallower gradients give similar but less
emphatic results). The point at which each pseudopod splits was identified using
a peak detection algorithm, and the daughters followed until one was retracted.
We then measured the initial angle of each new pseudopod relative to the
direction of the attractant gradient. Figure 5A shows that the simulated cells use
selection like real cells (see Figure 5B, data from [6] Figure 3A)—new
pseudopods that are pointing well away from the correct direction are nearly
always retracted, while pseudopods that point up-gradient are more likely to be
retained. The pseudopod selection mechanism is therefore operating for simulated
cells as in live cells.
The only obvious difference between the simulated and real data is in the
distribution of new pseudopod directions—the “rabbit ears” in
the real cells are caused by unequal bifurcation in which the smaller pseudopod
points off to one side. Simulated cells bifurcate symmetrically, and thus the
distribution of new pseudopods is more even.
The second measured pseudopod-centred mechanism of chemotaxis is directional
bias. During bifurcation, new pseudopods can only be generated in a narrow range
of angles either side of the parent. However, the mean orientation of new
pseudopods is biased slightly towards the attractant source [7]. This
makes the path qualitatively similar in the presence or absence of an
attractant, but biases accumulate over time and steer the cell. To test whether
our simulated cells used this mechanism, we repeatedly reoriented the stimulus
as the cell turned (Movie S5). This caused the cell to move in
circles. Figure 5C shows
that pseudopod bias in the modelled cells is similar to the experimentally
measured bias (Figure 5D,
replotted from [7]). The mean position of new pseudopods was biased about
15° towards the stimulus. Note that (as in real cells) pseudopods are biased
whether they steer the cell towards or away from the stimulus, emphasising that
the timing and general location of pseudopod production are not altered by the
stimulus. Rather, in both simulation and reality, cell-autonomous processes
control the rate and general site at which pseudopods are made and the general
area they emerge, and chemoattractant signalling fine-tunes this behaviour.
For a more detailed analysis of how bifurcations are affected by the attractant
gradient, we simulated inaccurate migration in shallow gradients and counted a
large number of bifurcations. We then measured the angle between the dominant
pseudopods before and after the split (schematic, Figure 5E). In simulations run with zero
external stimulus, the mean change in the absence of signal is about 55°. We
then measured how this angle varied for cells migrating at different angles
relative to the attractant gradient. When simulated cells were moving towards
the attractant source, the mean change dropped (Figure 5F) to about 30°; as the angle
between cell and attractant gradient increased, the mean angle between
successive dominant pseudopods also increased by a ratio of about 1° of
pseudopod per 3° of additional orientation away from the chemoattractant. At
70° between the gradient and the new pseudopod, the mean angle between
successive dominant pseudopods was not altered. Thus the model predicts that the
pseudopod split angle is smoothly biased by the attractant direction, in a way
that partially compensates for the tendency of new pseudopods to direct the cell
away from the attractant and which will steer the cell towards the attractant
source over a number of turns.
To compare the simulations with real cells, we examined the data generated by
quantitation of movies of cells turning in shallow gradients (Figure 5G; new data, extracted
from the same data set as examined in [7]). Again, the correlation
between simulated and real data is surprisingly good. Our model thus predicts
new data as effectively as it recreates the multiple known features of migration
and chemotaxis described previously.
As discussed earlier, we used an undefined model because the large number of
incompletely defined pathways makes them require multiple biologically
improbable presumptions. Furthermore, many pathways that were thought essential
turn out to be dispensable for chemotaxis [44],[45]. However, all of the
components required to drive the simulation have physiological equivalents. As
discussed earlier, the core activating term corresponds to actin activation,
most likely through the Arp2/3 complex. At least three positive feedback loops
of the type we use have been described—direct autocatalysis of actin,
actin polymerization generating templates for Arp2/3 complex activation, and
actin activation of PI 3-kinase.
Our pseudopod-centred mechanism efficiently couples gradient sensation to
migration, overcoming a long-term problem with chemotaxis models [46]. The
similarity between the behaviour of modelled and real cells is astounding,
especially given the conceptual simplicity of the model and the robustness of
the model to changes in parameters. Two apparently separate mechanisms of
chemotaxis—pseudopod selection and orientation bias—both emerge from
the same simple model, and the complex patterns of actin polymerization and
depolymerization following sudden stimuli are also clearly observed without
multiple signalling pathways. This emphasises that much of the described complex
behaviour of cells is likely to be an emergent property derived from relatively
simple pathways.
This implies that future understanding of chemotaxis will require a change in
experimental approach. Current research often focuses on how external signals
are amplified and processed, and separately on pathways that initiate new actin
and new pseudopods. The success of our pseudopod-centred model suggests that a
greater emphasis on the physiological mechanisms of pseudopod evolution, and how
chemoattractants modulate them, will yield greater fundamental insight.
A complete description of the numerical methods used is far beyond the scope of
this article and is fully presented in reference [25]. In brief, equations
(1)–(3) in Figure 2B
are approximated on the evolving cell perimeter using an Arbitrary Lagrangian
Eulerian surface finite element method using piecewise linear elements. Time
integration is achieved using a semi-implicit approach. The computed activator
profile is used to drive a mechanical model of the protrusive and retractive
forces exerted on the cell membrane. Movement of the cell is obtained using a
level set method and a moving Cartesian mesh. Calculations are performed using
the level set toolbox in MATLAB [24].
The fourth equation in Figure
2B, defining the signal, is different from Meinhardt's [20]. In the
original Meinhardt model, the location on the cell membrane with the highest
receptor occupancy is used—without specification of how it is
computed—to centre an assumed sinusoidal variation of the external signal.
That model, unlike ours, therefore bypasses a key question in chemotaxis.
Instead we relate the signal to the local proportional receptor occupancy, with
additional random terms representing noise in the pseudopod system and in the
receptor signalling system.
The cells in Figure 3D are
Dictyostelium AX3 cells, developed for 4 h and imaged exactly as described in
[6]. For
fluorescence microscopy, similar cells were transfected with an extrachromosomal
vector expressing GFP-lifeact and imaged using an Olympus confocal microscope
with a 60×1.4 NA objective.
Pseudopod angles were measured using Quimp3 [8] from the same dataset that
was used in [7].
|
10.1371/journal.pgen.1008209 | Time of day and network reprogramming during drought induced CAM photosynthesis in Sedum album | Plants with facultative crassulacean acid metabolism (CAM) maximize performance through utilizing C3 or C4 photosynthesis under ideal conditions while temporally switching to CAM under water stress (drought). While genome-scale analyses of constitutive CAM plants suggest that time of day networks are shifted, or phased to the evening compared to C3, little is known for how the shift from C3 to CAM networks is modulated in drought induced CAM. Here we generate a draft genome for the drought-induced CAM-cycling species Sedum album. Through parallel sampling in well-watered (C3) and drought (CAM) conditions, we uncover a massive rewiring of time of day expression and a CAM and stress-specific network. The core circadian genes are expanded in S. album and under CAM induction, core clock genes either change phase or amplitude. While the core clock cis-elements are conserved in S. album, we uncover a set of novel CAM and stress specific cis-elements consistent with our finding of rewired co-expression networks. We identified shared elements between constitutive CAM and CAM-cycling species and expression patterns unique to CAM-cycling S. album. Together these results demonstrate that drought induced CAM-cycling photosynthesis evolved through the mobilization of a stress-specific, time of day network, and not solely the phasing of existing C3 networks. These results will inform efforts to engineer water use efficiency into crop plants for growth on marginal land.
| Crassulacean acid metabolism (CAM) photosynthesis represents an important adaptation to arid environments as CAM plants take up CO2 at night when evapotranspiration rates are lower. Genomes and large-scale datasets are available for several plants with constitutive CAM activity, but they provided little insight on how this trait evolved from C3. Here we sequenced the CAM-cycling plant Sedum album, which switches from C3 to CAM photosynthesis under drought conditions. We performed a global gene expression analysis sampling every two hours over 24 hours in C3 (well-watered) and CAM (drought) conditions. This comparative approach allowed us to identify components of the CAM pathway that were previously unidentified in constitutive CAM plants such as pineapple and orchid. Our results reveal a massive time of day specific rewiring of the transcriptional networks in Sedum where only 20% of cycling genes overlap between the C3 and CAM conditions. This time of day reprogramming results in broad network changes linking stress pathways and photosynthesis under CAM.
| Drought is the most pervasive abiotic stress and plants have evolved diverse strategies to mitigate the effects of water deficit [1]. Most water loss in plants occurs through transpiration as a byproduct of daytime stomata mediated CO2 uptake. Crassulacean acid metabolism (CAM) plants have evolved an alternative carbon assimilation pathway to store CO2 nocturnally when evapotranspiration rates are lower [2]. CAM plants store and concentrate atmospheric or respiratory CO2 at night through the carboxylation of phosphoenolpyruvate (PEP) by the enzyme phosphoenolpyruvate carboxylase (PPC). The resulting four carbon acid, oxaloacetate, is subsequently reduced to malate by malate dehydrogenase (MDH), and then transported to the vacuole as malic acid, producing the characteristic nighttime acidification observed in CAM plants [3]. During the day, malic acid is decarboxylated to release the CO2 for fixation by Rubisco. Because of this temporal separation, CAM plants have remarkably high water use efficiency, and use roughly 35% less water than C4 plants and up to 80% less water than comparable C3 species [4, 5]. These traits make CAM an attractive model for engineering improved water use efficiency and drought tolerance into crop plants that may be grown on more marginal land prone to seasonal droughts[6].
The CAM pathway is highly plastic and occurs along a continuum ranging from plants that are predominately C3 with weak or inducible CAM activity to constitutive species with highly optimized and efficient CAM. This diversity reflects the multiple origins of CAM and its utility in different environments [7, 8]. The biochemistry and physiology of CAM was worked out in species from the Crassulaceae, Cactaceae, and Bromeliaceae that exhibited strong constitutive CAM regardless of environmental conditions [9, 10]. Constitutive CAM plants often inhabit arid and semi-arid environments or live in epiphytic habitats subject to seasonal drought. In regions with seasonal or periodic drought, facultative and CAM cycling CAM species often display a flexible system of switching from C3 to CAM under drought conditions. Facultative CAM species have high nocturnal stomatal conductance under drought and CAM-cycling plants display typical C3 diel stomatal conductance, but re-fix respiratory CO2 at night. The tremendous variation of CAM is partially explained by convergent evolution, since CAM has evolved independently at least 40 times across 35 diverse plant families [7]. Model CAM species have emerged for several families and a recent wealth of genetic and genomic resources have advanced our understanding of CAM pathway evolution. The genomes of three constitutive CAM species have been sequenced including pineapple (Ananas comosus) [11], the orchid Phalaenopsis equestris [12], and the eudicot Kalanchoe fedtschenkoi [13]. Draft genomes of the facultative CAM orchids Dendrobium catenatum [14] and D. officinale [15] are also available, as are transcript and expression studies across the Agavoideae [16].
Facultative CAM and CAM-cycling species provide an excellent system for dissecting the molecular basis of CAM as we can directly compare expression dynamics between C3 and CAM gene networks. Comparisons between C3 and CAM states can be used to identify metabolite, physiology, and gene expression changes associated with CAM evolution. Transcriptome and metabolite surveys of the facultative CAM species Talinum triangulare identified a set of core CAM pathway genes and candidate transcription factors mediating photosynthetic plasticity [17]. Though this study captured changes throughout drought progression, limited temporal resolution did not capture the circadian components of facultative CAM. Time series microarray data from ice plant (Mesembryanthemum crystallinum) also identified components of CAM induction, and suggested that there is a circadian regulated 4–8 hour phase shift underlying the C3 to CAM shift [18]. Taking an evolutionary approach comparing cycling genes in C3, C4 and CAM (Agave) plants, it was found that CAM photosynthesis resulted from both accelerated evolution of specific protein domains as well as reprogramming of diel networks [19]. However, it is still unclear how these time of day networks interact with the circadian clock and how the C3 to CAM switch is mediated within a single system.
The circadian clock has evolved to optimize the daily timing, or phase, of cellular biology with the local external light and temperature cycles[20]. In the model C3 species Arabidopsis thaliana, between 30–50% of genes cycle under diel conditions with control by at least three evolutionarily conserved transcriptional modules defined by the morning (ME: CCACAC; Gbox: CACGTG), evening (EE: AATATCT; GATA: GGATA), and midnight (TBX: AAACCCT; SBX: AAGCCC; PBX: ATGGGCC) [21]. Moreover, this time of day transcriptional network ensures development and environment-specific growth through opposing phytohormone influences[22], and is conserved across distantly related C3 and C4 species [23, 24]. The circadian clock directly controls components of the CAM pathway and is thought to coordinate the temporal oscillations of CAM. PPC is activated nocturnally via phosphorylation by the circadian clock activated phosphoenolpyruvate carboxylase kinase (PPCK) [25]. Silencing of PPCK in Kalanchoë fedtschenkoi not only reduces CAM activity by ~66%, but also perturbs central components of the circadian clock, suggesting PPCK may be essential for clock robustness [26]. Well-characterized CAM pathway genes in pineapple have acquired novel clock associated cis-elements compared to their orthologs from C3 and C4 species, suggesting a broad control of CAM by the core circadian clock [27, 28]. Expression patterns of clock genes are conserved under C3 and CAM mode in ice plant [29] but the role of the clock in facultative and weak CAM induction remains unclear.
Here we generated a draft genome for the CAM-cycling species Sedum album using long read Single Molecule Real-Time (SMRT) sequencing and surveyed high-resolution temporal gene expression, metabolite, and physiology changes across a diel time course in well-watered (C3) and drought (CAM-cycling) conditions. This comparative approach revealed a massive time of day specific rewiring of the transcriptional networks in S. album where only 20% of cycling genes overlap between the two conditions. This work demonstrates for the first time that a phase shift in the core circadian clock underpins a massive time of day rewiring that enables S. album to switch to CAM photosynthesis and overcome water stress.
We assembled a draft genome of S. album to serve as a foundation resource to understand the genome-wide changes associated with the transition from C3 to drought induced CAM photosynthesis. We estimated the genome size to be 611 Mb based on flow cytometry, which was significantly larger than a previously reported value for S. album (142 Mb)[30]. Karyotype analysis suggests S. album is tetraploid (2n = 4x = 68) with a deduced diploidized genome size of ~305 Mb (S1A Fig). This is consistent with the kmer based genome size estimate, which revealed a heterozygous peak (first) with the full genome at 799 Mb and monoploid (second peak) at 256 Mb (S1B Fig). Because of this complexity, we utilized a PacBio based SMRT sequencing approach to build the draft genome. We generated 4.4 million PacBio reads collectively spanning 33.7 Gb or 55x genome coverage. Raw PacBio reads were error corrected and assembled using the two leading PacBio assemblers: Falcon [31] and Canu [32]. Canu was able to accurately phase the highly similar homeologous regions and produced an assembly of 627 Mb across 15,256 contigs with an N50 of 47kb, designated as the S. album V2 genome. The average nucleotide similarity of homeologous regions is 99.6% based on the Canu assembly, suggesting S. album is either autotetraploid or allotetraploid with two highly similar genomes. Falcon collapsed the homeologous regions into a single haplotype with a total assembly size of 302 Mb across 6,038 contigs and an N50 of 93kb. The Canu and Falcon based contigs were polished to remove residual errors using high-coverage Illumina data with Pilon [33].
The S. album assemblies were annotated using the MAKER-P pipeline [34]. We utilized transcripts assembled from the timecourse RNAseq data (described below), and protein datasets from other angiosperms as evidence for ab initio gene prediction. After filtering transposon derived sequences, the tetraploid and diploidized assemblies contained 93,910 and 44,487 gene models respectively. The two resulting polished assemblies were highly complete with a BUSCO score of 97%. The tetraploid assembly is highly fragmented with lower contiguity, and partial gene models may explain the greater than 2x number of genes compared to the diploid assembly. We identified homeologous genes between the diploid and tetraploid assembly and compared their relative expression. Homeologs within the tetraploid assembly have low sequence divergence and most RNAseq reads map ambiguously to both homeologs, leading to similar transcript quantification for each homeolog in the pair. Comparison of expression between homologous genes in the tetraploid and diploid assemblies revealed a strong correlation across all timepoints collected from well-watered (C3) or drought (CAM) timecourses (C3 r = 0.927, CAM r = 0.901; S2 Fig). Given the similarity in expression and for simplicity of downstream analyses, we proceeded with the diploidized assembly produced from Falcon, and this version is referred to as the S. album V3 genome.
We surveyed synteny and gene level collinearity between S. album and the closely related constitutive CAM plant Kalenchoe fedtschenkoi. Synentic regions between the two genomes were largely collinear, with conserved gene content and order (Fig 1). Roughly 71% of the S. album V3 genome had detectable synteny with K. fedtschenkoi, but this is likely an underestimation given the fragmented nature of both assemblies. Syntenic depth for each S. album region ranged from 1–5 to each region of K. fedtschenkoi, reflecting the two shared whole genome duplications events in the Crassulaceae [13] and polyploidy in S. album (S3 Fig). Collinear regions were larger in K. fedtschenkoi compared to S. album, highlighting the relatively compact nature of the S. album genome (Fig 1). To estimate the divergence time of the S. album subgenomes, we calculated Ks (synonymous substitutions per synonymous site) between homoeologous gene pairs based on synteny with K. fedtschenkoi. We identified a single peak with a median Ks of 0.026, corresponding to an estimated divergence time of ~853,000 years (S4 Fig). This supports a relatively recent polyploid origin of S. album.
We surveyed patterns of gene family dynamics across the genomes of seven CAM, five C3 and two C4 plants to identify expanded and contracted gene families in S. album. After normalizing for ploidy, we identified 33 gene families that were uniquely contracted in S. album and 41 that are uniquely expanded compared to the other species (S1 Table). Among the contracted gene families in S. album are ABA responsive proteins, sucrose transporter 2, Aluminum activated malate transporter 1, and several flavonoid biosynthesis pathways. Among the expanded gene families in S. album are various transporters, photosynthesis related proteins, and early light induced proteins, among others (S1 Table).
When S. album is subjected to drought conditions, it switches from robust uptake of CO2 during the light phase to low rates of nocturnal carbon assimilation [35], which is consistent with weak CAM-cycling induction (Fig 2). We collected parallel diel timecourse data under well-watered (C3) and drought (CAM-cycling) conditions to identify genes, cis-elements and networks associated with the switch from C3 to CAM-cycling. The CAM-cycling timecourse was collected after 14 days of sustained drought when the CAM pathway was induced to conserve water usage (Fig 2A). The well-watered (C3) S. album had mild diel titratable acid and stomatal aperture changes, whereas nocturnal acid accumulation and reduced stomatal aperture were observed under drought (Fig 2B and 2C). Sedum has relatively low rates of carbon assimilation under drought conditions [36], so it is difficult to speculate on stomatal conductance based on aperture alone. However, the stomatal aperture was reduced under drought compared to well-watered samples and with no significant changes in nocturnal stomatal aperture. Together, this suggests that S. album is a drought-inducible CAM-cycling plant, and much of the nocturnal carbon assimilation is likely from recycled respiratory CO2, as previously reported [37].
Parallel sampling of C3 (well-watered) and CAM-cycling (14-day drought) plants were collected every two hours over a 24-hour diel timecourse in triplicate for RNA sequencing and metabolite measurements. Differentially expressed genes between pairwise C3 and CAM-cycling timepoints varied from 702 to 3,245 genes, reflecting massive transcriptional reprogramming (S1 Table). GO enrichment analysis of these differential expressed genes indicated that most of the up-regulated genes under drought and CAM induction were related to abscisic acid responses, oxidative stress, and water deprivation at both day and night. Genes responding to ethylene were mainly enriched at night (S2 Table). Down-regulated differential expressed genes in the CAM samples were enriched with genes involved in cell wall metabolism, organization and biogenesis reflecting the reduced growth under drought conditions. There was also an over-representation of down-regulated genes related to photosynthesis at night (10pm-4am) in drought stressed S. album.
After filtering genes with low expression (total TPM < 5), 29,372 genes were used to construct a differential co-expression network utilizing a weighted correlation network analysis (WGCNA) approach [38]. This approach grouped genes into 32 and 25 co-expression modules for the C3 and CAM-cycling networks respectively. Preservation of modules between the C3 and CAM networks was low (Fig 3A), consistent with massive rewiring of the expression network under drought stress and CAM induction. The core CAM pathway genes (described in more detail below) were found in different modules in the CAM-cycling network, suggesting regulation of CAM and C3 pathways through distinct transcriptional programs and network rewiring.
Photosynthesis and stress related pathways were co-regulated in the drought induced CAM-cycling co-expression network. The C3 network module 3 and CAM-cycling network module 7 were enriched in photosynthesis-related GO terms such as photosynthetic electron transport chain (GO:0022900), response to light stimulus (GO:009416), and photosynthesis light harvesting in photosystem I (GO:0009765; S3 and S4 Tables). CAM-cycling network module 7 was also enriched with GO terms related to stress response (GO:0006950) and oxidative stress (GO:0006979) compared to the orthologous module in the C3 network, which had no enrichment. GO terms related to photosystem II assembly and function were over-represented in the C3 network module 3 but not in the corresponding CAM-cycling module 7. This indicates photosystem I and electron transport chain pathways are co-regulated with stress related genes under drought but not photosystem II. This is likely due to drought induced damage on photosystem II and the tightly coordinated regulation of stress responsive and photosynthesis pathways to mitigate photooxidative and drought associated damage. This hypothesis is further supported by the enriched GO terms in CAM-cycling network module 15 and 19, which contain ATP metabolic process, ATP hydrolysis coupled protein transport, and superoxide metabolic processes (S3 and S4 Tables). These GO terms were not enriched in any well-watered, C3 network modules. Reactive oxygen species scavengers accumulate during water deficit to prevent oxidative damage [39], and scavengers such as superoxide dismutase are up-regulated in S. album under drought stress (S5 Fig). The comparison of gene co-expression networks between well-watered (C3) and drought conditions (CAM-cycling) support the tight regulation between photosynthesis and stress responses.
In pineapple, core CAM pathway genes have diel expression patterns along with enriched circadian associated cis-elements [27, 28]. Most drought related genes are circadian regulated [40] and proper diurnal expression is associated with growth and resilience under water deficit [41–43]. Therefore, we also estimated the periodicity, phase (peak time of expression) and amplitude differences of each S. album gene under either C3 or CAM-cycling using the JTK_CYCLE program [44]. There were 10,278 (35%) and 11,946 (41%) genes cycling under C3 and CAM-cycling conditions respectively (Table 1). These numbers are similar to those reported for Arabidopsis [21] and other species [24] under diel conditions of 12 light/12 hrs dark (LD) and thermocycles (HC: hot/cold). Surprisingly, only 22% (6,480) genes cycled under both conditions, and of these shared cycling genes, 39% (2,504) had a change in amplitude and 75% (4,579) had a change in phase. In addition, almost 50% of genes displayed a unique pattern of cycling in only one condition. These results corroborate the differential expression and network analysis, which is consistent with a massive time of day rewiring of the transcriptional machinery in S. album under drought and during the switch to CAM-cycling photosynthesis.
To understand the shifts in the S. album circadian clock during C3-CAM conversion, we looked at cycling expression patterns (S5 Table, S6 Fig). First, we found an expansion of core clock genes in S. album compared to Arabidopsis, with large changes in circadian clock associated 1/late elongated hypocotyl (CCA1/LHY, 7 vs. 2), PRRs (10 vs. 4), early flowering 3 (ELF3, 5 vs. 2) and lux arrhythmo (LUX, 7 vs. 5) and additional copies of gigantea (GI) and cryptochrome 1 (CRY1) (S6 and S7 Tables). Many of these expansions were shared with K. fedtschenkoi [13], pineapple[27], and orchid[12], suggesting gene expansion may be linked to the circadian regulation of CAM pathways (S7 Table). While none of the CAM species have a copy of the AtPRR9 ortholog, there were two PRRs only found in S. album and K. fedtschenkoi (S7 Fig). In general, the phase of the cycling of S. album circadian clock gene orthologs matched those of Arabidopsis [21, 24] but they showed a phase shift under drought stress (S6 Table, S6 Fig). While GI, CCA1/LHY and ELF3 maintained the phase of expression between C3 and CAM in S. album, the amplitude of GI and ELF3 decreased under CAM (25–50%) and CCA1/LHY amplitude increased (25%) under CAM. In contrast, most of the PRR1s, PRR5s (but not PRR7/3s) displayed a 1–4 hr phase shift (later) under CAM-cycling conditions, which is consistent with the global 1–4 hr phase shift (S5 Table). By comparing the phase and amplitude changes observed in drought treated S. album with the drought treated C3 plant Brassica rapa and two CAM systems (A. comosus and K. fedtschenkoi) (S6 Table), we were able to untangle the circadian rhythmicity change unique to drought treated S. album. Shared amplitude changes between S. album and B. rapa may represent universal drought responses and those that are unique to S. album represent lineage specific drought responses and changes related to CAM induction. We identified several core clock genes with divergent changes in phase and amplitude. For instance, CRY1 genes were cycling in both conditions of S. album but only under well-watered conditions of B. rapa (S6 Table), suggesting the circadian rhythm of CRY1 was conserved in S. album under drought stress. In addition, flavin-binding, kelch repeat, f box (FKF) displayed cycling pattern in all three CAM plants surveyed but not in the two conditions of B. rapa.
Interestingly, we observed a phase shift where genes with rhythmic expression in well-watered conditions were over-represented at phases 11–17 while genes with rhythmic expression under drought were enriched in later phases (phases 18–23) (Fig 4A). This observed phase changing is not solely caused by drought or CAM photosynthesis but an interaction of both, as drought treated B. rapa and CAM performing A. comosus have different responses (Fig 4A). Therefore, the phase shift of S. album genes under drought is possibly caused by the involvement of both circadian and drought-related cis-element, and expression shift of the circadian clock genes mentioned above. We identified a complete conservation of the time of day overrepresentation of the diel and circadian-related cis-elements, consistent with what we have found in other plant systems[21, 23, 24]. In addition, the time of day overrepresentation of both the evening specific cis-elements, evening elements and TBX, displayed a 1–4 hr phase shift between C3 and CAM-cycling, while the morning specific elements ME and Gbox did not (S8A Fig). Furthermore, there were many more significant evening-specific cis-elements under CAM that were distinct from conserved elements (S8B Fig), including 8 novel drought/CAM-specific cis-elements in S. album that are enriched in specific phases of drought but not well-watered (Fig 4B, S8 Table). These elements were mainly enriched at the later phase, which is consistent with the phase shift of drought-specific cycling genes we observed. While all eight elements were over-represented in drought-specific cycling genes (S9 Table), elements CAM-1 and CAM-7 were also enriched in cycling genes that had higher amplitude in drought condition (Fig 4C). These eight novel cis-elements support our findings that the time of day networks are completely rewired under drought induced CAM photosynthesis, and that is not just a phasing of the C3 networks but instead a new drought and CAM specific network.
In addition to the novel cis-elements, genes with CAM and drought specific cycling or higher amplitude in CAM were enriched with the stress-mediated ABRE (YACGTGGC) and ABRE-like (BACGTGKM) cis-element in their 2kb upstream promoter sequences (Table 1). CAM specific rhythmic genes were also enriched with the MYB transcription factor binding (MACCWAMC) cis-elements. In contrast, cycling genes with higher amplitude under C3 were enriched with the circadian related evening element [45] and MYC2 transcription factor binding site. To understand the biological functions of genes in each category, we performed a GO enrichment test. Genes with cycling expression in C3 only were mainly enriched with GO terms related DNA replication, translation and protein related metabolic process. Enriched GO terms in genes with CAM specific cycling were mainly related to transcription and gene expression regulation by RNA. Genes that are cycling in both conditions, but with higher amplitude in CAM were over-represented by GO terms involved in stress response including water deprivation (S10 Table).
Genes in the core CAM pathway belong to large gene families shared across plant genomes, and most have functions unrelated to CAM activity. Independent CAM lineages show evidence of convergent amino acid substitutions in core enzymes [13], but independent recruitment of the same orthologs remains untested for most CAM genes. Through comparing diel expression patterns under C3 and CAM induction, we identified S. album genes associated with the core CAM pathway and putative vacuolar transporters involved in metabolite transport. We also compared the annotated S. album CAM genes with their pineapple and K. fedtschenkoi orthologs to identify similarities among the three sequenced CAM plants with diel expression data [13, 27, 28]. Putative CAM pathway genes in S. album had higher expression (PPC, PPCK, NADP+-ME, MDH, PPDK) or become oscillating (PPCK, NAD-ME) under drought induced CAM-cycling compared to C3 (Fig 5, Table 2). Core CAM pathway genes were identified in the constitutive CAM plants K. fedtschenkoi and pineapple using a similar approach [13, 27].
Beta carbonic anhydrase (β-CA) converts CO2 to HCO3- in the first step of nocturnal carboxylation in CAM photosynthesis. None of the annotated β-CA genes in S. album had high nocturnal expression under drought induced CAM-cycling compared to C3 (S6 Fig). Low β-CA expression was also observed in the drought inducible CAM plant Talinum triangulare [17] and constitutive CAM species Yucca aloifolia [46]. In contrast, β-CA from pineapple and K. fedtschenkoi had strong nocturnal expression, which may support a higher conversion rate of CO2 to HCO3- (S9 Fig). Nocturnal carbon assimilation in the weak CAM plant S. album likely occurs at a much lower rate than constitutive CAM species. Thus, β-CA may not be essential to provide additional bicarbonate for phosphoenolpyruvate carboxylation in the dark period during CAM-cycling in S. album.
Phosphoenolpyruvate carboxylase (PPC) and Phosphoenolpyruvate carboxylase kinase (PPCK) are the core CAM genes that mediate the fixation of CO2 at night. Under CAM induction, these genes had up to 73x fold higher expression than C3. PPC (Sal_001109) has high expression during the day while PPCK genes (Sal_021277, Sal_045582, Sal_050851) were expressed mainly nocturnally (Fig 5). PPC transcript expression was slightly shifted compared to pineapple (Aco010025.1) and K. fedtschenkoi (Kaladp0095s0055.1), while PPCK expression pattern was more conserved across the three CAM species (Fig 6A). Although there are multiple copies of PPC in each plant genome, S. album and K. fedtschenkoi recruited the same PPC ortholog for CAM (Fig 6B). Given the unique monocot and eudicot specific duplications of PPC, it is not possible to assess if the same PPC was recruited in monocot and eudicot CAM plants. Based on the conserved diurnal expression of PPC, we surveyed cis-element enrichment across CAM, C4, and C3 orthologs to identify cis-elements related to circadian regulation. The morning element [21] circadian clock motif was present in the PPC promoters from all three CAM plants, as well as the two C4 plants (Fig 6B). In addition, TCP and MYB transcription factor binding sites were also present in the three CAM PPC promoters.
Based on expression dynamics between conditions, diurnal decarboxylation in S. album is likely driven by chloroplastic NADP+-malic enzyme (NADP+-ME, Sal_015733) and pyruvate orthophosphate dikinase (PPDK, Sal_044029; Fig 5). K. fedtschenkoi also utilizes the ME and PPDK pathway but pineapple primarily utilizes the phosphoenolpyruvate carboxykinase pathway (Aco010232.1) for decarboxylation [13, 27]. Nocturnally fixed CO2 is converted to malate and stored in the vacuole to avoid cytosolic pH fluctuations and prevent feedback inhibition of PPC [10]. While the proton pumps required to establish a gradient in the vacuole are well-characterized [47], the malate transporters remain uncharacterized in CAM plants. The most likely transporters responsible for malate influx and efflux are homologs of the Arabidopsis clade II aluminum activated malate transporter (ALMT 3/4/5/6/9) and tonoplast dicarboxylate transporter respectively [28, 48, 49]. However, under CAM-cycling conditions, no S. album ALMT genes shared a conserved transcript pattern with pineapple and K. fedtschenkoi. The tonoplast dicarboxylate transporter ortholog in S. album had peak expression in the early morning in both C3 and CAM-cycling mode, but the expression drops to almost zero at dusk (ZT12) in CAM-cycling. The sudden expression decrease coincided with the titratable acid accumulation (Fig 1B). This expression pattern was similar in pineapple and K. fedtschenkoi where their expression decreases sharply (>10 fold reduction) before sunset and gradually increases before sunrise (S10 Fig). This supports a potential conserved role of tonoplast dicarboxylate transporter as a malate exporter across CAM plants.
To elucidate the potential regulators of CAM-cycling photosynthesis, gene regulatory networks were constructed and compared between the C3 and CAM timecourses and putative activators and suppressors controlling core CAM genes were identified (Table 3; Fig 3B and 3C). Core CAM pathway genes shared few interactors between C3 and CAM-cycling supporting independent networks. PPC expression in the CAM-cycling network was associated with the appearance of 12 putative activators and the disappearance of 2 suppressors. The three copies of phosphoenolpyruvate carboxylase kinase (PPCK) showed similar regulatory patterns. Interestingly, one of the PPCK copies (Sal_021277-RA) interacted with the ABA biosynthesis gene 9-cis-epoxycarotenoid dioxygenase (Sal_006713-RA), an ABA exporter (Sal_009291-RA) and an ABA receptor regulator (Sal_007192-RA) under C3 conditions and this interaction disappeared when S. album switched to CAM-cycling under drought. PPCK phosphorylates PPC at night to reduce the allosteric inhibition of PPC by malate. This phosphorylation process is under circadian control [26] and the changes of interactions of core clock genes with PPCK are evident in our gene interaction networks. Two CCA1/LHY gene copies (Sal_047094-RA and Sal_051928-RA) interacted with PPCK (Sal_021277-RA) under both conditions but two additional CCA1/LHY copies (Sal_050684-RA and Sal_051929) only interacted with PPCK under well-watered condition while CCA1/LHY_C1 (Sal_047093-RA) and PPCK only interacted in drought treated S. album. The only other CAM core gene that had putative gene interactions with core clock gene CCA1/LHY was malic enzyme (ME), where some of these interactions were only observed in the drought network (S11 Table). Malate dehydrogenase (MDH) was potentially controlled by 312 activators and 2 suppressors, including phosphoglycerate mutase (Sal_015714-RA), which is involved in glycolysis and gluconeogenesis. Four additional glycolysis and gluconeogenesis pathway genes (Sal_003892-RA, Sal_043275-RA, Sal_048516-RA, and Sal_050408-RA) were also predicted to be involved in activating ME expression in CAM photosynthesis. In addition, starch synthesis pathway genes including ADP glucose pyrophosphorylase large subunit (Sal_005475-RA and Sal_005955-RA) and starch branching enzyme (Sal_041721-RA) were potential activators for ME and pyruvate orthophosphate dikinase (PPDK) respectively. Our gene regulatory networks demonstrate a direct link between CAM photosynthesis and carbohydrate metabolism (Fig 3B, c) and the drought induction of CAM photosynthesis requires the rewiring of large number of existing genes, similar to patterns observed in pineapple [27].
Nocturnal synthesis of PEP is essential for maintaining high levels of CAM activity. PEP is supplied through the degradation of transitory starch or soluble sugars [50], and mutants deficient in starch degradation have impaired CAM activity [51]. Transitory starch accumulated in S. album throughout the day and was depleted nocturnally in both C3 and CAM, but the total amount of starch synthesized was higher when S. album was operating in CAM-cycling mode (Fig 7). A large proportion of genes in the starch biosynthesis pathway (phosphoglucoseisomerase, starch branching enzymes, granular bound starch synthase and glucan water dikinase) had higher expression in CAM-cycling than C3 (S11A Fig). Starch is mainly synthesized in the stroma from glucose 6-phosphate (G6P) supplied from the Calvin Benson cycle through the conversion of F6P to G6P via glucose-6-phosphate isomerase (PGI). G6P can also be imported from the cytosol through a glucose 6-phosphate/phosphate translocator (GPT). The chloroplastic PGI is tightly regulated by multiple factors in Arabidopsis and this can limit the flux of starch synthesis [52, 53]. When the required flux of starch synthesis exceeds the limit set by PGI (such as the switch to CAM), plants may bypass the rate limiting plastidic PGI pathway by exporting carbon as triose phosphates. Carbon from cytosolic PGI can then be remobilized to the plastid as hexose phosphate using the GPT. Genes involved in this PGI bypass pathway were upregulated in CAM compared to C3 including the triose phosphate/phosphate translocator (Sal_024485-RA) (S11B Fig). Most of the transcripts for genes involved in the branch of the pentose phosphate pathway from glucose 6-phosphate to ribulose 5-phosphate (G6P shunt) in both the plastid and cytosol were down-regulated in the CAM mode (S11C Fig). Together, these data suggest that metabolism in CAM-cycling S. album is shifted to increase starch synthesis while avoiding carbon loss through the G6P shunt, which is consistent with the hypothesis previously proposed by Sharkey and Weise [54].
There are plastidic GPTs in plants, however they are not normally expressed in autotrophic tissue. It is hypothesized that this transporter is not expressed to prevent the pool of G6P from reaching the Km threshold of glucose-6-phosphate dehydrogenase in the chloroplast [54, 55]. A large pool of G6P in the chloroplast could result in a futile cycle and loss of carbon as CO2 through the G6P shunt [54]. A S. album ortholog to the Arabidopsis GPT2 gene (Sal_018693-RA) had peak daytime expression in both C3 and CAM-cycling but showed drastically increased expression (ranged from 11 to 205-fold higher) in CAM-cycling condition (Fig 8, S12 Table). A similar diurnal expression pattern was observed in the constitutive CAM plants pineapple and K. fedtschenkoi, but not the C3 plant Arabidopsis (Fig 8B). In contrast, the S. album GPT1 ortholog had constitutive expression in both C3 and CAM-cycling, and is likely not associated with CAM activity (S13 Table). Expression dynamics between CAM and C3 indicates that the starch synthesis from G6P in S. album might be regulated at multiple pathways including starch synthesis, plastidic PGI bypass, G6P shunt, as well as triose phosphate and G6P transporters as summarized in Fig 8C.
At night, transitory starch is broken down into either maltose and glucose via the hydrolytic degradation pathway, or to G6P via the phosphorolytic degradation pathway [56]. It is thought that starch degradation shifts from hydrolytic to phosphorolytic pathway during the transition from C3 to CAM in ice plant [56, 57]. Chloroplasts isolated from the ice plant under C3 photosynthesis mainly export maltose during starch degradation and switch to G6P export under CAM. We measured the total G6P and maltose contents of S. album across the C3 and CAM diurnal timecourses. C3 performing S. album had a higher abundance of maltose than CAM-cycling plants at night but the G6P content was similar in both C3 and CAM-cycling plants (Fig 7). We also examined the transcript level of genes proposed to be involved in the two starch degradation pathways (hydrolytic and phosphorolytic) [56]. Homologs of Arabidopsis glucan phosphorylase (LSF2), β-amylase 6 and β-amylase 9 were expressed higher in well-watered conditions but glucan phosphorylase (SEX4), β-amylase 2, β-amylase 3, β-amylase 7 and isoamylase 3 were expressed strongly under drought (S11D Fig). Hydrolytic pathway genes were not universally upregulated in either photosynthetic condition. This indicates that either S. album does not switch its starch degradation pathways during the C3-CAM conversion, or the regulation of this shift was not controlled at the transcriptional level.
Facultative CAM and CAM-cycling plants are optimized for rapid growth under favorable C3 conditions and sustained resilience under drought through CAM related water conservation. This reversible mechanism makes facultative CAM and CAM-cycling plants an ideal model to dissect the genetic basis of CAM and for engineering improved water use efficiency and drought tolerance in crop plants. Engineering inducible CAM photosynthesis into C3 crops could improve water use efficiency and resilience under prolonged drought stress while maintaining high yields when field conditions are ideal. Establishing high quality genomic resources are an essential foundation for downstream functional genomics of CAM. The complex polyploidy of S. album presented a computational challenge, but through a combination of long read sequencing and optimized assembly we produced a high quality, diploidized reference.
High-resolution diurnal expression data for both C3 (well-watered) and CAM-cycling (drought) provided a comparative system for identifying genes and pathways involved in CAM. Utilizing a comparative differential co-expression approach, we found a strong link between stress responses and photosynthesis, supporting the tight regulation of CAM photosynthesis by drought networks. The CAM-cycling and C3 co-expression networks have a low degree of overlap, supporting massive reprogramming of genes related to drought and CAM photosynthesis induction. This is further supported by the gene regulatory networks, which showed major shifts in gene interaction under CAM induction. In addition, the large-scale rewiring of gene rhythmicity and phase underlies the drought protective mechanism in S. album. The time of day reprogramming is comprised of phase and amplitude change of core clock genes, and expression regulation under known and novel drought/CAM-specific cis-elements, which in turn lead to a phase difference between well-watered and drought cycling genes. The genome level phase shift of transcript expression is uniquely observed in S. album and this pattern was not observed in drought treated B. rapa and the constitutive CAM plant A. comosus. This highlights the complexity of CAM-cycling induction and the tight coordination of multiple metabolic pathways including stress responses, photosynthesis, and circadian rhythm.
We observed conserved diurnal transcript expression of core CAM genes (PPCK, ME, PPDK) and potential transporters among independent CAM lineages, but patterns are not conserved for every gene (i.e. PPC, MDH). Though PPC is highly expressed in all CAM plants, the diel oscillation varies between species. PPC orthologs from the same clade (PPC-1 and not PPC-2) are always recruited for CAM related carboxylation [58] and the copies from the CAM species surveyed have common cis-regulatory elements related to circadian regulation. Based on our transcript data, β-CA likely has minimal function in CAM-cycling S. album compared to the constitutive CAM pathway where strong nocturnal expression of β-CA might be essential to rapidly assimilate and store carbon. The non-cycling expression of β-CA was also reported in the inducible CAM plant Talinum triangulare [17]. We hypothesize that the nocturnal β-CA expression is coupled with the high nocturnal stomatal conductance in constitutive CAM plants, which enhances the nocturnal CO2 fixation rate. β-CA involved in normal metabolic processes in the cytosol is probably sufficient for CAM activity in weak, low assimilating species, but enzymatic activity remains to be surveyed. β-CA is not essential for efficient operation of the C4 pathway in maize [59], but the dispensability of β-CA in CAM is untested.
Transitory starch content is high in CAM plants to provide enough substrate to regenerate PEP. CAM induced expression changes suggest S. album increases the transitory starch content through multiple pathways. Starch biosynthesis pathway genes have increased expression under CAM-cycling compared to C3. During CAM-cycling, G6P transporters and triose phosphate transporters are upregulated and genes involved in the G6P shunt are downregulated. The diurnal induction of GPT transcripts and transport activity under CAM has been reported in ice plant previously [60, 61]. Our high-resolution transcriptome data showed that expression of GPT2 increased after CAM-cycling induction while GPT1 had similar expression in both conditions. Therefore, GPT2 is the likely controller of G6P translocation in CAM-cycling photosynthesis. In C3 plants, the G6P pool in the stroma is 3 to 20 times lower than the cytosol [62, 63]. We thus speculate that CAM plants activate GPT2 in the daytime to import G6P into the stroma for transitory starch synthesis, this would bypass the plastidic PGI which can be kinetically limiting to starch synthesis and result in higher stromal G6P concentration. Interestingly, the S. album GPT2 had a low but consistent diurnal oscillation under C3 conditions while the Arabidopsis GPT2 has low constitutive expression. Inducible CAM plants might have C3-CAM intermediate transcript expression for certain genes in order to rapidly switch between photosynthetic modes in response to abiotic stress.
Concentrations of total cellular G6P and maltose and transcript expression of the starch degradation pathway genes revealed the complexity of PEP regeneration during the C3-CAM transition. The minimal change of G6P content between well-watered and drought conditions could be due to high cytosolic G6P content masking the G6P produced through phosphorolytic starch degradation. It could also indicate that there is no increase in phosphorolytic starch degradation during C3-CAM conversion. The increase in transcript expression of many hydrolytic pathway genes might suggest that there is no switch of starch degradation pathways from hydrolytic to phosphorolytic in S. album, contrasting pattern observed in ice plant [60, 61]. However, the starch degradation patterns observed in ice plant were assayed with isolated chloroplasts, and in vivo degradation patterns may differ when concentrations of G6P and maltose in both the chloroplast and cytosol become kinetically relevant. Alternatively, the shift of starch degradation pathways could be species specific.
In S. album, C3 to CAM-cycling switching results from a complete rewiring of the time of day networks. We found that there was a specific drought/CAM network that leverages a novel set of cis-elements. While the circadian cis-elements are completely conserved under C3 conditions, an alternative transcriptional module is leveraged to engage CAM photosynthesis. Underlying this alternative pathway may be a new repertoire of core circadian genes since we observed a significant increase in these genes as well as distinct phasing and amplitude under CAM-cycling conditions. Core circadian genes have been shown to be retained after whole genome duplication (WGD) and these new gene copies may provide a means to augmented stress responses.
Drought responsive and CAM pathways are intrinsically linked in facultative and CAM-cycling plants. Disentangling these two processes would ideally require a C3 Sedum outgroup with no detectable CAM-cycling activity. CAM occurs along a continuum within Sedum, ranging from CAM-cycling and facultative to constitutive[36]. Sedum is highly polyphyletic[64] and the C3 Eurasian species are in phylogenetically distinct clades, making detailed comparisons with CAM lineages challenging. Furthermore, CAM anatomy and gene expression predate CAM origins within Yucca and more broadly across Agavoideae [46, 65], and CAM-like signatures of gene expression under drought may be observed in C3 Sedum species. The phase shift during CAM-cycling induction in S. album was not observed in similar drought timecourses from B. rapa [43], and few genes with cycling expression are overlapping. This suggests the massive shifts of expression in S. album are driven by CAM related activities and not simply conserved drought responses. More detailed comparisons are needed to parse pathways related to CAM and drought in S. album. Genomic and high-resolution transcriptomic data from S. album provide a valuable foundation for downstream functional genomics work on the molecular basis of drought induced CAM-cycling photosynthesis. These pathways may be useful for engineering improved water use efficiency and drought tolerance into C3 or C4 crop plants.
Sedum album plants were vegetatively propagated via cuttings in growth chambers under 12-hour photoperiod with day/night temperatures of 24°C and 20°C respectively. The light period started at 8am (ZT0) and ended at 8pm (ZT12). To induce CAM activity, plants were subjected to moderate drought by withholding water for 14 days. Well-watered (C3) control plants were watered every 2 days for the duration of the experiment. For the 24-hour diurnal timecourse experiments, leaves from 5 pots of randomly sampled plants were pooled for each replicate and three biological replicates were collected for each timepoint. The well-watered (C3) 6am (ZT22) and drought (CAM) 8am (ZT0) samples had two and one replicates respectively due to loss of samples. Samples were collected every 2 hours over the 24-hour experiment. All tissues were frozen into liquid nitrogen immediately and stored at -80C.
Carbohydrate and titratable acidity data was collected for each of the samples used in the C3 and CAM timecourse experiment. Carbohydrate and titratable acidity measurements were taken from the same tissue samples used for RNAseq experiment. Frozen leaf tissue (0.1–0.5g) was ground, weighed and transferred into 1.5mL tubes. 500uL of 3.5% perchloric acid was added to the leaf tissue and mixed by vortexing. The resulting supernatant (perchloric acid extract) was neutralized to pH 7.0 using neutralizing buffer (2M KOH, 150mM Hepes, and 10mM KCl). The samples were then frozen to precipitate salts, centrifuged, and the supernatant was transferred to a new 1.5mL tube for carbohydrate assays. Sucrose, glucose, fructose and fructose-6-phosphate, β-maltose, and glucose-6-phosphate were measured. The pellet was resuspended in 1mL of 80% ethanol and mixed by vortexing followed by two rounds of washing. The pellet was air-dried for 30 min, resuspensed in 200mM KOH and incubated at 95°C for 30 min. 1M acetic acid was used to adjust the pH to 5.0 and 5 ul of an enzyme cocktail containing 5 units α-amylase (E-ANAAM Megazyme, Bray, Wicklow, Ireland) and 6.6 units amylogucosidase (E-AMGDF Megazyme) was added to each tube and incubated at room temperature for 24 hours. The resulting supernatant containing glucose was transferred to a new tube for starch content measurement using the glucose assay.
Glucose content was measured using a 96 well plate in a Filter Max F5 plate reader (MDS Analytical Technologies, Sunnyvale CA, USA) at 340 nm with an NADP(H)-linked assay. Wells were filled with 200 μl of 150 mM Hepes buffer pH 7.2 containing 15 mM MgCl2, 3 mM EDTA, 500 nmol NADP, 500 nmol ATP and 0.4 units glucose-6-phosphate dehydrogenase (G8529 Sigma St. Louis MO, USA). Five μl of sample was added to each well and the reaction was started by adding 0.5 units of hexokinase (H4502 Sigma). Absolute glucose content was determined using an extinction coefficient of 6220 L mol-1 cm-1 for NADPH at 340 nm[66]. Sucrose content was assayed as reported above with 20U of invertase (Sigma-aldrich, I4505) and 0.5U of hexokinase. Other carbohydrates (β-maltose, glucose-6-phosphate, glucose, fructose and fructose-6-phosphate) were measured as previously reported by Weise et al. [63, 67]. All enzymes used were purchased from Sigma-Aldrich and 4 unit of maltose phosphorylase was used.
To measure titratable acidity, ~0.3g of fresh ground tissue was mixed with 3mL of 80% ethanol and boiled at 80°C for 60 min. The supernatant was cooled to room temperature and titrated with 0.1N sodium hydroxide until an endpoint pH of 8.3. Titratable acid (in μ Eq per gram of fresh weight) was calculated as volume of 0.1N sodium hydroxide X 0.1 X 1000 / Fresh weight (gram).
Epidermal peels from well-watered (C3) and drought stressed (CAM) S. album leaves were prepared according to Wu et al. [68]. Microscope slides were visualized using a Nikon Eclipse Ni-Upright microscope with 40X differential interference contrast objective lens and images were captured with Nikon DS-Fi3 camera. Stomatal width and length was measured using program NIS-Elements and recorded in an Excel file. A minimum of 10 individual stoma from 5 leaves were examined for each time point (ZT06 and ZT22) and condition (C3 and CAM-cycling). A Student’s t-test was performed to test for significant changes in stomatal aperture.
High molecular weight (HMW) genomic DNA was isolated from well-watered S. album leaf tissue for PacBio and Illumina sequencing. HMW gDNA was isolated using a modified nuclei preparation [69] followed by phenol chloroform purification to remove residual contaminants. PacBio libraries size selected for 25kb fragments on the BluePippen system (Sage Science) and purified using AMPure XP beads (Beckman Coulter). The 25kb PacBio libraries were sequenced on a PacBio RSII system with P6C4 chemistry. In total, 4 million PacBio reads were sequenced, collectively spanning 33.7 Gb or 55x genome coverage. Illumina DNAseq libraries were constructed from the same batch of HMW gDNA using the KAPA HyperPrep Kit (Kapa Biosystems) followed by sequencing on an Illumina HiSeq4000 under paired end mode (150 bp). In total, 38 Gb of Illumina data was generated for error correction, representing ~62x coverage.
RNA was extracted using the Omega Biotek E.Z.N.A. Plant RNA kit according to the manufacturer’s protocol. RNA quality was examined on a 1% agarose gel and RNA concentration was quantified using the Qubit RNA HS assay kit (Invitrogen, USA). 2μg of total RNA was used to construct stranded RNAseq library using the Illumina TruSeq stranded total RNA LT sample prep kit (RS-122-2401 and RS-122-2402). Multiplexed libraries were pooled and sequenced on HiSeq4000 using paired end 150nt mode.
Given the complex polyploidy of S. album, several long read assembly algorithms were tested for their ability to resolve and collapse homeologous regions. Raw PacBio reads were error corrected using Falcon (V0.2.2)[31] and Canu (V1.4) [70]. Parameters for Falcon were left as default. The following parameters for Canu were modified: minReadLength = 2000, GenomeSize = 612Mb, minOverlapLength = 1000. The Canu assembly accurately separated the homeologous regions and produced an assembly of 627 Mb across 15,256 contigs with an N50 of 47kb. The Canu assembly graph was visualized in Bandage [71] and most of the nodes (contigs) were highly interconnected. The assembly graph complexity is likely caused by a combination of polyploidy and heterozygosity. This Canu based assembly is referred as tetraploid genome version of S. album and named as Sedum V2 genome. Homeologous regions in the Falcon based assembly were highly collapsed and the total assembly size was 302 Mb with 6,038 contigs and an N50 of 93kb. This assembly is about half of the estimated genome size, and represents a diploidized version of the tetraploid S. album genome. We named this Falcon version as diploid genome version and designated as Sedum album V3 genome. We assessed the completeness of the tetraploid and diploidized assemblies using the benchmarking universal single-copy orthologs (BUSCO; v.2) [72] with the plant specific dataset (embryophyta_odb9). ~97% of the 1,440 plant specific genes were identified in both S. album assemblies, supporting they contained an accurate representation of the gene space.
Contigs from the Falcon and Canu assemblies were polished to remove residual errors with Pilon (V1.22) [33] using 62x coverage of Illumina pared-end 150 bp data. Illumina reads were quality-trimmed using Trimmomatic [73] followed by aligning to the assembly using bowtie2 (V2.3.0)[74] with default parameters. The total alignment rate of Illumina data was 92.3% for the Falcon assembly and 93.1% for Canu, suggesting both were largely complete. Pilon parameters were modified as followed and all others were left as default:—flank 7,—K 49, and—mindepth 15. Pilon was run four times to remove any residual errors.
The S. album diploid and tetraploid assemblies were annotated using the MAKER-P pipeline [34]. A custom library of long terminal repeat [75] retrotransposons was constructed for repeat masking. LTR harvest (genome tools V1.5.8) [76] and LTR Finder (v1.07) [77] were used to predict putative LTRs and this candidate list was refined using LTR retriever (v1.8.0) [78]. Parameters for LTR harvest and LTR finder were assigned based on suggestions from the LTR retriever package. This high quality library was used as input for RepeatMasker (http://www.repeatmasker.org/) [79] with implementation in the MAKER pipeline. RNAseq data from the C3 and CAM-cycling timecourses was used as transcript evidence. Representative transcripts from the RNAseq data were assembled using Trinity [80] with default parameters. Protein sequences from Arabidopsis [81] and the UniprotKB plant databases [82] were used as protein evidence. Ab initio gene prediction was done using SNAP [83] and Augustus (3.0.2) [84] with two rounds of training. Transposable element derived gene models were filtered using a library of representative transposases.
To identify homologs between the diploidized and tetraploid genome assemblies, a reciprocal blast approach was used. First, blastp was run using the tetraploid gene model sequences as query against the diploid gene models with an e-value cutoff of 0.001, maximum target sequence of 1, HSP ≥ 100, protein similarity ≥60%, and match length > 50 amino acids. A reciprocal blastp search was performed with the reversed query and database using the same parameters except without maximum target sequence option. Only gene pairs retained in both blastp results are identified as homologs between the two genome versions.
Expression profiles between the tetraploid and diploid assemblies were compared to assess if the diploid assembly contained a representative gene set for downstream analyses. Expression levels (in TPM) of the RNA-seq samples were quantified using Kallisto v0.43.0[85] with the tetraploid and diploid gene model sets. Expression values from all timepoints collected under C3 or CAM were averaged and plotted for homologous gene pairs in the diploid and tetraploid genome assemblies. Pearson correlation of all homologous gene pairs were calculated to assess the transcript expression similarity between two genomes. Based on the high expression correlation, the diploid version of the assembly and annotation was used for downstream analyses.
Paired end raw reads were trimmed using Trimmomatic v0.33 [73] to remove adapters and low quality bases. Quality trimmed reads were pseudo-aligned to the S. album gene models to quantify expression using Kallisto v0.43.0[85]. Default parameters were used for Kallisto and 100 bootstraps were run per sample. Transcript expression was quantified in transcripts per million (TPM) and an averaged TPM from the three replicates was used for gene co-expression and single gene analyses. Differential expressed genes of each timepoint between C3 and CAM samples are identified using R program sleuth. Likelihood ratio test and Wald test were used and only genes with q-value < 0.05 and b-value > |1| in both tests were categorized as differentially expressed (DE) gene.
The timecourse RNAseq data was clustered into gene co-expression networks using the R package WGCNA [38]. Prior to network construction, genes were filtered based on TPM and any gene with total TPM < 5 across all samples or 25% of datapoints have zero expression was removed. In total, 29,372 genes were used to construct separate networks for the well-watered (C3) and drought (CAM-cycling) timecourses. Parameters for the C3 network were as follows: power = 7, networkType =“signed", corType = "bicor", maxPOutliers = 0.05, TOMType = "signed", deepSplit = 3, mergeCutHeight = 0.01. For the CAM-cycling network, the following parameters were used: power = 10, networkType = "signed", corType = "bicor", maxPOutliers = 0.05, TOMType = "signed", deepSplit = 3, mergeCutHeight = 0.15.
The JTK_CYCLE program [44] was used to detect rhythmic expression patterns of genes across the two timecourses. JTK_CYCLE was implemented in the R package MetaCycle (v. 1.0.0) [86]. The 29,372 genes used to construct co-expression networks were tested for rhythmicity under the C3 and CAM-cycling conditions. Genes with Bonferroni adjusted p-value < 0.05 were classified as cycling and changes in amplitude were assessed by subtracting the amplitude under C3 from that for CAM-cycling. Gene with amplitude difference less than 3 was categorized as having no changes between the two conditions. The same 29,372 genes used in WGCNA and JTK_CYCLE analysis were used for gene interaction network construction using default setting of program CMIP [87, 88]. The threshold for the C3 and CAM networks was 0.47 and 0.52, respectively. The gene interaction pairs between C3 and CAM networks were compared. Gene interacting pairs that were found in the C3 network but not in the CAM network were classified as “repressor” and gene interactions that were predicted in the CAM network but not in C3 were categorized as “activator”.
For circadian rhythmicity, JTK_CYCLE analysis across the four species (S. album, B. rapa, A. comosus and K. fedtschenkoi) was performed as stated above. The B. rapa well-watered and drought datasets, were downloaded from NCBI GEO GSE90841 and only the second 24 hours of data from the well-watered and drought conditions were used for analysis. For A. comosus, the leaf green tip and white base datasets published in [28] were used and K. fedtschenkoi leaf timecourse dataset were downloaded from the NCBI SRA database (BioSample SAMN07453940-SAMN07453987). An expression matrix table was provided by Dr. Xiaohan Yang (Oak Ridge National Laboratory, USA). These expression datasets were filtered using the same criteria as S. album as stated. For phase enrichment analysis, Bonferroni adjusted p-value of each phase (0–23) was calculated using R program and phases with an adjusted p-value ≤ 0.05 was classified as enriched.
Gene Ontology (GO) terms of S. album protein sequences were annotated using InterProScan 5 [89]. GO terms for genes without InterProScan annotations were inferred using corresponding GO term from the top BLAST hit to Arabidopsis orthologs. The GO terms from both methods were merged and 33,362 genes had at least one annotated GO term. GO enrichment analysis was performed using the R package TopGO [90] with Fisher’s exact test and Bonferroni adjusted p-values.
OrthoFinder (v1.1.9) [91] was used to identify gene families between sequenced CAM and representative C3 and C4 species. The CAM species Sedum album, Phalaenopsis equestris, Dendrobium catenatum, Dendrobium officinale, Apostasia shenzhenica, Ananas comosus, and Kalanchoe fedtschenkoi, C3 species Brassica rapa, Oryza sativa, Solanum lycopersicum, Populus trichocarpa and Arabidopsis thaliana and C4 species Sorghum bicolor, Setaria italica were used for orthogroup identification. Protein sequences from each species were downloaded from Phytozome V12. Orthofinder was run using default parameters. Only orthologs from the CAM species pineapple and K. fedtschenkoi were included in downstream analyses as timecourse data for orchid (Phalaenopsis equestris) is currently unavailable.
To assess the phylogenetic relationships between orthologs, protein sequences of genes within the same orthogroup identified by OrthoFinder were extracted and aligned with MUSCLE (v.3.8.31)[92] and maximum likelihood phylogenetic trees were built using RAxML (v8.2.10)[93] with bootstraps of 100 and the–m PROTGAMMAJTT flag.
Orthogroups were further analyzed to identify expansions and contractions unique to S. album compared to other CAM as well as C3 and C4 species. The number of genes for each species was normalized by the total number of genes in the genome to account for differences in gene number and ploidy between species, as previously described [94]. Orthogroups were classified as contracted with a ratio of less than 0.2 in S. album compared to other species, and orthogroups with a ratio greater than 3 in S. album compared to other species were classified as expanded.
cis-element analysis was carried out as previously described [21]. The S. album promoters (500, 1000, 2000 bp) were parsed using the gene models. Gene lists were generated for each phase as called by JTK_CYCLE with a cycling significance cut-off of 0.05 and for specific conditions: CAM only cycling, C3 only cycling, C3 and CAM cycling with equal expression, C3 and CAM cycling with greater C3 expression, C3 and CAM cycling with greater CAM expression, Overrepresentation of cis-element was calculated for all of 3–8 mers using the ELEMENT program [95] with a p-value cut off of 0.05.
|
10.1371/journal.pcbi.1006020 | Deploying digital health data to optimize influenza surveillance at national and local scales | The surveillance of influenza activity is critical to early detection of epidemics and pandemics and the design of disease control strategies. Case reporting through a voluntary network of sentinel physicians is a commonly used method of passive surveillance for monitoring rates of influenza-like illness (ILI) worldwide. Despite its ubiquity, little attention has been given to the processes underlying the observation, collection, and spatial aggregation of sentinel surveillance data, and its subsequent effects on epidemiological understanding. We harnessed the high specificity of diagnosis codes in medical claims from a database that represented 2.5 billion visits from upwards of 120,000 United States healthcare providers each year. Among influenza seasons from 2002-2009 and the 2009 pandemic, we simulated limitations of sentinel surveillance systems such as low coverage and coarse spatial resolution, and performed Bayesian inference to probe the robustness of ecological inference and spatial prediction of disease burden. Our models suggest that a number of socio-environmental factors, in addition to local population interactions, state-specific health policies, as well as sampling effort may be responsible for the spatial patterns in U.S. sentinel ILI surveillance. In addition, we find that biases related to spatial aggregation were accentuated among areas with more heterogeneous disease risk, and sentinel systems designed with fixed reporting locations across seasons provided robust inference and prediction. With the growing availability of health-associated big data worldwide, our results suggest mechanisms for optimizing digital data streams to complement traditional surveillance in developed settings and enhance surveillance opportunities in developing countries.
| Influenza contributes substantially to global morbidity and mortality each year, and epidemiological surveillance for influenza is typically conducted by sentinel physicians and health care providers recruited to report cases of influenza-like illness. While population coverage and representativeness, and geographic distribution are considered during sentinel provider recruitment, systems cannot always achieve these standards due to the administrative burdens of data collection. We present spatial estimates of influenza disease burden across United States counties by leveraging the volume and fine spatial resolution of medical claims data, and existing socio-environmental hypotheses about the determinants of influenza disease disease burden. Using medical claims as a testbed, this study adds to literature on the optimization of surveillance system design by considering conditions of limited reporting and spatial aggregation. We highlight the importance of considering sampling biases and reporting locations when interpreting surveillance data, and suggest that local mobility and regional policies may be critical to understanding the spatial distribution of reported influenza-like illness.
| Seasonal influenza represents an important public health burden worldwide, and even within a single year, there is substantial variation in disease burden across populations [1–3]. On the other hand, pandemic influenza, which has the potential to cause millions of fatalities, is characterized by even more uncertainty in spatio-temporal risk. Traditional influenza surveillance is guided by the World Health Organization’s global standards for the collection of virological and epidemiological influenza surveillance data [4]. Epidemiological surveillance systems play an important role in our understanding of influenza dynamics and are used to identify seasonal influenza disease burden, severity, epidemic onset and seasonality, but they often suffer from reporting delays and limited, opportunistic sampling of the population.
Sentinel surveillance for influenza-like illness (ILI) is one such system that passively estimates influenza morbidity. Select general practitioners or health care facilities (“sentinels”) report aggregate counts of ILI to a centralized public health agency as an efficient means of collecting high quality data by focusing resources on a few population-representative sites [5]. The European Influenza Surveillance Network (EISN) collates sentinel ILI data from over 30 European countries, while the U.S. Centers for Disease Control and Prevention’s (CDC) ILINet surveillance system recruits roughly 2,000 sentinel physicians to submit reports on the percentage of patient visits with ILI weekly throughout the year [6–8]. While such sentinel surveillance systems are sufficient to provide situational awareness of national-level influenza activity, the coarseness of such data limit its use in local decision-making. Additionally, WHO recommends that choice in sentinel sites should consider population representativeness, geographical representation, patient volume, feasibility, and the data needs and goals of the surveillance system [4]. However, few ILI surveillance systems meet these criteria as they are limited by few incentives (e.g., data feedback from higher-level agencies, additional support for laboratory testing) and hampered by the administrative burden of data collection. Indeed, past studies have identified discrepancies across surveillance systems [9], and have investigated strategies to limit practitioner-based biases and improve capture of true population patterns in sentinel surveillance [10–12].
Medical claims represent an alternative potential source of passive ILI surveillance data with larger volume, fewer reporting delays, and finer spatio-temporal resolution than many traditional surveillance systems [13]. Additionally, medical claims data do not require additional administrative burden or voluntary reporting to a surveillance agency. We acknowledge that it may not be possible to combine these medical data streams directly into public health systems without further consideration of the ethical and privacy concerns of integrating health data at fine spatial resolutions [14]. In the meantime, however, we can leverage these features of medical claims and combine them with statistical models to explore the most informative design of passive surveillance systems and to test the robustness of ecological inference from opportunistic samples of health-associated big data.
We cannot, however, rely solely on the volume and resolution of big data to address surveillance data gaps; statistical models for ILI surveillance should also utilize information from known factors of spatial heterogeneity in influenza transmission and disease burden. Many studies have examined the relationship of environmental factors [15–21], transmission dynamics [22, 23], demography and contact patterns [24–32], immune landscapes [33, 34], and influenza type and subtype circulation [6, 29, 35–39] on influenza disease burden, although few have compared the relative importance of these mechanisms (except [40]). In addition, it is important to consider the possibility that individual patient behavior may bias the reporting of ILI disease burden, thus driving observed spatial heterogeneity. The association between poverty and social determinants [41–46], access to care, care-seeking behavior, and health insurance coverage [47–49], and reported ILI disease burden has been treated extensively elsewhere.
Surveillance system design may also contribute to the biased observation of ILI disease burden [11]. Current national sentinel systems in Australia, China, the United States, and Europe capture patients seen by 5% to less than 1% of active physicians in a given population, and while these systems strive to represent population demography, spatial distributions, and patient volume as accurately as possible, this is not always possible [4, 50, 51]. Theoretical work suggests influenza disease burden detection could be optimized if population coverage, care-seeking rates, and geographic access to care are considered in sentinel site choice [10, 11, 52, 53]. Non-traditional data with surveillance potential such as medical claims may enhance the estimation of attack rates through improved population coverage, better discriminate the duration of heightened epidemic activity and public health need through its real-time reporting, and improve our prediction of surge capacity needs with finer spatial resolution data. We note, however, that consideration of measurement biases is even more important as these digital data streams are opportunistic; they have greater volume and coverage in the population, but their measurement biases are less well studied [14]. Fortunately, non-traditional systems are often accompanied with metadata that provides context about the data coverage and user demographics, thus enabling explicit treatment of these potential flaws.
In this study, we developed a Bayesian hierarchical influenza surveillance model that accounts for transmission, environmental, influenza-specific, and socioeconomic factors, as well as measurement processes underlying spatial heterogeneity in reported influenza-like illness across counties in the United States. This model leveraged a large-scale and highly-resolved dataset of passive ILI surveillance from medical claims, and we validated the model results using ILI sentinel surveillance from CDC. Next, we probed the robustness of this ecological inference under limited data availability in order to mimic the potential conditions of real-world sentinel surveillance systems and to improve one primary goal of surveillance —the end-of-season estimation of disease burden. Our results highlight the relative contributions of surveillance data collection and socio-environmental processes to disease reporting, and emphasize the importance of considering surveillance system design and measurement biases when using surveillance data for epidemiological inference and prediction.
Using medical claims data representing 2.5 billion visits from upwards of 120,000 health care providers each year (see Methods: ‘Medical claims data’), we modeled influenza disease burden across U.S. counties for flu seasons from 2002-2003 through 2008-2009 and the 2009 pandemic using a hierarchical Bayesian modeling approach (see Methods: ‘Model structure’ and ‘Statistical analysis’). Our goal is to use this approach to simultaneously validate our surveillance data source and provide improved spatial surveillance of influenza burden based on socio-environmental and health behavior predictors. With these Bayesian models, we then study the impact of common surveillance limitations.
Our study considered six disease burden response variables: two measures of influenza disease burden (epidemic intensity and epidemic duration) in three populations (total population, children 5-19 years old, and adults 20-69 years old) across multiple seasons. We define epidemic intensity as a relative risk measure of population-normalized and detrended ILI activity above an epidemic baseline (details in Methods: ‘Defining influenza disease burden’). We define epidemic duration to be the number of weeks of ILI activity above an epidemic baseline. While total population models represent broad surveillance efforts to capture ILI activity in the community, the child and adult models may represent networks of school- or workplace-based surveillance systems. There were 13 county-level, 2 state-level and 4 HHS region-level predictors in the complete model (Table 1); all predictors were the same across response variables except care-seeking behavior, which was specific to the age group in the response (see Methods: ‘Predictor data collection and variable selection’). Analogous models considered influenza disease burden solely during the 2009 H1N1 pandemic. All model estimates of disease burden are openly available on GitHub at https://github.com/bansallab/optimize-flu-surveillance.
We validated our surveillance models of medical claims data in two ways. First we compared the model fits to CDC ILI and laboratory-confirmed surveillance data (details in Methods: ‘Model assessment and validation’). We then verified that significant socio-environmental factors identified by our models are consistent with past influenza studies.
The outputs of our statistical models provide improved surveillance of U.S. county-level disease burden due to influenza-like illness from 2002 to 2010. In this section, we explore broad temporal and spatial trends of seasonal ILI, the burden of seasonal ILI among children and adults, and the burden of ILI during the 2009 H1N1 pandemic.
Leveraging the large volume and spatial resolution of our data, we sought to examine the robustness of our model predictions and inference in order to assess their suitability for disease surveillance and prediction. First, we compared our estimation of epidemic intensity when using analogous models at the county and state spatial units of analysis. These comparisons recall hypothetical scenarios where inference from state-level surveillance data might inform county-level decision making in the absence of resolved county-level data. Next, two model sequences were designed to simulate different flu sentinel surveillance systems —fixed-location sentinels, where the same sentinel locations reported data every year, and moving-location sentinels, where new sentinel locations are recruited each year. A third model sequence considered the specificity of inference and model predictions to certain inclusion of historical data, thus providing insight into the generalization of our model to epidemic forecasting. We examine these applications for the total population epidemic intensity model, and ten replicates were performed for each model with missingness to generalize findings beyond that of random chance.
Reliable surveillance systems are at the heart of public health preparedness, mitigation and response. In this study, we opportunistically use an administrative data source to inform influenza spatio-temporal patterns and surveillance design. Our medical claims data represented an average of 24% of all U.S. health care visits to approximately 37% of all health care providers across 95% of U.S. counties during flu season months in our study period (increasing to 38%, 70% and 96%, respectively, by 2009). We pair these data with a Bayesian hierarchical modeling approach which enables “borrowing information”, the efficient incorporation of spatial dependence and group indicators for spatial and temporal random effects. The high resolution and coverage of our data combined with this spatial statistical approach allowed us to contribute to influenza surveillance in three ways: (a) enhance fine-grain mapping of disease burden from influenza-like illness to guide local influenza preparedness and control; (b) inform the future treatment of digital data streams as a measurement process for infectious disease surveillance; and (c) systematically explore surveillance design choices. Moreover, our surveillance model enables the generation of synthetic datasets that capture realistic spatial distributions of ILI, which can be used in models to inform the design of control strategies and surveillance systems.
In the process of our model validation, we also consider the relative importance of 16 environmental, demographic, or socio-economic factors in predicting influenza spatial heterogeneity. This makes ours the first large-scale influenza study to simultaneously consider multiple hypotheses across spatial scales (with the exception of work in review by Chattopadhyay et al [40]), and generates a new set of hypotheses on drivers of influenza spread. Our results strengthen the epidemiological link between humidity and influenza transmission and survival in temperate regions by finding strong negative associations between absolute humidity and both epidemic intensity and duration [54, 55]. These associations were not simply influenced by the strong spatial dependence of humidity —the relative effect of this predictor remained consistent when we removed the model’s spatial dependence term (Section 2.3 in S1 Appendix) and considered humidity as the sole model predictor (Fig AO in S1 Appendix). Charu et al. suggests that humidity may not provide additional information beyond a well-calibrated model of human mobility [56], and our work suggests that humidity among other factors is necessary to capture the end-of-season spatial heterogeneity in influenza disease burden. We also observed that higher estimated prior immunity was associated with greater epidemic intensity and longer epidemic durations. As larger epidemics induce more antigenic drift in subsequent seasons, we suggest that this drift renews population susceptibility every season, even on small spatial scales [57]. Finally, while higher vaccination coverage among toddlers was associated with lower epidemic intensity, we were surprised to note that higher vaccination coverage among elderly was associated with longer epidemics.
While statistical results may be neither interpreted as causative evidence nor are free from the possibility of spurious associations, future validations of our findings on influenza epidemiology will become more possible as high volume data sources achieve wider availability and tests of multiple hypotheses become more prevalent [40]. From the perspective of surveillance operations, we acknowledge the limitations of including many predictors with disparate data sources in our model; nevertheless, we gained additional epidemiological knowledge from the multiple predictor comparisons and note that all of the data we used were publicly available annually and at the county scale. In the future, comparisons of inference between models may enable us to posit new hypotheses for epidemiological study (e.g., vaccination of the elderly provides a protective effect among more susceptible and highly connected populations like children) (Fig AI in S1 Appendix).
Our model provides fine-scale, high coverage surveillance of ILI in the United States, allowing for a better understanding of influenza spatio-temporal patterns. Through an examination of significant group effects, we observed that South Atlantic states may experience longer and more acute seasons than other parts of the U.S during both seasonal and pandemic influenza scenarios and across ILI surveillance for children and adults. Our results also suggest that county-level spatial dependence and state effects explain a substantial part of the variation in epidemic intensity, while county-level spatial dependence and season effects best capture variation in epidemic duration. The explanatory power of county spatial dependence for surveillance models in both measures adds evidence to the importance of local mobility in the spatial spread and distribution of influenza disease burden [26, 56]. Moreover, we posit that state groupings explained variation in epidemic intensity because state-level policy recommendations and laws drive the probability for influenza infection and seeking of insured healthcare. For instance, influenza vaccination guidelines and access to free vaccinations are driven by local policy recommendations, and insurance policies are tied to state-level rules and regulations. Additional evidence for this hypothesis comes from our 2009 pandemic model where state effects also played a large role in explaining the variance in the data. On the other hand, variation in epidemic duration was better captured by season-level effects, and fixed effects that varied more between seasons than within them (e.g., influenza A/H3 and B circulation) were significant, similar to other studies [1]. We hypothesize that the duration of heightened ILI activity is more closely tied to population-level susceptibility and the identities of the predominantly circulating strains —factors that are likely to vary more across seasons than across space.
Our work uniquely captures factors of the measurement process, highlighting biases and disparities in healthcare-based influenza surveillance. We found that locations with greater poverty had lower influenza disease burden, in contrast to previous evidence for heightened rates of influenza-related hospitalizations, influenza-like illness, respiratory illness, neglected chronic diseases, and other measures of poor health among populations with greater material deprivation [43, 44, 47, 58–63]. Differences in socio-economic background may change recognition and therefore reporting of disease symptoms [46, 58]. Material deprivation and lack of social cohesion have also been implicated in lower rates of health care utilization for ILI, which would reduce the observation of influenza disease burden in our medical claims data among the poorest populations [44, 60]. When we artificially removed counties from our model (fixed-location sentinels) or subset our data into age groups, measurement factors associated with health care-seeking behavior more strongly explained the variation in epidemic intensity among the remaining observations (Fig 4, Fig AI in S1 Appendix). These two results together suggest that statistical inference from opportunistic data samples may avoid some types of reporting biases when the coverage or volume of data achieves a minimum threshold, in response to concerns posed in [14]. Increases to claims database coverage or care-seeking behavior may reduce reporting biases by increasing the representativeness of a given location’s sample, thus highlighting the importance of collecting and using metadata from opportunistic sources of epidemiological data.
Equipped with our model, we investigated the impact of surveillance system structure. We present the concept of a network of sentinel locations, in contrast to sentinel physicians or hospitals, which may be composed of administrative units (e.g., counties) that are chosen for either their representativeness of the larger population or their status as an outlier (e.g., match or failure to match locations in Fig 4, respectively). The ability for our model to estimate relatively accurate estimates of influenza burden across increasingly missing data suggests that routine sentinel surveillance in fixed locations may be more accurate for interpolating ILI disease burden among uncovered areas than surveillance across changing locations, even when fewer locations may be surveyed. Our framework enables sentinel counties to have flexible physician recruitment strategies, provided that county health departments can achieve target population coverage levels. Moreover, the improved performance of fixed-location surveillance systems is operationally ideal; as counties and physicians are retained as sentinels over long periods of time, we may expect the quality and consistency of reporting to improve. The accuracy of our surveillance model broke down at roughly 70% missingness among sentinels in fixed locations, which translates to fewer than 950 sentinel counties reporting data. While there are fewer sentinel counties than sentinel physicians in ILINet (approximately 2,000), we note that our county data represents aggregate reports from many healthcare providers. Indeed, the volume of visits captured by ILINet corresponded roughly to 5% of reporting counties in our medical claims data, and this level of missingness provided poor disease burden estimates for approximately 10-30% of counties in the best-case sentinel design (i.e., fixed-locations).
Our work contributes to our understanding of optimal population capture through surveillance by suggesting a framework that best maintains surveillance system design over multiple flu seasons [10–12]. Previous work acknowledges that spatial scales of aggregation alter statistical inference and statistically-identified drivers of disease distributions [64, 65]. Our aggregated state surveillance models adequately captured the high epidemic intensity risk among counties in the South Atlantic, similar to other studies of spatial scale [66], but they over-estimated epidemic intensity among low-risk states, thus suggesting that these types of surveillance models may be useful for public health preparedness but less optimal for the allocation of limited resources. Nevertheless, we observed that larger discrepancies between state- and county-level surveillance models were associated with greater within-state heterogeneity in disease burden, suggesting perhaps that the spatial aggregation of data may have minimal effects on epidemiological inference and policy-making if populations and socio-environmental determinants are relatively homogeneous within a given spatial unit (Fig Q in S1 Appendix). Overall, state surveillance models seemed more prone to over-estimate than under-estimate county-level disease burden, suggesting that inference from state surveillance data is best limited to populous counties in a given state (Fig P in S1 Appendix). Future work is needed to better understand surveillance-associated aggregation biases in order to expand the utility of aggregate scale surveillance data in local contexts.
Given the growing availability of health-associated big data in infectious disease surveillance [13, 67], we emphasize the importance of collecting relevant metadata on system coverage and reporting, while considering the ethical and privacy implications of using these data at fine spatial resolutions [14]. In the future, statistical surveillance modeling may become standard methodology to inform the choice of sentinel locations with non-traditional high-volume digital health data, improve the long-term design of disease surveillance systems, and enhance the development of syndromic surveillance in developing countries [68].
Weekly visits for influenza-like illness (ILI) and any diagnosis from October 2002 to April 2010 were obtained from a records-level database of US medical claims managed by IMS Health and aggregated to three-digit patient US zipcode prefixes (zip3s), where ILI was defined with International Classification of Diseases, Ninth Revision (ICD-9) codes for: direct mention of influenza, fever combined with respiratory symptoms or febrile viral illness, or prescription of oseltamivir. Medical claims have been demonstrated to capture respiratory infections accurately and in near real-time [69, 70], and our specific dataset was validated to independent ILI surveillance data at multiple spatial scales and age groups and captures spatial dynamics of influenza spread in seasonal and pandemic scenarios [56, 71, 72]. Please see Section 1 in S1 Appendix for a statement on ethics and data access.
We also obtained database metadata from IMS Health on the percentage of reporting physicians and the estimated effective physician coverage by visit volume; these data were used to generate “measurement” predictors (Table 1). ILI reports and measurement factors at the zip3-level were redistributed to the county-level according to population weights derived from the 2010 US Census ZIP Code Tabulation Area (ZCTA) to county relationship file, assuming that ZCTAs that shared the first three digits belonged to the same zip3. These metadata indicated that our medical claims database represented roughly 24% of visits for any diagnosis from approximately 37% of all health care providers across 95% of U.S. counties during influenza season months, averaged over the years in our study period.
We performed the following data processing steps for each county-level time series of ILI per population (Section 7 in S1 Appendix): i) Fit a LOESS curve to non-flu period weeks (flu period defined as November through March each year) to capture moderate-scale time trends (span = 0.4, degree = 2); ii) Subtract LOESS predictions from original data to detrend the entire time series; iii) Fit a linear regression model with annual harmonic terms and a time trend to non-flu period weeks [16]; iv) Counties were defined to have an “epidemic” in a given flu season if at least two consecutive weeks of detrended ILI observations exceeded the ILI epidemic threshold during the flu period (i.e., epidemic period) [73]. The epidemic period was the maximum length consecutive period where detrended ILI exceeded the epidemic threshold during the flu period. The epidemic threshold was the upper bound of the 95% confidence interval for the linear model prediction. Counties with a greater number of consecutive weeks above the epidemic threshold during the non-flu period than during the flu period were removed from the analysis; v) Disease burden metrics were calculated for counties with epidemics.
Two measures of influenza disease burden were defined for each county. For a given season and county: We define attack rate as the sum of population-normalized and detrended ILI during the epidemic period found above (and shifted by one to accommodate the likelihood distribution). Our epidemic intensity measure is defined as the standardized ratio of this attack rate and the expected attack rate. The expected attack rate is calculated as the population-weighted mean of the observed attack rates, and is a model offset described under ‘Model structure’. Epidemic duration was defined as the number of weeks in the epidemic period and counties without epidemics were assigned the value zero.
Models and data were processed separately for the 2009 H1N1 pandemic season and for state-level epidemic intensity (details in Sections 2.5 and S2.6 in S1 Appendix respectively).
Quantifiable proxies were identified for each hypothesis found in the literature, and these mechanistic predictors were collected from probability-sampled or gridded, publicly available sources and collected or aggregated to the smallest available spatial unit among US counties, states, and Department of Health and Human Services (HHS) regions for each year or flu season in the study period, as appropriate (Table 1, Section 6 in S1 Appendix).
We selected one predictor to represent each hypothesis according to the following criteria, in order: i) Select for the finest spatial resolution; ii) Select for the greatest temporal coverage for years in the study period; iii) Select for limited multicollinearity with predictors representing the other hypotheses, as indicated by the magnitude of Spearman rank cross-correlation coefficients between predictor pairs. We also compared the results of single predictor models and our final multi-predictor models as another check of multicollinearity (Section 6 in S1 Appendix). For the modeling analysis, if a predictor had missing data at all locations for an entire year, data from the subsequent or closest other survey year were replicated to fill in that year. If a predictor data source was available only at the state or region-level, all inclusive counties were assigned the corresponding state or region-level predictor value (e.g., assign estimated percentage of flu vaccination coverage for state of California to all counties in California). Predictors were centered and standardized prior to all exploratory analyses and modeling, as appropriate. Interaction terms comprised the product of their component centered and standardized predictors. Data cleaning and exploratory data analysis were conducted primarily in R [74]. Final model predictors are described below, and our hypotheses for each predictor are described in Table 1. All cleaned predictor data are available upon request.
We present the most common version of our model structure here. The generic model for county-year observations (for i counties and t years) of influenza disease burden yt is:
y t | μ t , τ ∼ f ( y t | μ t , τ ) (1)
where yt = (y1t, …, ynt)′ denotes the vector of i = 1, …, n county observations across t = 1, …, T years included in the model (Eq 1). We modeled the mean of the observed disease burden magnitude (μt), where f(yt|μt, τ) is the distribution of the likelihood of the disease burden data, parameterized with mean μt = (μ1t, …, μnt)′ and precision τ (where precision is the inverse of variance), as appropriate to the likelihood distribution.
The proposed determinants of disease burden were modeled as:
g ( μ i t ) = E i t + α + ∑ p = 1 m X i t p β p + γ i + ζ j [ i ] + η k [ i ] + ν t + ϕ i + ϵ i t (2)
where g(.) is the link function, α is the intercept, there are m socio-environmental and measurement predictors (i.e., Xt1, …, Xtm), where Xt1 = (X1t1, …, Xnt1)′, and Eit is an offset of the expected disease burden, such that Eq 2 models the relative risk of disease (μit/Eit) in county i, common in disease mapping [78–80]. Group terms at the county, state j, region k, and season t levels (γi, ζj[i], ηk[i], νt, respectively) and the error term (ϵit) are independent and identically distributed (iid).
Geographical proximity appears to increase the synchrony of flu epidemic timing [81, 82], while connectivity between cities has been linked with spatial spread in the context of commuting and longer distance travel [83–86]. We modeled county spatial dependence ϕi with an intrinsic conditional autoregressive (ICAR) model, which smooths model predictions by borrowing information from neighbors [87]:
ϕ i | ϕ j , - i , τ ϕ ∼ Normal ( 1 ξ i ∑ i ∼ j ϕ j , 1 ξ i τ ϕ ) , (3)
where ξi represents the number of neighbors for node i, ϕj,−i represents the neighborhood of node i, which is composed of neighboring nodes j (neighbors denoted i ∼ j). The precision parameter is τϕ (Eq 3).
The goals of our modeling approach were to i) estimate the contribution of each predictor to influenza disease burden, ii) predict disease burden in locations with missing data, and iii) improve mapping of influenza disease burden. We performed approximate Bayesian inference using Integrated Nested Laplace Approximations (INLA) with the R-INLA package (www.r-inla.org) [88, 89]. INLA has demonstrated computational efficiency for latent Gaussian models, produced similar estimates for fixed parameters as established implementations of Markov Chain Monte Carlo (MCMC) methods for Bayesian inference, and been applied to disease mapping and spatial ecology questions [90–94].
Log epidemic intensity was modeled with a normal distribution, and log epidemic duration was modeled with a normal distribution without the offset term in Eq 2. Consequently, we note that all epidemic intensity models examine the relative risk of disease burden, while epidemic duration models examine the duration in weeks. Multi-season models included all terms in Eq 2. Model coefficients were interpreted as statistically significant if the 95% credible interval for a parameter’s posterior distribution failed to include zero.
To assess model fit, we examined scatterplots and Pearson’s cross-correlation coefficients between observed and fitted values for the epidemic intensity and epidemic duration total population surveillance models. The epidemic intensity model fit the data well and the Pearson’s cross-correlation coefficient between the observed and fitted mean relative epidemic intensity was R = 0.86 (Section 2 in S1 Appendix). The epidemic duration model fit relatively well, and the Pearson’s cross-correlation coefficient between the observed and predicted mean number of epidemic weeks was R = 0.94 (Section 4 in S1 Appendix).
We also examined scatterplots of standardized residuals and fitted values; standardized residuals were defined as ( y - μ y ^ ) / σ y ^, where μ y ^ is the fitted value posterior mean and σ y ^ is the fitted value standard deviation. Residual plots for the epidemic intensity and duration models may be found in Sections 2 and 4 in S1 Appendix, respectively.
For each disease burden measure, we compared models with no spatial dependence, county-level dependence only, state-level dependence only, and both county and state-level dependence. The goal of the county-level dependence was to capture local population flows, while state-level dependence attempted to capture state-level flight passenger flows (details in Section 2 in S1 Appendix). We determined that models with only county-level spatial neighborhood structure best fit the data after examining the Deviance Information Criteria (DIC) values and spatial dependence coefficients of the four model structures, further supporting evidence in [56]. County-level spatial structure was subsequently used in all final model combinations. We report results from models with county-level dependence only.
We assessed the contribution of each set of group effects (i.e., season, region, state, county, county spatial dependence, observation error) to model fit by comparing the mean precision estimates for the terms, where precision is the inverse of variance. Effects with a smaller precision captured a greater magnitude of variability in the data.
We examined the added value of county-level information relative to state-level information by comparing the aggregation bias between county and state surveillance models. Here, we defined aggregation bias as the difference between fitted log epidemic intensity from state and county surveillance models. Positive values mean that the state model overestimates risk relative to the county model, and vice versa.
For model validation, we compared model fitted values for epidemic intensity with CDC ILI and laboratory surveillance data, which are derived from approximately 2,000 ILI-reporting sentinel physicians and 100,000-200,000 respiratory specimens annually (details in Section 2 in S1 Appendix). We assessed model robustness through additional cross-validation and out-of-sample validation analyses; the total population epidemic intensity model was refit where 20%, 40%, 60%, 80%, 90%, 95%, and 97.5% of all county observations were randomly replaced with NAs (sentinels in fixed locations), and where 20%, 40%, 60% and 80% of model observations were stratified by season and randomly replaced with NAs (sentinels in moving locations). We also refit three models where one, three, and five of seven flu seasons were randomly chosen and completely replaced with NAs (inclusion of historical data). To account for variability due to random chance, models were replicated ten times each with different random seeds. For each sequence of missingness, we performed out-of-sample validation by comparing the mean fitted values to the true observed values for all data that were randomly removed across seasons and replicates (Section 3.4 in S1 Appendix). We then compared the magnitude and significance of socio-environmental and measurement drivers, and the posterior distributions of county-season fitted values. Fitted value distributions were noted as significantly different (i.e., values did not match) if the interquartile ranges for two fitted values failed to overlap with each other (Section 3.2 in S1 Appendix). The results described in “Sentinel surveillance design” use methods identical to this analysis and may be interpreted additionally as model sensitivity and robustness.
Model estimates of disease burden, summary statistics for predictors, and their associated model codes are openly available on GitHub at https://github.com/bansallab/optimize-flu-surveillance. All processed predictor data are available upon request.
|
10.1371/journal.pgen.1006361 | The Skp1 Homologs SKR-1/2 Are Required for the Caenorhabditis elegans SKN-1 Antioxidant/Detoxification Response Independently of p38 MAPK | SKN-1/Nrf are the primary antioxidant/detoxification response transcription factors in animals and they promote health and longevity in many contexts. SKN-1/Nrf are activated by a remarkably broad-range of natural and synthetic compounds and physiological conditions. Defining the signaling mechanisms that regulate SKN-1/Nrf activation provides insights into how cells coordinate responses to stress. Nrf2 in mammals is regulated in part by the redox sensor repressor protein named Keap1. In C. elegans, the p38 MAPK cascade in the intestine activates SKN-1 during oxidative stress by promoting its nuclear accumulation. Interestingly, we find variation in the kinetics of p38 MAPK activation and tissues with SKN-1 nuclear accumulation among different pro-oxidants that all trigger strong induction of SKN-1 target genes. Using genome-wide RNAi screening, we identify new genes that are required for activation of the core SKN-1 target gene gst-4 during exposure to the natural pro-oxidant juglone. Among 10 putative activators identified in this screen was skr-1/2, highly conserved homologs of yeast and mammalian Skp1, which function to assemble protein complexes. Silencing of skr-1/2 inhibits induction of SKN-1 dependent detoxification genes and reduces resistance to pro-oxidants without decreasing p38 MAPK activation. Global transcriptomics revealed strong correlation between genes that are regulated by SKR-1/2 and SKN-1 indicating a high degree of specificity. We also show that SKR-1/2 functions upstream of the WD40 repeat protein WDR-23, which binds to and inhibits SKN-1. Together, these results identify a novel p38 MAPK independent signaling mechanism that activates SKN-1 via SKR-1/2 and involves WDR-23.
| Oxidative stress is the result of imbalanced control of reactive oxidative species in cells, is a common occurrence during aerobic metabolism, and must be managed to limit cellular damage and disease. Many details about the signaling mechanisms utilized by animal cells in response to pro-oxidants remain to be discovered. We provide evidence that the signaling mechanisms that activate SKN-1, a master regulator of detoxification genes and aging in the model organism C. elegans, may vary in response to different oxidative stressors. From a genome-wide genetic screen, we identify highly conserved genes skr-1/2 as central regulators of the gene response to pro-oxidants. During exposure to oxidants, SKR-1/2 function upstream from SKN-1 and a direct SKN-1 repressor named WDR-23. These results provide new insights into our understanding of SKN-1 regulation and lay the foundation for future studies to define in detail novel signaling pathways that respond to pro-oxidants.
| Reactive small molecules are common in natural environments and are produced as byproducts of oxygen metabolism. Reactive small molecules in excess can cause oxidative damage with widespread detrimental effects, but also function as signaling molecules for normal physiological processes [1]. Appropriate response to and regulation of these compounds is crucial as aberrant accumulation has been implicated in early onset of aging along with many pathological states that include metabolic syndromes, neurological disorders, and cancer [2,3,4]. In C. elegans, the cap ‘n’ collar transcription factor family member SKN-1 is homologous to mammalian Nrf2 and functions to promote longevity and resistance to a wide range of environmental stressors [5].
In response to a wide-range of reactive small molecules, SKN-1/Nrf transcription factors translocate into the nucleus and bind to response elements in target genes to activate a conserved detoxification response [6,7,8]. In mammals, Keap1 represses basal Nrf2 activity through a direct interaction that promotes ubiquitylation and degradation, and there is strong support for a model in which small molecules directly react with Keap1 releasing Nrf2 from repression [6,7,8]. Additional Keap1-independent signaling mechanisms exist that are less-defined [9].
Genetic tractability and the conserved nature of the SKN-1/Nrf response have made C. elegans an important model for regulation of this pathway [5]. C. elegans has also been instrumental for defining SKN-1 as a central determinant of aging and longevity [10,11,12] and is being used to understand the role of SKN-1 in antiparasitic drug resistance [13,14,15]. Although C. elegans lacks a close Keap1 homolog, it is repressed under basal conditions by an analogous mechanism via the WD40 repeat protein WDR-23, which binds to and inhibits SKN-1 [15,16]. The protein kinases AKT-1/2, SGK-1, and GSK-3 also function to inhibit SKN-1 under basal conditions [11,17]. A number of protein kinases have been implicated in activation of SKN-1 (MKK-4, IKKɛ-1, NEKL-2, and PHDK-2), although it is not known if any of these regulate SKN-1 directly [18]. During oxidative stress, the p38 MAPK signaling cascade directly phosphorylates and promotes nuclear accumulation of SKN-1 in cells of the intestine [19], a tissue thought to be a primary site for detoxification; p38 MAPK is also required for activation of SKN-1 in the intestine during infection [20,21]. A recent study demonstrated that TIR-1, Toll/interleukin-1 receptor domain protein, functions upstream from p38 MAPK during exposure to an oxidant [22]. Although protein kinases, particularly p38 MAPK, are clearly important, it is not known if this one mechanism is responsible for activation of SKN-1 by all the diverse reactive small molecules known to strongly activate the pathway.
We demonstrate here that the kinetics of p38 MAPK activation and tissues with SKN-1::GFP accumulation vary with different pro-oxidants that all elicit a strong SKN-1 dependent detoxification response. Using genome-wide RNAi screening, we identified SKR-1/2 as required for the core SKN-1 transcriptional response to diverse pro-oxidant compounds. SKR-1/2 are highly conserved orthologues of Skp1, a component of many protein complexes including the Skp-Cullin-F box ubiquitin ligase (SCF) that regulates cell cycle progression and differentiation [23,24]. Loss of skr-1/2 strongly and specifically attenuates induction of SKN-1 dependent genes independent of p38 MAPK signaling and reduces survival of pro-oxidants. SKR-1/2 functions upstream of WDR-23 and influences the accumulation of a WDR-23::GFP fusion protein in nuclei. We hypothesize that this newly identified pathway regulates SKN-1 activity by modulating WDR-23 function.
SKN-1 activation has been shown to be mediated through direct phosphorylation by PMK-1 (p38 MAPK) in response to oxidative stress induced by arsenite and during pathogen infection (Pseudomonas aeruginosa) [19,20,21]. SKN-1 residues phosphorylated by PMK-1 are required for nuclear accumulation, a step sometimes correlated with induction of SKN-1 dependent detoxification genes. Arsenite, paraquat, juglone, and acrylamide are a diverse set of small molecules that all strongly activate SKN-1 dependent detoxification genes [16,19,25]. To test if PMK-1 responded similarly to these different compounds, we measured phosphorylation of PMK-1 at its kinase activation residue (Y180/182) over a four hour time course following exposure to arsenite, juglone, and acrylamide (Fig 1). Arsenite increased PMK-1 phosphorylation levels strongly at all time points, and juglone and acrylamide caused smaller transient increases during short term exposure (5–60 min). Interestingly, PMK-1 phosphorylation levels decreased with acrylamide after 3 and 4 h (Fig 1).
We also measured phosphorylation of PMK-1 with replicates to confirm effects at specific time points and to test paraquat; doses and durations used were all sub-lethal (S1 Fig) yet strongly activate SKN-1 dependent genes (5 mM arsenite for 1 h, 35 mM paraquat for 2 h, 38 μM juglone for 3 h, and 7 mM acrylamide for 4 h). An increase in phosphorylation of PMK-1 was observed with arsenite and paraquat and a decrease was confirmed for 4 h of acrylamide exposure (Fig 2A). Total levels of PMK-1 were not altered with any treatment or duration (Figs 1 and 2A), indicating that changes to p-PMK-1 were at the posttranslational level. Given that we used whole animal lysates, these results may not reflect PMK-1 phosphorylation kinetics in all tissues. However, our results do demonstrate that the overall patterns of PMK-1 phosphorylation can vary between conditions that all strongly activate SKN-1.
We next used real-time RT-PCR (qPCR) in deletion mutants of pmk-1 with and without skn-1(RNAi) to assess the requirement of the p38 MAPK pathway after exposure to 5 mM arsenite or 38 μM juglone for 1 h. As expected, five detoxification genes directly regulated by SKN-1 were strongly activated by both compounds in N2 wild-type worms and this was partially dependent on pmk-1 and largely dependent on skn-1 (Fig 2B). Although all four gst detoxification genes were partially dependent on pmk-1, they were still induced. Given that the pmk-1(km25) allele we used is a deletion considered to be null, these results suggest that there are mechanisms that can compensate for loss of PMK-1 in these contexts (e.g., PMK-1 paralogs or other pathways).
Arsenite induces strong nuclear accumulation of SKN-1b/c::GFP fusion proteins in the intestine, which is dependent on the p38 MAPK pathway [19]. Although SKN-1 protein accumulation in nuclei is thought to be one mechanism of pathway activation [19], genetic and environmental conditions have been identified that activate SKN-1 target genes without causing detectable SKN-1b/c::GFP accumulation implicating other yet-to-be defined regulatory mechanisms [26,27,28]. We scored the percentage of worms with three levels of intestinal SKN-1b/c::GFP nuclear accumulation with arsenite, azide, paraquat, juglone, and acrylamide exposure to determine if SKN-1 accumulation varies with SKN-1 inducer. The LD001 strain we used expresses fusion proteins of SKN-1c and b variants, but not the longer SKN-1a variant [6]. Consistent with previous reports, we found that 5 mM arsenite, 5 mM azide, and 35 mM paraquat induced high levels of nuclear SKN-1b/c::GFP localization in the intestine (Fig 3A). Conversely, nuclear SKN-1b/c::GFP was not detected in the intestine with short (5–15 min, S2 Fig) or long-term (up to 5 h of 38 μM juglone or 24 h of 7 mM acrylamide, Fig 3A) exposure to juglone or acrylamide even though these treatments strongly activate SKN-1 dependent detoxification genes in the same tissue [16,25] (S3 Fig).
Scoring of stress-inducible SKN-1b/c::GFP localization is typically limited to the intestine, which is credited with being the primary site of detoxification and has the largest nuclei that are readily visible. However, we and others [25] observe strong induction of SKN-1 dependent detoxification gene reporters in other tissues, particularly the hypodermis (S3 Fig). Careful observation revealed SKN-1b/c::GFP accumulation in nuclei throughout worms exposed to arsenite (Fig 3B and 3C and S4A Fig); these other nuclei are most obvious in the head and tail regions, which have less autofluorescence than areas around the intestine. No accumulation was observed in intestine, head, or tail regions of worms fed skn-1 dsRNA (S4A Fig) verifying specificity of the signal. We also scored nuclear accumulation with a transgene that covers all forms of SKN-1 including the long SKN-1a variant, skn-1op::GFP [11]. As shown in S4B Fig, SKN-1op::GFP generally responded similar to SKN-1b/c::GFP with no intestinal accumulation with juglone.
The head nuclei SKN-1b/c::GFP signals were also consistently induced by azide, paraquat, and juglone (Fig 3B and 3C); head nuclei SKN-1b/c::GFP signals were also induced by acrylamide, but this took several hours. Although identification of all tissues with SKN-1b/c::GFP is difficult, at least some of the signal appears to be in hypodermal cell nuclei based on location and morphology (S4C Fig) when compared to images of hypodermal specific markers [29]. The pmk-1(km25) allele was crossed into the SKN-1b/c::GFP strain to test the requirement of p38 MAPK for the accumulation of hypodermal SKN-1. As expected, SKN-1b/c::GFP failed to accumulate in the intestine of pmk-1 worms under arsenite exposure (Fig 4). Alternatively, loss of pmk-1 only slightly reduced the number of worms with SKN-1b/c::GFP accumulation in the head region (Fig 4).
To determine the physiological role of SKN-1 in the hypodermis and intestine, we next measured the effects of hypodermis and intestine-specific skn-1(RNAi) on juglone survival. As shown in S5 Fig, loss of skn-1 from either tissue reduced survival consistent with skn-1 mediating resistance in both tissues even though SKN-1::GFP accumulation is not obvious in the intestine during exposure to juglone. Therefore, juglone and acrylamide are able to activate a robust SKN-1 dependent detoxification gene response without detectable SKN-1::GFP nuclear accumulation in the intestine, and SKN-1b/c::GFP accumulation during stress can be partially decoupled from pmk-1 in tissues other than the intestine.
To identify new regulators of the SKN-1 detoxification response, we utilized a C. elegans strain harboring an integrated Pgst-4::GFP transcriptional reporter in a genome-wide RNAi screen for genes that regulate gst-4 expression during exposure to juglone. gst-4 is a phase 2 detoxification gene regulated directly by SKN-1 under numerous conditions, and it has been shown to be a reliable reporter for SKN-1 transcriptional activity [25,27,30,31]. We screened approximately 19,000 dsRNA clones and identified 10 genes that when silenced consistently reduced juglone-induced Pgst-4::GFP fluorescence. RNAi of seven of these genes caused developmental defects or general sickness. After retesting these by initiating silencing at the L4 larval stage rather than L1 to bypass developmental requirements, there were a total of six RNAi clones that strongly reduced Pgst-4::GFP fluorescence. These were skr-1/2, C01B10.3, pad-1, mdt-15, ifb-1, and uba-1 (Fig 5A); mdt-15 has previously been reported to function with SKN-1 [32]. We tested the candidate novel regulators with two other stressors to assess specificity. None of the novel gst-4 regulators had a significant effect on induction of a heat shock reporter, Phsp-16.2::GFP (S6A Fig), but ifb-1, uba-1, etf-1, and D1081.8 were required for induction of an osmotic stress reporter, Pgpdh-1::RFP (S6B Fig).
We next tested the candidate genes for genetic interactions with wdr-23 and skn-1 using strains carrying the Pgst-4::GFP reporter and a wdr-23(tm1817) loss of function allele or a skn-1(k1023) gain of function allele [10]. Both strains have constitutive activation of Pgst-4::GFP that is suppressed by skn-1(RNAi). dsRNA targeting the candidates had little or no effect on Pgst-4::GFP with the exception of ifb-1 and uba-1 in wdr-23(tm1817) (Fig 5B); although not confirmed in these experiments, skr-1/2(RNAi) consistently causes a partial embryonic lethality phenotype. These findings suggest that skr-1/2, C01B10.3, and pad-1 likely function upstream of skn-1 and wdr-23.
skr-1/2(RNAi) (Skp-related) caused the strongest inhibition of Pgst-4::GFP with juglone, and most closely resembled the pattern observed for skn-1(RNAi) (Fig 5A). SKR-1/2 was previously shown to be required for longevity extension in daf-2 insulin and IGF-1-like receptor mutants [33], but has not previously been reported to function as a regulator of stress responses. Because the neighboring skr-1 and skr-2 genes are recent duplications that are 83% identical at the nucleotide level, skr-1 dsRNA likely silences both skr-1 and skr-2 [24]. Consistent with this, skr-1 dsRNA reduces both skr-1 and skr-2 mRNA (S1 Table) and skr-2 dsRNA has the same effect on Pgst-4::GFP as skr-1 dsRNA (S7A Fig). We therefore refer to ‘skr-1/2(RNAi)’ when using skr-1 dsRNA.
C. elegans has a total of 21 SKR proteins that are orthologous to the single yeast and mammalian Skp1 proteins, which are one of the four core components of the highly conserved SCF ubiquitin-ligase complex. Skp1 interacts directly with cullin and F-box proteins, with the latter functioning to selectively recruit substrate targets for ubiquitin ligation [23]. To address whether other skr genes or components of the SCF complex are also required for gst-4 activation in response to juglone, we performed an RNAi screen against a sub-library of available skr dsRNA clones (skr-3, 5, 7, 8, 9, 10, 11, 12, 13, 15, 17, 18, 19, 20, and 21) and cul-1, which is the only cullin that is known to interact with SKR-1 and 2 [23,24]. Unlike skr-1/2, none of the other skr clones or cul-1 had a significant effect on Pgst-4::GFP induction after juglone exposure (S7B Fig). Due to a general sickness and larval arrest of cul-1(e1756) null mutants [34], we were unable to confirm the cul-1(RNAi) results using a mutant. To enhance RNAi, we also tested eri-1 RNAi hypersensitive worms fed cul-1 dsRNA for two generations. Embryonic lethality (a well-established cul-1 phenotype [35]) was observed in second generation worms fed cul-1 dsRNA, but there was still no effect on Pgst-4::GFP when induced with juglone in second generation worms that were able to develop to the L4 stage (S7C and S7D Fig).
To investigate the role of skr-1/2 in response to reactive compounds, we tested its requirement for Pgst-4::GFP induction during exposure to arsenite, paraquat, juglone, and acrylamide. skr-1/2(RNAi) strongly inhibited induction of Pgst-4::GFP with juglone, paraquat, and acrylamide and had a smaller effect with arsenite (Fig 6A); a role for skr-1/2 in Pgst-4::GFP induction by arsenite was confirmed in a more sensitive plate reading assay at four doses (Fig 6B). The requirement for skr-1/2 was also tested with qPCR of four gst mRNAs directly controlled by SKN-1 under control conditions and after exposure to 38 μM juglone for 3 h (Fig 6C). Loss of skr-1/2 only reduced expression of one of the gst genes under control conditions. Juglone activated all four gst mRNAs (P<0.001, Fig 6C), and as expected, skn-1(RNAi) strongly reduced induction of all four gst genes in juglone. Silencing of skr-1/2 also strongly inhibited induction of all four gst genes by juglone by 57–74%. Therefore, skr-1/2 is partially required for induction of core SKN-1 dependent detoxification genes by diverse reactive compounds.
To obtain a global view of gene regulation by skr-1/2, we extracted RNA from worms grown on control, skr-1/2, and skn-1(RNAi) after exposure to juglone and identified differentially expressed genes (DEG) using unbiased whole-genome RNA sequencing (RNA-seq). The FOXO transcription factor daf-16 is also widely involved in detoxification responses [36] and a previous study implicated skr-1/2 in regulation of a DAF-16 dependent gene and in promoting longevity in daf-2 mutant worms, which have elevated DAF-16 activity [33]. Therefore, we also included daf-16(RNAi) to evaluate the specificity of gene regulation by skr-1/2. We identified 309 (174 up, 135 down) DEG by juglone, 273 (142 up, 131 down) DEG by skr-1/2(RNAi), 1241 (517 up, 737 down) DEG by skn-1(RNAi), and 562 (281 up, 281 down) DEG by daf-16(RNAi) (Fig 7A and 7B, S8A and S8B Fig and S1 Table).
A heat map of all genes up-regulated by juglone is shown in Fig 7A demonstrating that many of the strongest up-regulated genes were dependent on both skn-1 and skr-1/2. Correlation coefficients are shown below the heat map for all comparisons demonstrating a much higher correlation of skr-1/2 with skn-1 (0.763) than with daf-16 (0.235) for the genes up-regulated by juglone. Of the 174 genes up-regulated by juglone, 78 were skn-1 dependent and 35 were skr-1/2 dependent (Fig 7B). Of the 35 skr-1/2 dependent juglone induced genes, almost all (32, or 91%) were also skn-1 dependent. Furthermore, 110 of the 131 (84%) total skr-1/2 dependent genes were also skn-1 dependent; 14.7% of all skn-1 dependent genes were skr-1/2 dependent. Compared to skn-1(RNAi), daf-16(RNAi) affected fewer overall genes but had a similar proportional overlap with skr-1/2 (13.5%, Fig 7B). Using DAVID analysis, skr-1/2 dependent genes were primarily enriched for glutathione mediated detoxification and metabolism, glucosyltransferase activities, and collagen/cuticle development (Fig 7C), while the genes up-regulated by skr-1/2(RNAi) were enriched for regulation of growth rate, collagen, and metal ion binding (S8C Fig). Genes up-regulated by skr-1/2(RNAi) had little overlap with genes influenced by juglone (S8B Fig).
To gain deeper insights into the gene expression effects of skr-1/2 relative to skn-1 and daf-16, we performed linear regression analysis on all genes that were down-regulated by skn-1(RNAi) or daf-16(RNAi), regardless of how they were affected by juglone, by plotting fold change with skr-1/2(RNAi) versus fold change by skn-1 or daf-16(RNAi) (Fig 7D). Fold-changes for skr-1/2(RNAi) were significantly correlated with fold-changes for all genes significantly downregulated by skn-1(RNAi) (slope = 0.265±0.02, R2 = 0.24, P < 0.0001) (Fig 7D). No correlation was observed between skr-1/2 and daf-16(RNAi) for all genes significantly downregulated by daf-16(RNAi) (slope = 0.08±0.05, R2 = 0.01, P = 0.07) (Fig 7D). A significant, but very small, correlation was found between skr-1/2 and skn-1(RNAi) for genes up-regulated by skn-1(RNAi), not for genes up-regulated by daf-16(RNAi) (S8D Fig). Taken together, these results reveal that skr-1/2 is required for expression of a subset of detoxification and extracellular matrix genes that are largely (84%) nested within a set of skn-1 dependent genes and not correlated with daf-16 dependent genes.
We next examined the requirement of skr-1/2 for juglone and arsenite resistance. In our hands, skn-1(RNAi) often does not have a reproducible effect on survival of juglone in worms that have not previously been exposed suggesting that basal activity may not play a large role in survival of an acute lethal juglone dose. We then pre-conditioned worms to a low and hormetic level of juglone (38 μM for 2 h), which strongly activates SKN-1 dependent genes (Fig 2B) and dramatically increases resistance [37], before measuring survival in a much higher lethal dose (125 μM). In these experiments, skn-1(RNAi) significantly decreased juglone survival compared to control worms in all three trials and skr-1/2(RNAi) significantly decreased survival in two of three trials (Fig 8A and S9 Fig). In 10 mM arsenite, either skn-1(RNAi) or skr-1/2(RNAi) significantly decreased survival compared to the control worms (Fig 8B and S9 Fig). These results show that skr-1/2 is required for oxidative stress resistance.
We next tested the effects of skr-1/2(RNAi) on juglone survival in two strains with greatly increased SKN-1 activity, wdr-23(tm1817) loss of function and skn-1(k1023) gain of function [10]. As expected, both of these strains were highly resistant over a range of juglone concentrations compared to N2, and skr-1/2(RNAi) did not have a reproducible effect on juglone resistance in N2 worms not previously exposed to juglone (Fig 8C). As expected given that skr-1/2 appears to regulate gst-4 upstream from wdr-23 (Fig 5B), skr-1/2(RNAi) did not reduce resistance of either wdr-23(tm1817) or skn-1(k1023) worms (Fig 8C).
Having established skr-1/2 as a requirement for SKN-1 dependent detoxification responses, we conducted additional experiments to gain insight into potential mechanisms. We demonstrated above that skr-1/2 likely regulates detoxification gene expression upstream from WDR-23 and SKN-1 (Fig 5B). To test if skr-1/2 is required for nuclear localization of SKN-1, we performed skr-1/2(RNAi) in a strain carrying the SKN-1b/c::GFP transgene and scored accumulation after exposure to juglone. Because we did not observe an increase of SKN-1b/c::GFP accumulation in the intestine nuclei by juglone (Fig 3A), we scored the effects of skr-1/2(RNAi) on SKN-1b/c::GFP accumulation in the head region. RNAi against skr-1/2 failed to prevent accumulation of SKN-1b/c::GFP in the head after juglone exposure (Fig 8D).
We next determined if skr-1/2 loss influences phosphorylation of PMK-1. RNAi against skr-1/2 had no effect on total or phosphorylated PMK-1 levels under basal conditions or when increased by arsenite (Fig 8E), a condition that strongly increased phosphorylated PMK-1 levels (Fig 1). Taken together, these data suggest that although skr-1/2 is required for activation of SKN-1 target genes by stress, it does so without observable changes to SKN-1 nuclear accumulation or p38 MAPK phosphorylation.
As mentioned earlier, skr-1 and skr-2 are very similar even at the nucleotide level (83% identical). In order to investigate if one or both are required for gst induction, we conducted qPCR experiments with skr-1(tm2391) and skr-2(ok1938) deletion mutants. skr-1(tm2391) and skr-2(ok1938) homozygotes are each maternal effect embryonic lethal and skr-1 mutants also display late larval arrest [34]; RNA was isolated by picking healthy homozygote larvae at the L3 and early L4 stages. The skr-1(tm2391) allele reduced mRNA levels of four juglone induced gst genes (S10A Fig). The skr-2(ok1938) allele only slightly reduced mRNA levels for two genes (gst-4 and 12) and increased induction of gst-10 and gst-30. Interestingly, skr-2(ok1938) had a ~40% reduction in skr-1 mRNA levels (S10B Fig), likely due to a partial deletion of the skr-1 promoter in this mutant; however, this reduction in skr-1 mRNA appeared to have minimal effect on the activation of skn-1 dependent genes in response to juglone. These results help confirm a role of skr-1 and are consistent with SKR-2 being less important than SKR-1 for SKN-1 pathway activation.
To test if overexpression of skr-1/2 alone is sufficient to increase expression of SKN-1 transcriptional targets, we generated a transgenic worm overexpressing a 4 kb DNA fragment that contained full genomic sequences of both skr-1 and skr-2 and an integrated Pgst-4::GFP reporter. Transgenic worms carrying the 4 kb DNA fragment overexpressed (oe) both skr-1 and skr-2 mRNA, but had wild-type levels of Pgst-4::GFP fluorescence (S11A and S11B Fig); qPCR confirmed that the mRNA levels of SKN-1 transcriptional targets gst-4 and gcs-1 were unaffected by skr-1/2 overexpression at basal conditions, and the overexpression also did not further activate SKN-1 target transcription during juglone exposure (S11C Fig). Therefore, overexpression of skr-1/2 is not sufficient to activate the SKN-1 stress response.
To determine the expression pattern of SKR-1, we generated transgenic worms expressing a SKR-1::GFP fusion protein. Consistent with a previous study [24], SKR-1::GFP was expressed throughout the worm, with SKR-1::GFP highly expressed in the intestine, pharynx, neurons, and spermatheca (S12 Fig); intracellular distribution of SKR-1 was not previously determined. Within the large cells of the intestine, we observed SKR-1::GFP in the cytosol and nuclei suggesting that the protein is broadly distributed within cells (Fig 9A). To test if SKR-1 is regulated under stress, we treated this transgenic worm to conditions that activate SKN-1. Exposure to juglone and arsenite did not result in obvious changes in expression or localization of SKR-1::GFP (S11D Fig).
To test whether SKR-1 might interact with key members of the SKN-1 pathway, we co-transfected either full length SKN-1c or WDR-23a fused to an N-terminal GST tag together with SKR-1 fused to an N-terminal V5 fusion tag in HEK293 cells and performed GST pull-downs. Pulldown of GST-SKN-1c failed to capture SKR-1 as determined by Western blot using a V5 monoclonal antibody (Fig 9B). Alternatively, SKR-1 was captured by pull-down of GST-WDR-23a. We also used a strain expressing an integrated transgene of wdr-23a cDNA fused to GFP [38], which rescues misexpression of a gst-4 transgene in wdr-23 mutants, to test if SKR-1/2 might regulate WDR-23 in vivo. As shown in Fig 9C, skr-1/2(RNAi) had a significant but very small effect on WDR-23::GFP fluorescence levels. Alternatively, skr-1/2(RNAi) doubled the proportion of worms with obvious nuclear localization in the hypodermis (Fig 9D). Obvious changes to intestinal WDR-23::GFP were not observed.
In C. elegans, the p38 MAPK signaling cascade activates SKN-1 in the intestine during oxidative stress induced by arsenite exposure [19] and during pathogen infection [20,21]. Here, we demonstrate striking variation in PMK-1 activation kinetics among diverse pro-oxidant and electrophilic compounds that all strongly induce SKN-1 dependent transcriptional responses. SKN-1 can also accumulate in the nuclei of tissues other than the intestine partially independent of PMK-1. To begin defining other regulatory pathways that may function parallel to PMK-1, we leveraged the genetic tractability of C. elegans to identify a novel role for the highly conserved Skp1 homologs skr-1/2 in the SKN-1 mediated stress response. SKR-1 interacts with and influences the localization of the SKN-1 repressor WDR-23, and loss of skr-1/2 inhibits the expression of skn-1 dependent detoxification genes and impairs survival during exposure to pro-oxidants. A revised working model for SKN-1 regulation by reactive small molecules that incorporates our findings is presented in Fig 10 and discussed below.
Phosphorylation of PMK-1 at activating residues by arsenite and pathogen exposure is associated with intestinal accumulation of nuclear SKN-1 consistent with the prevailing model in which PMK-1 phosphorylates SKN-1 in the intestine during oxidative stress and causes nuclear accumulation [19]. Interestingly, PMK-1 phosphorylation kinetics varied greatly between different sub-lethal pro-oxidant and electrophile exposures that all strongly activate SKN-1 dependent detoxification gene expression (Fig 1). Acrylamide even decreased PMK-1 phosphorylation levels during chronic exposure, but caused strong and sustained Pgst-4::GFP induction (Figs 1 and 2 and S3 Fig). We also observed strong SKN-1 dependent detoxification gene activation in pmk-1 deletion mutants during sub-lethal exposure to arsenite and juglone (Fig 2) and SKN-1 nuclear accumulation in tissues other than the intestine (Fig 3 and S4 Fig) and in pmk-1 deletion mutants (Fig 4).
Importantly, SKN-1 dependent gene induction was partially dependent on pmk-1 (Fig 2B), which is consistent with a recent study showing that sek-1 and pmk-1 are required for induction of a downstream target of SKN-1 in neurons (nlg-1) that promotes survival of juglone [39]. We also cannot rule out the possibility that SKN-1 accumulates in the intestine with juglone at levels below what we are able to detect. Our data establish that pro-oxidants and electrophiles are capable of activating SKN-1 and its downstream detoxification genes outside the intestine and with varying degrees of SKN-1::GFP nuclear accumulation and PMK-1 activation suggesting the presence of parallel or compensatory mechanisms; e.g., other kinases downstream from NSY-1/SEK-1 (such as PMK-2 or PMK-3) or other unknown pathways. It is also possible that other transcription factors could help compensate.
SKR homologs function in SCF multi-subunit E3 ubiquitin ligases that are conserved from yeast to humans. In yeast and humans, a single SKR named Skp1 forms the SCF complex with Rbx, Cul1, and various F-box proteins to promote protein ubiquitination [40,41]. Skp1 also functions as a scaffold in protein complexes independently of SCF [42,43,44,45]. In C. elegans, 21 skr genes have been identified [23,24]. Loss of function of either skr-1 or 2 results in embryonic arrest that is characterized by excessive cell numbers and hyperplasia [23,24]. Interestingly, skr-1/2(RNAi) delays degradation of SKN-1 during embryonic development leading to a delay in development [46]. SKR-1/2 was previously shown to be required for longevity in daf-2 mutants and not wildtype worms [33], but a role in stress responses has not been reported. In this study, we identified a new role for skr-1/2 in permitting skn-1 dependent stress responses in larval and adult stages of C. elegans.
skr-1/2 were required for transcriptional induction of a SKN-1 dependent gst-4 reporter by diverse pro-oxidants at sub-lethal doses (Fig 6) and for resistance to lethal doses of juglone and arsenite (Fig 8A and 8B). We found that skr-1/2(RNAi) did not reduce levels of total or phosphorylated PMK-1 (Fig 8E), and in our RNA-seq data, knockdown of skr-1/2 did not alter mRNA levels of core MAPK genes (nsy-1, sek-1, or pmk-1) (S1 Table). Although it remains possible that skr-1/2 could affect protein levels of NSY-1 or SEK-1, these data raise the possibility of SKR-1/2 functioning by a different mechanism.
The function of skr-1/2 in our study was shown to be tightly linked to SKN-1. Knockdown of skr-1/2 did not block heat-shock or osmotic stress transcriptional response reporters (S6 Fig) and skr-1/2 dependent genes were enriched for functions in detoxification and cuticle collagen (Fig 7), which are both also enriched among skn-1 dependent genes [12,30,47]. Furthermore, our RNA-seq data demonstrated that a strikingly large majority (84%) of skr-1/2 dependent genes were also skn-1 dependent and that expression changes caused by skr-1/2 and skn-1 RNAi were correlated (Fig 7). skr-1/2 had no correlation with genome-wide daf-16 dependent expression (Fig 7).
The SCF ubiquitin ligase complex has previously been reported to negatively regulate Nrf2 protein levels in human cells; phosphorylation of Nrf2 by glycogen synthase kinase 3 promotes ubiquitin mediated Nrf2 degradation by the SCF complex via its interaction with the β-transducin repeat-containing protein (β-TrCP) F-box [48]. Our results show that skr-1/2 Skp1 homologs in C. elegans positively regulate skn-1 mediated stress responses. RNAi against the gene encoding the central scaffold of the C. elegans SCF complex, cul-1, failed to mimic the effects of skr-1/2(RNAi) on regulation of the gst-4 promoter during exposure to juglone (S7 Fig). However, we cannot rule out a role for cul-1 because of the possibility of residual CUL-1 function after RNAi treatment.
C. elegans and Drosophila melanogaster each contain expanded families of Skp1 homologs whereas yeast and vertebrates, including humans, have only one member [23,24]. Having multiple Skp1 homologs has been hypothesized to permit the evolution of more flexible and variable functions. Even in yeast and human cells where single homologs are present, evidence has been provided for Skp1 functioning independently of the SCF complex to regulate membrane protein recycling, kinetochore function, ion transport, and protein degradation [42,43,44,45,49]. In all these cases, the molecular role of Skp1 appears to be as a scaffold to assist in protein complex assembly.
In C. elegans, SKN-1 is regulated directly by WDR-23 and PMK-1 [16,19]. Our data do not support a positive association between SKR-1/2 and PMK-1 function as skr-1/2(RNAi) did not reduce PMK-1 protein levels or its stress activated phosphorylation (Fig 8E). Instead, SKR-1 has the potential to interact with WDR-23 and the requirement of skr-1/2 for skn-1 dependent gene induction was abolished in a wdr-23 null mutant (Fig 5B). WDR-23 is a direct repressor of SKN-1 that is present in many cells [16,25,50], and loss of wdr-23 promotes stress resistance and a long-lived phenotype that is skn-1 dependent [10]. SKR-1 was present throughout cells including within some nuclei (Fig 9A) where it could interact with WDR-23. Our genetic interaction and WDR-23::GFP results (Figs 5B, 9C and 9D) are consistent with SKR-1/2 functioning to negatively regulate WDR-23 accumulation in nuclei likely by serving as a scaffold to influence protein complex assembly (Fig 10). SKN-1 dependent genes are extremely sensitive to changes in WDR-23 function [16] and one possible mechanism is for SKR-1/2 to negatively regulate WDR-23 protein levels via ubiquitination. In this model, loss of skr-1/2 would increase WDR-23 levels and repression of SKN-1.
Importantly, we did not observe accumulation of either SKN-1b/c or SKN-1op(a/b/c) reporters in the intestine after juglone exposure and skr-1/2 RNAi did not impair SKN-1b/c::GFP accumulation in head nuclei (Fig 8D). Although SKR-1/2 could influence SKN-1 accumulation at levels that we were not able to detect, it is also possible that SKR-1/2 and WDR-23 may be able to regulate SKN-1 by alternative mechanisms such as DNA binding or transactivation activity. Future experiments to investigate these alternatives and to identify other SKR-1 and WDR-23 interacting proteins required for the detoxification response could help clarify the molecular mechanisms of WDR-23 and SKN-1 regulation. Induction of SKN-1 detoxification genes can also be decoupled from nuclear accumulation in the intestine during genetic impairment of translation and fatty acid metabolism pathways [51,52]. It remains to be seen if SKR-1/2 plays a role under these conditions.
Our study provides evidence that PMK-1 phosphorylation and SKN-1 nuclear localization are differentially regulated in response to different reactive small molecule exposures that activate SKN-1 dependent detoxification responses. We also identify SKR-1/2 as required for SKN-1 dependent detoxification gene induction in response to diverse pro-oxidants and electrophiles that may interact with and regulate WDR-23. SKN-1 is activated by a broad range of different conditions and our study highlights the fact that distinct upstream regulatory pathways are present that may permit tailored responses to oxidative and reactive small molecule stressors. These findings and the identification of SKR-1/2 as a key regulator lay the foundation for defining a novel SKN-1 regulatory mechanism that involves WDR-23.
C. elegans strains were grown and maintained at 20°C using standard methods [53], unless noted otherwise. The following strains were used: wild-type N2 Bristol, VP596 dvIs19 [pAF15(Pgst-4::GFP::NLS)]; vsIs33[Pdop-3::RFP]), QV65 gpIs [Phsp-16.2::GFP]; vsIs33, CF1580 daf-2(e1370) III; muIs84 [(pAD76) sod-3p::GFP + rol-6], QV25 wdr-23(tm1817); eri-1(mg666)IV; dvIs19, QV130 skn-1(k1023); dvIs19, KU25 pmk-1(km25)IV, KU4 sek-1(km4)X, LD1 ldIs7 [skn-1B/C::GFP + pRF4(rol-6(su1006))], LD1008 [skn-1(operon)::GFP, rol-69su1006)], VP537 eri-1(mg666)IV; dvIs19, GR1373 eri-1(mg366)IV, VP604 kbIs24 [Pgpdh-1::dsRed2;Pmyo-2::GFP;unc-119] X, QV254 zjEx114 [SKR-1::GFP; Pmyo-2::tdTomato], QV256 zjEx115 [skr-1/2 gDNA; Pmyo-2::tdTomato; Pmyo-3::dsRed], QV288 pmk-1(km25)IV; ldIs7, VC1439 skr-2(ok1938) I/hT2 [bli-4(e937) let-?(q782) qIs48] (I;III), CU6110 skr-1(tm2391) I/hT2 I;III; +/hT2 I;II, and KHA116 unc-119(ed3); chuIs116 [wdr-23p::wdr-23a(cDNA)::GFP, unc-119].
RNAi was performed as described previously [16] by feeding worms strains of E. coli [HT115(DE3)] that are engineered to transcribe double stranded RNA (dsRNA) homologous to a target gene. For screening, L1 larvae of VP537 worms were grown in liquid medium with dsRNA-producing bacteria for 3 days, and subsequently exposed to 38 μM juglone for 4 h and screened for Pgst-4::GFP expression with a Zeiss Stemi SV11 microscope. The entire ORFeome RNAi feeding library (Open Biosystems, Huntsville, AL) was screened with additional missing clones supplemented from the original genomic RNAi feeding library (Geneservice, Cambridge, United Kingdom). Clones that resulted in reduced Pgst-4::GFP were rescreened three additional times, clones with positive scores in all three trials were considered novel regulators of gst-4 and subsequently sequenced for identification. With the exception of the genome-wide RNAi screen, all subsequent RNAi experiments were performed on nematode growth media (NGM) plates that were made with the addition of 50 μg mL-1 carbenicillin, 0.2% lactose, and seeded with appropriate HT115 RNAi bacteria and grown overnight before use. RNAi feeding was initiated at either synchronized L1 larvae stage, or at L4 to young adult stage for RNAi clones that displayed developmental effects when fed from L1. Bacteria with plasmid pPD129.36 were used as a control for non-specific RNAi effects. This control plasmid expresses 202 bases of dsRNA that are not homologous to any predicted C. elegans gene.
To visualize SKN-1b/c::GFP, LD1 worms at L1/L2 stages were incubated with 5 mM sodium arsenite for 1 h, 5 mM sodium azide for 5 min, 35 mM paraquat for 2h, 38 μM juglone for 5–15 min, 1, 3, and 5 h, or 7 mM acrylamide for 5–15 min, 1, 4, 12, or 24 h. Worms were then washed and anesthetized with 5 mM levamisole. Anesthetized worms were mounted on 2% agarose pads and visualized and imaged using an Olympus BX60 microscope and Zeiss AxioCam MRm camera. Differential interference contrast and fluorescent images of SKN-1b/c:GFP and SKN-1op::GFP were taken. Both grayscale images taken by the GFP and RFP channels were merged into green and red channels respectively to produce a composite image using ImageJ (NIH). When needed for clarity, brightness and contrast adjustments were made equally to images from the same color channel and within the same experiment. Accumulation of SKN-1b/c::GFP and SKN-1op::GFP expression in intestinal nuclei was scored as in previous studies [19], while accumulation of SKN-1b/c::GFP and SKN-1op::GFP in the head regions were scored as either negative, referring to no observation of GFP signals in the head region, or positive, referring to GFP signals observed throughout the head region.
To visualize Pgst-4::GFP, young adult worms were treated with 5 mM sodium arsenite for 1 h in liquid NGM (and were washed 3x with NGM buffer and allowed to recover for 3 h on NGM agar), 35 mM paraquat for 2 h in liquid NGM (recovery for 2 h), 38 μM juglone in liquid NGM for 3 h (recovery for 1 h), or 7 mM acrylamide in liquid NGM for 4 h (no recovery). These conditions were all sub-lethal (S1A Fig) but strongly induce Pgst-4::GFP fluorescence (Fig 5A). For Pgst-4::GFP scoring, low refers to little to no GFP observed throughout the worm, medium refers to GFP signals observed only at the anterior and posterior ends of the worm, and high refers to GFP signals observed throughout the body. To visualize Phsp-16.2::GFP, young adult worms were exposed to 33°C for 1 h followed by 5 h recovery at 20°C; for Pgpdh-1::RFP, young adult worms were exposed to 250 mM NaCl for 24 h. To visualize effects of oxidants on SKR-1::GFP expression and localization, young adult worms were exposed to 38 μM of juglone for 3 h or 5 mM arsenite for 1 h.
Quantitative real-time RT-PCR was used to measure mRNA levels in L4 to young adult stage worms fed with appropriate RNAi as described previously [16]. Worms were incubated with 5 mM sodium arsenite or 38 μM juglone in liquid NGM for 1 or 3 h at 20°C with gentle shaking. Total RNA from 200–300 worms was isolated with a Quick-RNA MicroPrep Kit (Zymo Research, Irvine, USA), and cDNA was synthesized using 1 μg of RNA with GoTaq 2-Step RT-qPCR System (Promega, Madison, WI, USA) following the manufacturer’s protocol. Quantitative real-time PCR was performed in 10 μL reactions in a Realplex ep gradient S Mastercycler (Eppendorf AG, Hamburg, Germany) with GoTaq Green Master Mix (Promega, Madison, WI) according to the manufacturer’s protocol. Data was analyzed by the standard curve method, with the housekeeping genes rpl-2 and cdc-42 used as internal reference controls. Primer sequences are available upon request.
Synchronized L4 to young adult stage N2 worms were treated with compounds in liquid NGM while gently shaking at 20°C. Each treatment was accompanied by a control group of worms incubated in NGM buffer for the same duration. After incubation with each stressor, worms were washed three times with NGM buffer, and approximately 1,000 L4 to young adult stage worms were lysed in homogenization buffer for each replicate (homogenization buffer content: 50 mM Tris Base pH 7.5, 150 mM NaCl, 0.1% SDS, 0.5% NaDeoxycholate, 1x Halt™ protease inhibitor cocktail, and 1x Halt phosphatase inhibitor cocktail (LifeTechnologies, Cat# 78430, #78420, Rockford, IL). Worms were sonicated for complete lysis by ultrasonication (Misonix XL 2000, Farmingdale, NY). Worm lysates were centrifuged at 13,000 x g for 10 min at 4°C and supernatants were normalized for protein concentration with BCA protein assay (Pierce, Cat#23227) and collected for SDS-PAGE electrophoresis. Equal volume of lysates totaling 30 μg of proteins were loaded and separated by SDS-PAGE, and detected by immunoblotting with phosphorylated PMK-1 (1:2000; Promega, Cat# V1211x), total PMK-1 antibody (1:1000; gift from K. Matsumoto [54]), and β-tubulin antibody (1:100; Developmental Studies Hybridoma Bank, Cat # E7) with methods as previously described [16].
HEK293 cells were cultured in Dulbecco modified Eagle medium supplemented with 10% fetal bovine serum, 4.5 g/l glucose, 584 mg/l L-glutamine, 100 mg/l sodium pyruvate, and 1 U/ml penicillin. Co-transfections were performed with Lipofectamine LTX with PLUS reagents (Life Technologies, Cat#1533810) using full length SKN-1c or WDR-23a cloned into pDEST27 vector and full length SKR-1 cloned into V5-DEST vector. Two days after transfection, cells were lysed with immunoprecipitation lysis buffer (Life Technologies, Cat # 87787) and pulldown was conducted using glutathione-Sepharose 4B at room temperature for 1 h (GE Healthcare, Cat#17-0756-01 Little Chalfont, United Kingdom). Beads from pulldowns were washed extensively with PBS+0.1% Triton-X and eluted with an equal volume of 2X SDS loading buffer by heating at 90°C for 5 minutes. Western blots were carried out as described above, with mouse anti-GST MAb (1:1,000; Santa Cruz Biotech, Cat #B-14 Dallas, TX), and mouse anti-V5 tag MAb (1:1000; Invitrogen, Cat #R960-25).
Juglone stress resistance assays were performed on synchronized worms at L4 stages fed HT115 bacteria. Worms were pretreated with 38 μM of juglone in liquid NGM for 2 h and subsequently transferred to 125 μM juglone for survival analysis similar to [37]. In the arsenite stress resistance assay, worms were incubated with 10 mM of arsenite in liquid NGM and analyzed for survival. Worms were considered dead if they did not display any movement in response to prodding with a thin wire. A total of three independent trials were performed for each survival assay.
N2 worms were synchronized via by hypochlorite treatment and grown on RNAi for two days. Worms were then incubated in either NGM buffer (control) or 38 μM juglone for 3 h. RNA was extracted from ~1,000–2,000 worms per sample using the RNAqueous-Micro Total RNA Isolation Kit (ThermoFisher Scientific, Cat#AM1931). Total RNA was sent to The Yale Center for Genome Analysis for 75 nucleotide single-end sequencing in an Illumina HiSeq 2000. Raw sequencing data was processed using the public Galaxy server and mapped to the C. elegans genome (ce10). Using the Cufflink package and CuffDiff application from Galaxy [55], FPKM (Fragments Per Kilobase of transcript per Million mapped reads) were calculated and tested for differential expression with a FDR score of 5%. Differentially expressed genes were clustered for GO (Gene Ontology) analysis by DAVID for gene functional classification [56].
Statistical significance was determined using Student’s T-test when two means were compared and a one-way analysis of variance (ANOVA) with Tukey’s or Bonferonni multiple comparison tests when three or more means were compared. Log-Rank tests in the OASIS online tool were used when survival curves were compared [57]. Chi-square tests were used to evaluate categorical data, and linear regression was used for testing correlations in gene expression. P values of <0.05 were taken to indicate statistical significance except for comparing more than two survival curves, in which Bonferonni adjustments were made to P values to account for repeated comparisons. Statistical significance is indicated in figures as *P<0.05, **P<0.01, ***P<0.001, and ns = not significant. Data is available at: https://figshare.com/s/fb874c6dc1aa9b77375c
|
10.1371/journal.pcbi.1005216 | Kinase Inhibition Leads to Hormesis in a Dual Phosphorylation-Dephosphorylation Cycle | Many antimicrobial and anti-tumour drugs elicit hormetic responses characterised by low-dose stimulation and high-dose inhibition. While this can have profound consequences for human health, with low drug concentrations actually stimulating pathogen or tumour growth, the mechanistic understanding behind such responses is still lacking. We propose a novel, simple but general mechanism that could give rise to hormesis in systems where an inhibitor acts on an enzyme. At its core is one of the basic building blocks in intracellular signalling, the dual phosphorylation-dephosphorylation motif, found in diverse regulatory processes including control of cell proliferation and programmed cell death. Our analytically-derived conditions for observing hormesis provide clues as to why this mechanism has not been previously identified. Current mathematical models regularly make simplifying assumptions that lack empirical support but inadvertently preclude the observation of hormesis. In addition, due to the inherent population heterogeneities, the presence of hormesis is likely to be masked in empirical population-level studies. Therefore, examining hormetic responses at single-cell level coupled with improved mathematical models could substantially enhance detection and mechanistic understanding of hormesis.
| Hormesis is a highly controversial and poorly understood phenomenon. It describes the idea that an inhibitor molecule, like an anti-cancer or anti-microbial drug, can inadvertently stimulate cell growth instead of suppressing it. This can have a profound effect on human health leading to failures in clinical treatments. Therefore, getting at the mechanistic basis of hormesis is critical for drug development and clinical practice, however molecular mechanisms underpinning hormesis remain poorly understood. In this paper we use a mathematical model to propose a simple and yet general mechanism that could explain why we find hormesis so widely in living systems. In particular, we discover that hormesis is present within a fundamental structure that forms a basic building block of many intracellular signalling pathways found in diverse processes including control of cell reproduction and programmed cell death. The benefits of our study are two-fold. Having simple molecular understanding of the causes of hormetic responses can greatly improve the design of new drug compounds that avoid such responses. Moreover, due to the fundamental nature of the newly proposed mechanism, our findings have a potential broad applicability to both anti-cancer and anti-microbial drugs.
| Hormesis is a phenomenon describing biphasic dose response relationships that exhibit low-dose stimulation and high-dose inhibition [1]. Many medical agents such as antibacterials, antifungals, and anti-tumour drugs have been found to display hormetic response [2] with the earliest observations dating back to 1800s. In particular, low concentrations of certain antifungals were found to stimulate fungal growth [3] or metabolism [4] while inducing toxicity at high concentrations. From the early 1920s the concept of low-dose stimulation and high-dose toxicity of various chemical elements with respect to bacterial growth was widely recognised [5]. We now know that bacteria can exhibit hormetic response to a wide range of antibiotic drugs, regardless of their mode of action [6]. This phenomenon is also found in tumour cells exposed to anti-tumour drugs. In fact, hormesis has been observed in an astonishingly broad range of tumour types including pancreatic, colon and breast (reviewed in [7]).
Despite the overwhelming body of research, some dating back a century, that documents hormetic responses to a broad range of compounds, their clinical significance has only relatively recently come to the fore [2]. The consequence of hormesis could have a profound effect for human health [8, 9]. Drug concentration generally varies substantially within the human body and as drug gets cleared, the associated low concentration can in turn stimulate pathogen or tumour growth. Therefore understanding the mechanistic basis of hormesis is vital for both drug development and clinical practice.
The vast majority of targets for antibiotics, antifungals and anti-tumour drugs fall into the following categories: enzymes, receptors, transporters and DNA/RNA and the ribosome [10]. However how such drug-target interactions lead to hormesis remains poorly understood. The biological explanations put forward are overcompensation after a disruption of homeostasis (reviewed in [11]), direct stimulatory response [12], superimposition of different monotonic dose-response curves [13], or heterogenic susceptibility of different tissues to the same stimuli [14]. These explanations provide understanding of hormesis at a phenotypic level but lack understanding at the molecular level. Some inroads have also been made with respect to mammalian cells focusing on drug mechanisms mediated via receptor and/or cell signalling pathways (reviewed in [7]). For example, biphasic dose response could occur through interaction of two different receptor subtypes that mediate/activate opposing stimulatory and inhibitory pathways via the same antagonist [15]. However, hormetic response is a built in feature of such receptor mediated mechanisms rather than an emergent property of the underlying biological system.
An area of research where understanding of the mechanisms giving rise to hormesis is particularly lacking involves enzyme-targeting drugs. Known as enzyme inhibitors, they are designed to block enzyme activity leading to disruption of bacterial cell wall [16], fungal membranes [17] and fungal cell wall [18] as well as programmed tumour cell death [19], to name a few. With regards to hormetic dose-responses to antibiotics, a recent study focusing on inhibition of a specific enzyme, Dihydropteroate synthase, suggested the involvement of bacterial quorum sensing [20]. To our knowledge, mechanisms behind hormetic dose-response to enzyme-inhibiting antifungals are not known.
In recent years kinase inhibitors, a subset of enzyme inhibitors, have been shown to be very effective therapeutic agents in a broad range of diseases, including cancers. Amongst other enzyme inhibitors, significant attention has been focused on those inhibiting the mitogen-activated protein kinase (MAPK) pathway [21–24], which is of fundamental importance to human health as abnormal regulations of MAPK contribute to tumour progression [25].
The observations of hormesis in MAPKs as a result of inhibition of BRAF oncogene are widespread: low doses of RAF inhibitors designed to cease tumour proliferation [26] can cause a paradoxical activation of tumour cell activity through undesired MAPK up-regulation [8, 9, 27–32]. Current explanations of hormetic responses induced by RAF kinase inhibition involve complex phenomena affecting regulatory mechanisms, feedback pathways or enzymatic activity [33], making them difficult to generalise. More generally, enzyme competition for the same substrate was recently proposed as a simpler mechanism giving rise to hormetic effects of enzyme-targeting Alzheimer’s drugs [34].
In this paper we put forward a novel, simple but general mechanism driving hormetic responses in systems where an inhibitor acts on an enzyme. We develop a mathematical model based on a basic building block in intracellular signalling, namely a dual phosphorylation-dephosphorylation motif, to which a kinase inhibitor is applied. In a broader context, dual-phosphorylation can be found in diverse processes such as circadian rhythms [35], virulence regulation [36, 37], mitotic entry [38], transcription [39, 40], cytokine production [40], as well as in MAPK pathways which regulate primary cellular activities in eukaryotes including proliferation and programmed cell death [41, 42].
The model demonstrates that under certain conditions the steady state amount of the double-phosphorylated protein substrate in the cycle can substantially increase at low inhibitor doses compared to the base level without inhibition. Therefore the dose-response curve of the double-phosphorylated substrate possesses a hallmark of hormesis: it is upward sloping at low inhibitor doses and downward sloping at high inhibitor doses. The existence of hormesis in our model depends on the mechanism of inhibition and the dissociation rates of the kinase-substrate-inhibitor complexes. We also found that the magnitude of hormetic responses depends on the substrate-kinase ratio in a non-monotone way.
The benefits of our study are two-fold. Our mechanism is based on a principal component of intracellular signalling pathways, and as such has a potential broad applicability. Moreover having simple molecular understanding of the causes of hormetic responses can greatly improve the design of new drug compounds that avoid such responses.
We consider a simple dual phosphorylation-dephosphorylation motif, whereby a distinct kinase protein is phosphorylating a separate protein substrate. Multiple phosphorylations can occur in close proximity or in diverse sites on a protein and here we focus on the former, instances of which can be found in activation of conventional MAPK enzymes [43], cell-cycle regulation via cyclin-dependent kinase 1 [44], regulation of other non-MAPK kinases [45] and ion channel trafficking [46]. The motif we consider is a subset of futile cycles [47, 48] also known as a single stage module in the context of MAPK pathways [49, 50]. Based on the experimental evidence for MAPK pathways [51–53] we assume that our motif follows a distributive mechanism consisting of two sequential phosphorylation steps and two sequential dephosphorylation steps that share the same intermediate mono-phosphorylated form. In particular, the protein substrate (C) is first converted into a mono-phosphorylated form (CP) and subsequently into a double-phosphorylated form (CPP), through a chain of reactions facilitated by a kinase (kin). Conversely CPP is converted back to CP which is subsequently converted to C, through a chain of reactions facilitated by a phosphatase (pho). In the distributive mechanism, the kinase(phosphatase) facilitates at most one phosphorylation (dephosphorylation) in each molecular encounter [48].
Therefore our dual phosphorylation-dephosphorylation motif can be described by the following reaction kinetic equations, which are a simplification of the reaction scheme described in [54]:
C+kin⇌k−1k1C·kin→k2CP+kin,CP+kin⇌k−3k3CP·kin→k4CPP+kin,CPP+pho⇌k−5k5CPP·pho→k6CP+pho,CP+pho⇌k−7k7CP·pho→k8C+pho.
(1)
Next we describe the assumptions behind the introduction of an inhibitor into Eq (1), based on the general modifier mechanism also known as hyperbolic or partial competitive inhibition [55]. We assume that the inhibitor (inh) is able to react with the kinase and the substrate-kinase intermediate complexes C · kin and CP · kin according to the following inhibition scheme:
kin+inh⇌drdfkin·inh,C·kin+inh⇌e−1e1C·kin·inh⇌e−2e2C+kin·inh,CP·kin+inh⇌e−3e3CP·kin·inh⇌e−4e4CP+kin·inh. (2)
The first- and second-order rates ki and ei in Eqs (1) and (2) and the association and dissociation rates df and dr in Eq (2) are considered dimensionless. In our system intermediate substrate-kinase-inhibitor complexes are able to dissociate into a substrate and kinase-inhibitor complex with forward e2, e4 and backward e−2, e−4 rates [56].
The model describing the time evolution of the substrate, kinase, phosphatase and inhibitor concentrations is based on the law of mass action and assumes the total conservation of mass holds for all four compounds. The details of the system of 9 differential equations and the corresponding analysis are presented in S1 Appendix. (with Supplementary Tables A1 and A2 containing model parameter values). This model system is studied under steady state conditions, that is, when all concentrations of reactants have reached a dynamic equilibrium. Numerical simulations are conducted with Matcont, a continuation package in MATLAB used for numerical bifurcation analysis of ODEs [57].
In the absence of an inhibitor, the double phosphorylation motif Eq (1) can possess either a single or two stable steady states of the doubly-phosphorylated form of the substrate CPP, [50, 54, 58, 59]. Therefore in our study we consider two cases: first, when the motif Eq (1) is monostable and second, when this motif is bi-stable.
In the case of a single stable steady state (CPP*) in the absence of an inhibitor, we find that CPP can exhibit biphasic (or hormetic) response to an inhibitor as illustrated in Fig 1. In particular, the observed dose-response curve in the presence of an inhibitor has an inverted U-shape: for sufficiently low inhibitor doses the computed steady-state values of CPP increase monotonically, while for sufficiently large inhibitor doses, the computed steady-state values of CPP monotonically decrease.
Moreover, by making simplifying assumptions that e−2 = e−4 = 0, e2 >>e1, e4 >>e3 and the inhibitor has fast off rate, we can analytically derive the slope of the dose-response curve, in other words the slope of the relationship between the steady-state value of CPP and the total amount of inhibitor at low doses (see S1 Appendix for details). This allows us to identify two primary factors necessary for the hormesis to be observed:
Note that the hormesis is still observed in numerical simulations when e−2, e−4 > 0 (Fig A1 in S1 Appendix).
In addition the above conditions (C1-C2) can also be used to forecast the presence of a hormetic dose response in the second case under our consideration, namely when in the absence of an inhibitor the motif Eq (1) has two stable steady states CPP,1* (Fig 2A) and CPP,2* (Fig 2B). In this case the numerical simulations predict that cells with high base level of double-phosphorylated substrate will respond differently to inhibition from the cells with low base level of double-phosphorylated substrate. In particular, cells with initially high levels of CPP (at steady state CPP,1*) will exhibit a monotone decreasing dose-response (Fig 2A) while cells with low initial levels of CPP (at steady state CPP,2*) will exhibit a hormetic response (Fig 2B).
The magnitude of hormetic response can differ between the mono- and bi-stable cases under consideration as illustrated in Figs 1 and 2B. In the mono-stable case the CPP value at dose inh* is approximately two-fold higher compared to the base level CPP* value in the absence of an inhibitor (Fig 1). In the bi-stable case the CPP value at dose inh* is approximately six-fold higher than the base level CPP,2* value in the absence of an inhibitor (Fig 2B).
In general, we find that the ratio of total mass of protein substrate to kinase mass influences the magnitude of hormetic response in a non-monotone way as shown in Fig 3. For sufficiently small substrate-kinase ratio, a hormetic response is not observed (absence of hormesis is labelled as 100% response in Fig 3 because the maximal response is equal to the baseline of no inhibition). However, the hormetic response increases sharply as the substrate-kinase ratio increases. Further increases of this ratio lead to a sharp decline in the magnitude of hormetic response, which continues to increase slowly for sufficiently large substrate-kinase ratios (see Fig 3 inset). Therefore, the magnitude of hormetic response peaks at intermediate values of the substrate-kinase ratio, as frequently observed in the MAPK pathway [60] for example, while hormesis is not observed for low substrate-kinase ratios.
Hormetic responses to enzyme-targeting drugs have been observed in both prokaryotes [20, 61, 62] and eukaryotes [8, 9, 27, 31, 32] but the mechanistic understanding behind such responses is still lacking. In this paper we focus on eukaryotic cells and propose a novel, simple but general mechanism that could give rise to hormesis in systems where an inhibitor acts on an enzyme.
At the core of our newly-proposed mechanism is one of the basic building blocks in intracellular signalling, the dual phosphorylation-dephosphorylation motif, found in diverse regulatory processes including MAPK pathways which control cell proliferation and programmed cell death in eukaryotes [41, 42]. We analytically derive conditions that lead to hormetic dose-response of the doubly-phosphorylated substrate in the presence of a kinase inhibitor. The conditions required for hormesis to be observed are surprisingly simple and involve two main factors: (C1) strong dissociation effect of intermediate substrate-kinase-inhibitor complexes and (C2) large dissociation rate of kinase-inhibitor complexes.
Crystallographic studies of kinase inhibitors bound to their targets demonstrate that a number of different conformational states can be induced. Type 1 kinase inhibitors are defined as binding the kinase in its active conformation and crystal structures of ternary complexes of ATP analogues bound with substrate peptides are reported (for review see [63, 64]). Indeed it is not uncommon for crystal systems of substrate peptide complexes to be used in Structure Based Design campaigns to develop Type 1 kinase inhibitors [65].
Given the fundamental nature of the dual phosphorylation-dephosphorylation motif and the relative simplicity of the derived conditions necessary to observe hormesis, why was this mechanism previously overlooked in theoretical literature? A further examination of the (C1) condition could provide a potential answer. In general, when considering partial competitive enzyme inhibition [55] as we do here, classical enzyme kinetics literature [55, 56] assumes not only equilibrium concentrations of different enzyme species but it also assumes that at those equilibrium concentrations there is no flux through substrate-kinase-inhibitor complexes. However, we find that in our study as flux decreases the maximum hormetic response also decreases (Fig 4) indicating that under the no-flux assumption, hormetic responses could be overlooked.
Once a new mechanism is proposed to explain a particular biological phenomenon, ideally it should be put to test. However, there are a number of difficulties associated with in vitro tests of our model predictions. First, biochemical assays involved with in vitro studies are not standardised and vary between research groups, making comparisons between already published observations difficult. Second, testing our model predictions requires measurements of single and double phosphorylation outputs, this could be problematic as antibody specificity required to distinguish these outputs might not readily be available. This would particularly be relevant for systems where phosphorylation sites are situated close together. Third, ensuring that the condition for observing hormesis e2, e4 > 0 is satisfied experimentally is challenging as kinase biochemical assays would not usually include phosphatase activity. Furthermore varying rates of reactions individually or measuring fluxes in such systems is equally difficult.
Having discussed difficulties associated with testing our model in reductionist in vitro systems, we next consider whether these difficulties could be overcome with a cell-based experimental systems. In particular, our model predicts that hormetic dose-response could be a wide-spread feature of MAPK pathways when exposed to enzyme inhibitors. However we argue here that the non-trivial biphasic dose-response associated with hormesis might often be overlooked when performing experiments at cell population level, as we now discuss.
Consider the case where in the absence of an inhibitor, the double phosphorylation motif Eq (1) possesses two stable steady states of the doubly-phosphorylated form of the substrate CPP. This means that tumour cells within a population can be grouped into two types: type-1 cells with ‘high’ CPP and type-2 cells with ‘low’ CPP. In reality these heterogeneous cell phenotypes can emerge not only due to multistability of the system [50, 54] but also due to stochastic fluctuations which lead to different concentrations of the the total protein substrate [66, 67]. In general, an untreated tumour is likely to harbour different proportions of cells in different phenotypic states [68].
We show that different cell types can respond differently to the presence of an inhibitor. Namely, our model predicts that in certain cases cells with initially high levels of CPP (at steady state CPP,1*) will exhibit a monotone decreasing dose-response (Fig 2A) while cells with low initial levels of CPP (at steady state CPP,2*) will exhibit a hormetic response (Fig 2B). This has an important consequence for measuring CPP at a population level as it is frequently done [69], as well as determining inhibitory concentrations (IC). Such consequences are best illustrated with the following example.
Let us assume, for example, that 88% of the tumour cells are type-1 cells and 12% of the tumour cells are type-2 cells. We can then simulate our model to generate dose response curves of CPP for both type-1 (Fig 5, green line) and type-2 (Fig 5, blue line) phenotypes. In addition, we can also numerically generate sampled values of the combined dose response of the entire population as would be measured, for instance, in a western blot or population-based imaging assay for C PP (Fig 5, red dots). By fitting a logistic curve to the sampled values of the combined dose response (Fig 5, red dashed line) we can estimate the inhibitor concentration causing 50% inhibition of the entire population, denoted IC50. However, the same inhibitor concentration has the opposing effects on the two sub-populations: while it inhibits type-1 cells, it actually stimulates type-2 cells. This can be observed by comparing steady-state values of CPP in the absence of inhibition (CPP,1* for type-1 and CPP,2* for type 2) to the steady-state values of CPP in the presence of the inhibitor (CPP,1** for type-1 and CPP,2** for type 2) at the IC50 concentration estimated for the entire population (Fig 5). In particular, the inhibition of type-1 cells can be seen from CPP,1*>CPP,1** while the stimulation of type-2 cells can be seen from CPP,2*<CPP,2**. Such unexpected stimulatory effects of the population-level IC50 exerted on type-2 sub-population could be further amplified when taking into account the imperfect drug penetration in a tumour [70]. In that case tumour cells would actually experience a lower inhibitor concentration IC*<IC50, which could lead to significant increases in steady-state values of CPP (denoted by CPP,2x in Fig 5), compared to the steady-state values of CPP in the absence of inhibition (denoted by CPP,2* in Fig 5). A numerical example with balanced type-1 and type-2 cell populations is presented in Fig A5 of S1 Appendix, showing that in this case it is also possible to mask the hormetic response at the population level, although the maximal hormetic response of the type-2 cells at the corresponding IC50 is substantially lower.
The presence of hormetic responses to an inhibitor which are masked at a population level could, therefore, complicate the interpretation of, and understanding gained from, preclinical models. Such complex sub-population effects have been noted for example in the NF-κB pathway, controlling DNA transcription, cytokine production and cell survival [71]. In particular, studies have shown that observing non-synchronous cells at a population level may under-represent oscillatory behaviour of nuclear shuttling [40, 72–74].
Examining hormetic responses at single-cell level could substantially improve detection rates as well as help identify mechanisms driving hormesis. However, while measuring and analysing single-cell bacterial dose response to antibiotics is already feasible [75], such methodology has rarely been implemented for studying dose-responses of tumour cells. Therefore, a wider application of single-cell dose-response techniques used for prokaryotes to tumour cells will greatly enhance our understanding of hormesis in cancer settings.
The conclusions of our study are based on the assumption that the dual phosphorylation-dephosphorylation motif presented in Eq (1) follows a distributive mechanism, whereby kinase (phosphatase) facilitates at most one phosphorylation (dephosphorylation) in each molecular encounter. This is motivated by the experimental evidence for MAPK pathways [51–53]. However, phosphorylation and dephosphorylation cycles can also follow a processive mechanism in which the kinase (phosphatase) facilitates two or more phosphorylations (dephosphorylations) before the final product is released [48]. In addition, a quasi-processive mechanism has been recently proposed to operate under the physiological condition of molecular crowding, which is a critical factor converting distributive into processive phosphorylation [76–78]. Our model can readily be extended to consider these alternative scenarios.
The findings presented here are relevant to applications in drug discovery relating to MAPK inhibition. Whereas inhibitors are specifically designed to target and suppress various stages in the MAPK pathways, the hormesis phenomenon leads to the opposite effect lowering the effectiveness of the compound and potentially leading to failure in the clinic [8, 9, 32]. Therefore, understanding mechanisms that lead to this undesired effect is important for designing inhibitors that would avoid them. Indeed, a recent study proposed a novel inhibitor, designed specifically to avoid MAPK activation at low-doses [79].
Our study could help achieve a similar goal. In particular, a straight forward approach to mitigate the risk of hormetic response is to favour inhibitor mechanisms of action for which this is impossible under our model. Protein substrate competitive inhibitors is one such example as these would generally, through steric hindrance, prohibit the formation of the necessary tertiary complex. In practice, structural biology can be employed to confirm that substrate and inhibitor complexes are mutually exclusive.
Overall, we argue that mathematical models are particularly useful tools in the drug-discovery process. Given the difficulties associated with measuring hormetic responses empirically be it with reductionist in vitro biochemical assays or cell based systems, the involvement of mathematical models in this process is of paramount importance. What we demonstrate here is that theoretical models classically make assumptions that immediately discount the possibility of observing hormetic responses in cell signalling pathways in the presence of inhibitors. Namely the assumption of no flux through substrate-kinase-inhibitor complex in motif Eq (2) is widespread in theoretical literature despite the lack of empirical support. It is, therefore, crucial that model assumptions are regularly challenged so that important behaviours are not overlooked.
|
10.1371/journal.pntd.0000917 | Revisiting the Immune Trypanolysis Test to Optimise Epidemiological Surveillance and Control of Sleeping Sickness in West Africa | Because of its high sensitivity and its ease of use in the field, the card agglutination test for trypanosomiasis (CATT) is widely used for mass screening of sleeping sickness. However, the CATT exhibits false-positive results (i) raising the question of whether CATT-positive subjects who are negative in parasitology are truly exposed to infection and (ii) making it difficult to evaluate whether Trypanosoma brucei (T.b.) gambiense is still circulating in areas of low endemicity. The objective of this study was to assess the value of the immune trypanolysis test (TL) in characterising the HAT status of CATT-positive subjects and to monitor HAT elimination in West Africa.
TL was performed on plasma collected from CATT-positive persons identified within medical surveys in several West African HAT foci in Guinea, Côte d'Ivoire and Burkina Faso with diverse epidemiological statuses (active, latent, or historical). All HAT cases were TL+. All subjects living in a nonendemic area were TL−. CATT prevalence was not correlated with HAT prevalence in the study areas, whereas a significant correlation was found using TL.
TL appears to be a marker for contact with T.b. gambiense. TL can be a tool (i) at an individual level to identify nonparasitologically confirmed CATT-positive subjects as well as those who had contact with T.b. gambiense and should be followed up, (ii) at a population level to identify priority areas for intervention, and (iii) in the context of HAT elimination to identify areas free of HAT.
| Human African trypanosomiasis (HAT) due to Trypanosoma brucei (T.b.) gambiense is usually diagnosed using two sequential steps: first the card agglutination test for trypanosomiasis (CATT) used for serological screening, followed by parasitological methods to confirm the disease. Currently, CATT will continue to be used as a test for mass screening because of its simplicity and high sensitivity; however, its performance as a tool of surveillance in areas where prevalence is low is poor because of its limited specificity. Hence in the context of HAT elimination, there is a crucial need for a better marker of contact with T.b. gambiense in humans. We evaluated here an existing highly specific serological tool, the trypanolysis test (TL). We evaluated TL in active, latent and historical HAT foci in Guinea, Côte d'Ivoire and Burkina Faso. We found that TL was a marker for exposure to T.b. gambiense. We propose that TL should be used as a surveillance tool to monitor HAT elimination.
| Human African trypanosomiasis (HAT) or sleeping sickness is caused by two subspecies of the protozoan flagellate Trypanosoma brucei. In West and Central Africa, T.b. gambiense causes the chronic form of sleeping sickness, while in East Africa, T.b. rhodesiense causes the more fulminant form [1]. T.b. brucei is normally not infectious to humans, like other species causing animal African trypanosomiasis (AAT) such as T. evansi, T. congolense, T. vivax and T. equiperdum.
After the successful control campaigns dating from 1930 to 1960, T.b. gambiense sleeping sickness re-emerged in the 1980s, with tens of thousands of cases treated every year. As a result of control activities, reported cases decreased to a mere 11,382 patients in 2006 [2] and to less than the symbolic number of 10,000 in 2009 [3]. However, along with decreasing incidence, disease control efforts may be discontinued, thus allowing the epidemic to build up again [2]. At present, two West African countries are endemic for HAT [2], [4], [5]. Guinea is the most affected with about 100 HAT cases reported annually from the coastal mangroves. In Côte d'Ivoire, control activities since the 1980s [6] have resulted in a low disease prevalence with a few tens of HAT cases annually, mainly from the Central West foci. In Togo, Ghana, Benin, Mali and Burkina Faso, no autochthonous cases have been reported over the last few years. Although the epidemiological situation remains unknown in several countries, including Liberia and Sierra Leone, HAT elimination in West Africa seems attainable.
Mass screening of the population at risk of T.b. gambiense is routinely performed using the card agglutination test for trypanosomiasis (CATT) on select individuals with antibodies against trypanosome antigens. CATT consists of bloodstream form trypomastigotes of T.b. gambiense variable antigen type (VAT) LiTat 1.3 purified from infected rat blood, fixed, stained and lyophilised [7]. When a drop of CATT reagent on a plastic card is mixed for 5 min with a drop of blood or diluted plasma or serum, the trypanosomes are agglutinated by antibodies that bind to the surface of the fixed cells resulting in a macroscopic agglutination reaction. Most of these antibodies will react with the VAT-specific epitopes on the cells. These highly immunogenic epitopes are present on the surface-exposed part of the densely packed variant surface glycoproteins (VSG). On living trypanosomes, only these VAT-specific epitopes are accessible for antibody binding. During the production of CATT reagent part of the VSG coat is shed and other epitopes on the VSG molecules that are not strictly VAT-specific, and from other surface proteins embedded between the VSGs, become available for antibody recognition and thus take part in the agglutination reaction [8]. This can lead to false-positive results, compromising the specificity of the test [9].
In the current elimination context in West Africa, when prevalence becomes low or transmission has stopped, the limited specificity of CATT becomes a considerable drawback because it results in low positive predictive values [10]–[12]. Recognising parasitologically unconfirmed but infected CATT-positive cases between many false-positives becomes problematic, since untreated, they may act as a reservoir. Molecular methods such as polymerase chain reaction (PCR), PCR-oligochromatography, NASBA-oligochromatography, real-time PCR and loop-mediated isothermal amplification method (LAMP) have been developed partly to resolve the problem of these unconfirmed CATT-positive subjects, but they also suffer from limited sensitivity, uncertain specificity and poor reproducibility depending on the genome sequence targeted [13]–[15].
In a previous study, the immune trypanolysis test (TL) was shown to be a promising tool to help better understand the phenomenon of nonconfirmed CATT-positive subjects [11]. Trypanosomes are able to change the VAT of their VSG by antigenic variation [16], [17]. During an infection, the host mounts an antibody responses against a variety of VATs [18]. Some VATs are expressed in T.b. gambiense (LiTat 1.3, LiTat 1.5 and LiTat 1.6), others in T.b. rhodesiense (ETat 1.2). Detailed studies of trypanosome VAT repertoires have been possible by introducing the TL, which consists of a suspension in complement-rich cavia serum of cloned bloodstream form trypomastigotes, all expressing the same VAT, incubated at 37°C with a test serum. Whenever serum antibodies bind to the VAT-specific epitopes on the trypanosome surface, the cells are lysed through antibody-mediated complement lysis. The TL test is considered 100% specific since in the given test conditions, only VAT-specific antibodies are able to cause lysis of the trypanosomes and such antibodies are absent in noninfected persons [19]. VAT repertoire studies of different trypanosome strains have revealed that some VATs, called predominant VATs, are recognised by almost all gambiense sleeping sickness patients, although exceptions do occur. Thus the VAT LiTat 1.3, corresponding to the main CATT antigen, is recognised by all the patients in Côte d'Ivoire while LiTat 1.5, representing a different VSG type, is recognised by almost all Nigerian patients. The combination of VATs LiTat 1.3+ LiTat 1.5+ LiTat 1.6 was able to detect 97% of the patients from eight different countries [19].
The objective of the present study was to evaluate the use of TL with T.b. gambiense VATs LiTat 1.3, LiTat 1.5 and LiTat 1.6 aiming at improved epidemiological surveillance of sleeping sickness in West Africa, with special interest in CATT-seropositive persons. TL was performed on plasma collected during medical surveys in several West African HAT foci. Our results argue that TL could be used (i) as a tool to identify CATT-positive subjects who experienced contact with T.b. gambiense and (ii) as a surveillance tool to monitor HAT elimination.
All samples were collected within the framework of medical surveys conducted by the national HAT control programmes (NCP) according to the respective national HAT diagnostic procedures. No samples other than those collected for routine screening and diagnostic procedures were collected for the purposes of the present study. All participants were informed of the objective of the study in their own language and signed a written informed consent form. Children less than 12 years old were excluded. For participants between 12 and 18 years of age, informed consent was obtained from their parents. This study is part of a larger project aiming at improving HAT diagnosis for which approval was obtained from WHO (Research Ethics Review Committee) and Institut de Recherche pour le Développement (Comité Consultatif de Déontologie et d'Ethique) ethical committees.
Specimens were collected in three West African countries (Guinea, Côte d'Ivoire and Burkina Faso) in foci with a different epidemiological HAT status (Figure 1).
All persons participating in the study were identified during active screening campaigns organised by the NCPs in Guinea, Burkina Faso and Côte d'Ivoire during HAT surveillance activities. Only subjects positive to the CATT/T.b. gambiense (CATT-B) performed on blood collected by finger prick and who had never received HAT-specific treatment were included in the study. For CATT-B-positive persons, blood was collected in heparinised tubes and a twofold plasma dilution series in CATT buffer was tested to assess the end titre, i.e. the highest dilution still positive (CATT-P). All persons included in the study underwent parasitological examinations by direct examination of the lymph node aspirate and/or mini-anion exchange centrifugation technique (mAECT) on blood [26]. Thus, three categories of study participants were defined for the purposes of the study:
HAT (patients): CATT-P end titer ≥1/8 and parasitologically confirmed;
SERO (seropositives): CATT-P end titer ≥1/8 but no parasites detected;
SUSP (suspects): CATT-B-positive and CATT-P <1/8 but no parasites detected.
The origin and numbers of participants in each group are detailed in Table 1. Left-over plasma specimens from the subjects were kept at −20°C during field activities, stored at −80°C in the Centre de Recherche Développement sur l′Elevage en zone Subhumide (CIRDES, Bobo-Dioulasso, Burkina-Faso) and sent on dry ice to the Institute of Tropical Medicine (ITM, Antwerpen, Belgium) where TL was performed blindly.
Cloned populations of T.b. gambiense VATs LiTat1.3, LiTat 1.5 and LiTat 1.6 and one T.b. rhodesiense VAT ETat 1.2R were used to test plasma as previously described [19]. Briefly, 25 µl of plasma was mixed with an equal volume of guinea pig serum, to which 50 µl of a 107 trypanosomes/ml suspension prepared from infected mouse blood was added. After 90 min of incubation at room temperature, the suspension was examined by microscopy (×250). Trypanolysis was considered positive when more than 50% of the trypanosomes were lysed. ETat 1.2R was a control for the absence of nonspecific trypanolytic activity of the test plasma. Plasma were considered positive (TL+) if positive with at least one of the three variants.
A total of 43,373 persons were screened with CATT/T.b. gambiense (Table 2). The highest HAT prevalence was observed in Dubreka-Boffa. Low prevalence was observed in Forécariah and Bonon. No patients were detected at the three study sites in Burkina Faso (Fol/Lor/Bat) and N'Zérékoré in Guinea. Seroprevalence of CATT-B and CATT-P end titer 1/8 or higherranged from 0.95 to 3.87% and from 0.31 to 1.21%, respectively, and were not associated with disease prevalence (CATT-B: r2 = 0.05, p = 0.91; CATT-P end titer ≥1/8: r2 = 0.26, p = 0.37).
The results of TL on the 287 subjects included in the study are summarised in Table 3. No serum lysed T.b. rhodesiense VAT ETat 1.2R, indicating the absence of non-antibody-related trypanolytic factors in the plasma samples.
All 71 HAT patients were TL+. All 18 SUSP from For/Lor/Bat in Burkina Faso were TL− while in Bonon 10 of 86 (11.6%) were TL+. Among the SERO subjects, 41 of 112 (36.6%) were TL+. Interestingly, the percentage of TL+ subjects in the SERO group was correlated with HAT prevalence (r2 = 0.84, p = 0.03) (Figure 2). It was the highest in the epidemic context of Dubreka/Boffa (15/17, 88.2%), lower in areas of lower HAT prevalence (18/30, 60% and 7/24, 29.2% in Forécariah and Bonon, respectively), whereas no TL+ subjects were detected in areas were HAT was absent, except for one subject in the Fol/Lor/Bat area in Burkina Faso. Table 4 presents the TL results per VAT for TL+ persons by study site. In Bonon, all HAT and SERO subjects were positive for all VATs. Among the ten SUSP, six were also positive for all VATs, whereas four showed different profiles. In Guinea, all 58 of 58 HAT were positive for LiTat 1.3 and LiTat 1.5 and only 38 of 58 were positive in LiTat 1.6. The same trends were observed for SERO: 32 of 33 were positive for LiTat 1.3 and LiTat 1.5 and only 12 of 33 were positive in LiTat 1.6.
This study shows that high prevalence of CATT-positive individuals can be found even in areas were transmission has stopped, presumably owing to false positivity. On the contrary, positivity of TL in SERO subjects was significantly correlated with HAT prevalence and not in nonendemic areas. Thus TL is a useful tool, both to define the epidemiological status of an area when no HAT cases are diagnosed and to improve the monitoring of CATT-positive subjects with no parasitological confirmation, who are currently left out of HAT control strategies in most endemic countries.
The HAT prevalence rates observed in this study are in agreement with recent data on HAT epidemiology in West Africa. Guinea was the most affected country, with 0.57% HAT prevalence in Dubreka-Boffa and 0.16% in the Forécariah focus. No HAT cases were diagnosed in the N'Zérékoré focus. With 0.18% prevalence, HAT is still endemic in the Bonon focus in Côte d'Ivoire. The disease did not re-emerge in the historical foci of Burkina Faso despite the return of agricultural workers from active HAT foci in Côte d'Ivoire since 2002 [25]. In West Africa, areas with disease prevalence approaching zero are becoming common, a recent trend observed in savannah areas [4]. In such areas, CATT-seropositive but parasitologically unconfirmed persons are encountered, making it difficult to evaluate the epidemiological status of the area and to determine what control measures should be applied at both the population and individual levels. This is clearly illustrated by the fact that SERO persons were found in all study sites but their number was not correlated with HAT prevalence (Figure 2). In the historical foci of Burkina Faso and in N'Zérékoré where transmission has stopped, SERO persons may be regarded as false-positives. Aspecific reactions in CATT may have different causes [14] such as cross-reactions with other infectious diseases or transient infections with T. b. brucei. Interestingly, the proportion of SERO subjects is highest in Bonon where pig breeding is widespread and where the prevalence of T.b. brucei in domestic pigs was reported to be around 70% [27]. Wild fauna and T.b. brucei are still present in South-west Burkina Faso [24], where the prevalence of SERO observed in this study is also relatively high. On the other hand, few domestic animals are kept in the Dubreka/Boffa and Forécariah foci in Guinea, where the proportion of SERO is the lowest but which display the highest HAT prevalence. Thus, although CATT is a good serological test for active screening, CATT seropositivity prevalence is not correlated with HAT prevalence, and CATT is not specific enough to evaluate whether T.b. gambiense is still circulating in a given area, which is of paramount importance in a disease elimination context.
TL was found to be highly sensitive (100% of HAT cases were TL+). Among SUSP and SERO persons, TL+ individuals were only found in areas with proven transmission, except one SERO person in the Fol/Lor/Bat focus in Burkina Faso, which is no longer active. It is noteworthy that this person had worked for 4 years in coffee and cacao plantations in a known HAT focus in Côte d'Ivoire where he may have been exposed to T.b. gambiense. Furthermore, a significant correlation was found between the percentage of TL+ SERO persons and the observed HAT prevalence. Our data therefore indicate that TL is a better marker of exposure to T.b. gambiense than CATT. The higher specificity of TL observed in this study is explained by the fact that only VAT-specific epitopes can react with antibodies in this test format. Other studies indicated that TL can be more sensitive than CATT since parallel testing with several VATs (LiTat 1.3, 1.5 and 1.6) may reveal infections by T.b. gambiense strains not expressing LiTat 1.3, the VAT used for CATT preparation [19], [28]. In addition, TL is based on antibody-mediated complement lysis and can therefore detect much lower antibody concentrations than CATT, which is based on agglutination reactions (unpublished data).
Assuming TL is a marker of exposure to T.b. gambiense, the existence in HAT endemic areas of persons harboring T.b. gambiense-specific antibodies detectable by TL but without detectable parasites may be explained by one of several hypotheses:
Concordant with TL positivity being a marker of contact with T.b. gambiense is the fact that lysis profiles to the different LiTat VATs tested were similar in HAT patients and SERO individuals in the different endemic areas. In Côte d'Ivoire, all SERO and HAT patients were positive for all three LiTat VATs. This was not the case in Guinea where only a fraction of HAT patients were positive for LiTat 1.6, as observed in SERO individuals.
At the individual level, TL can represent a tool for NCPs to identify among CATT-positives those who should be followed up by CATT and parasitological investigations until CATT becomes negative or the person is confirmed as a HAT patient. Whether these seropositives should undergo treatment remains an open question as long as their role in HAT transmission is unknown. We are currently carrying out follow-ups of SERO subjects. Preliminary results indicate that SERO TL+ individuals maintain a strong serological response (CATT and TL) over time, whereas SERO TL− subjects become CATT-negative within several months. Furthermore, HAT patients confirmed during these follow-ups were all from the SERO TL+ cohort (Bucheton, personal communication). In some countries, treatment of unconfirmed persons with CATT-P titers ≥1/16 is already recommended [10]. TL on these persons may avoid unnecessary treatments, as suggested by the nine, five and eight persons in Côte d'Ivoire, Guinea and Burkina Faso, respectively, who had CATT-P titers ≥1/16 but were TL− (data not shown).
At the population level, TL performed on CATT-positive individuals could be a valuable decision tool for NCPs to plan control measures (Figure 3). In active HAT foci, the priority is cutting transmission through active screening and vector control, thus SERO TL+ individuals should be monitored and treated when they become parasitologically positive. Also in areas without HAT, the presence of SERO TL+ cases should sound an alarm, since they indicate contact with T.b. gambiense. Continued surveillance of such areas is therefore strongly indicated. In areas without HAT and without SERO TL+ cases, HAT transmission may be considered absent and surveillance may be suspended unless a special event, such as population movement, occurs.
From a practical point of view, the implementation of TL in NCP is hampered by its technological requirements (cryobiology and laboratory animal facilities, availability of VAT-specific control sera, etc.). An alternative test that is applicable in the field and that allows combining several VATs in a single test is the indirect agglutination test LATEX/T.b. gambiense [31]. Unfortunately, the purified native VSGs used as antigens in the LATEX/T.b. gambiense bear non-VAT-specific epitopes that can lead to false-positive reactions as in the CATT. Investigations to eliminate these non-VAT-specific epitopes in rapid diagnostic tests for HAT are ongoing. In the meantime, adaptation of TL for testing blood collected on filter paper is underway. This would facilitate specimen storage and shipment from the field to the laboratory, as was done for T. evansi [32]. Furthermore, as in West Africa almost all TL+ persons are positive in LiTat 1.3, TL with this VAT alone may be sufficient for surveillance purposes in this region. During a WHO meeting held in Bamako in June 2009 with representatives of disease-endemic countries and partners involved in HAT control in West Africa, sleeping sickness control managers welcomed the performance of TL and stated their willingness to collect plasma specimens from SERO cases detected during medical surveys, and send them to CIRDES where TL is now available.
In conclusion, application of the TL test within the framework of medical surveys has provided a better picture of HAT epidemiology in West Africa, thanks to a better characterisation of parasitologically unconfirmed CATT positive subjects. The proportion of TL+ subjects among CATT+ individuals was associated with active HAT foci and can thus be used as a marker for exposure to T.b. gambiense even in areas were no HAT cases are currently being diagnosed. TL could thus be regarded as a decision tool for NCPs to decide when surveillance or disease control should be stopped in a given area. The results presented here encourage further investigations, including other endemic countries; into the significance of TL as a tool for improve the knowledge of HAT epidemiology and control.
|
10.1371/journal.pbio.2006842 | Gut microbiota diversity across ethnicities in the United States | Composed of hundreds of microbial species, the composition of the human gut microbiota can vary with chronic diseases underlying health disparities that disproportionally affect ethnic minorities. However, the influence of ethnicity on the gut microbiota remains largely unexplored and lacks reproducible generalizations across studies. By distilling associations between ethnicity and differences in two US-based 16S gut microbiota data sets including 1,673 individuals, we report 12 microbial genera and families that reproducibly vary by ethnicity. Interestingly, a majority of these microbial taxa, including the most heritable bacterial family, Christensenellaceae, overlap with genetically associated taxa and form co-occurring clusters linked by similar fermentative and methanogenic metabolic processes. These results demonstrate recurrent associations between specific taxa in the gut microbiota and ethnicity, providing hypotheses for examining specific members of the gut microbiota as mediators of health disparities.
| Understanding microbiota similarities and differences across ethnicities has the potential to advance approaches aimed at personalized microbial discovery and treatment, particularly those involved in ethnic health disparities. Here, we explore whether or not self-declared ethnicity consistently varies with gut microbiota composition across 1,673 healthy individuals in the United States. We find subtle but significant differences in taxonomic composition between four ethnicities, and we replicate these results across two study populations. Within the gut microbiota of Americans, there are at least 12 microbial taxa, which reproducibly vary in abundance across ethnicities. These taxa tend to correlate in abundance and metabolic functions and overlap with previously identified taxa that are associated with human genetic variation. We discuss the roles these taxa play in digestion and disease and propose hypotheses for how they may relate to ethnic health disparities. This study highlights the need to consider and potentially account for ethnic diversity in microbiota research and therapeutics.
| The human gut microbiota at fine resolution varies extensively between individuals [1–3], and this variability frequently associates with diet [4–7], age [6, 8, 9], sex [6, 9, 10], body mass index (BMI) [1, 6], and diseases presenting as health disparities [11–14]. The overlapping risk factors and burden of many chronic diseases disproportionally affect ethnic minorities in the United States, yet the underlying biological mechanisms mediating these substantial disparities largely remain unexplained. Recent evidence is consistent with the hypothesis that ethnicity associates with variation in microbial abundance, specifically in the oral cavity, gut, and vagina [15–17]. To varying degrees, ethnicity can capture many facets of biological variation including social, economic, and cultural variation, as well as aspects of human genetic variation and biogeographical ancestry. Ethnicity also serves as a proxy to characterize health disparity incidence in the US, and while factors such as genetic admixture create ambiguity of modern ethnic identity, self-declared ethnicity has proven a useful proxy for genetic and socioeconomic variation in population scale analyses, including in the Human Microbiome Project (HMP) [18–20]. Microbiota differences have been documented across populations that differ in ethnicity as well as in geography, lifestyle, and sociocultural structure; however, these global examinations cannot disconnect factors such as intercontinental divides and hunter–gatherer versus western lifestyles from ethnically structured differences [21–23]. Despite the importance of understanding the interconnections between ethnicity, microbiota, and health disparities, there are no reproducible findings about the influence of ethnicity on differences in the gut microbiota and specific microbial taxa in diverse US populations, even for healthy individuals [6].
Here, we comprehensively examine connections between self-declared ethnicity and gut microbiota differences across more than a thousand individuals sampled by the American Gut Project (AGP, N = 1375) [24] and the HMP (N = 298) [6]. Previous studies demonstrated that human genetic diversity in the HMP associates with differences in microbiota composition [25], and genetic population structure within the HMP generally delineates self-declared ethnicity [20]. Ethnicity was not found to have a significant association with microbiota composition in a Middle Eastern population; however, factors such as lifestyle and environment that influence microbiota variation across participants was homogenous compared to the ethnic, sociocultural, economic, and dietary diversity found within the US [26]. While ethnic diversity is generally under-represented in current microbiota studies, evidence supporting an ethnic influence on microbiota composition among first generation immigrants has been recently demonstrated in a Dutch population [27]. The goal of this examination is to evaluate, for the first time, if there are reproducible differences in gut microbiota across ethnicities within an overlapping US population, as ethnicity is one of the key defining factors for health disparity incidence in the US. Lifestyle, dietary, and genetic factors all vary to different degrees across ethnic groups in the US, and it will require more even sampling of ethnic diversity and stricter phenotyping of study populations to disentangle which factors underlie ethnic microbiota variation in the AGP and HMP.
We first evaluate gut microbiota distinguishability between AGP ethnicities (Fig 1A, family taxonomic level, Asian-Pacific Islanders [N = 88], Caucasians [N = 1237], Hispanics [N = 37], and African Americans [N = 13]), sexes (female [N = 657], male [N = 718]), age groups (years grouped by decade), and categorical BMI (underweight [N = 70], normal [N = 873], overweight [N = 318], and obese [N = 114]) (Demographic details in S1A Table). Age, sex, and BMI were selected as covariates because they are consistent across the AGP and HMP data sets. Additionally, 31 other AGP categorical factors measuring diet, environment, and geography were compared for pairwise differences between two ethnicities using proportions tests, and very few (10/894) tests significantly varied (S1 Table additional sheets). Interindividual gut microbiota heterogeneity clearly dominates; however, analyses of similarity (ANOSIM) reveal subtle but significant degrees of total microbiota distinguishability for ethnicity, BMI, and sex but not for age (Fig 1B, Ethnicity; Fig 1C, BMI; Fig 1D, Sex; Fig 1E, Age) [28]. Recognizing that subtle microbiota distinguishability between ethnicities may be spurious, we independently replicate the ANOSIM results from HMP African Americans (N = 10), Asians (N = 34), Caucasians (N = 211), and Hispanics (N = 43) (S2A Table, R = 0.065, p = 0.044). We again observe no significant distinguishability for BMI, sex, and age in the HMP. Higher rarefaction depths increase microbiota distinguishability in the AGP across various beta diversity metrics and categorical factors (S2B Table), and significance increases when individuals from over-represented ethnicities are subsampled from the average beta diversity distance matrix (S2C Table). Supporting the ANOSIM results, Permutational Multivariate Analysis of Variance (PERMANOVA) models with four different beta diversity metrics showed that while all factors had subtle but significant associations with microbiota variation when combined in a single model, effect sizes were highest for ethnicity in seven out of eight comparisons across beta diversity metrics and rarefaction depths in the AGP and HMP (S2D Table). We additionally test microbiota distinguishability by measuring the correlation between beta diversity and ethnicity, BMI, sex, and age with an adapted BioEnv test (S2E Table) [29]. Similar degrees of microbiota structuring occur when all factors are incorporated (Spearman Rho = 0.055, p-values: Ethnicity = 0.057, BMI < 0.001, Sex < 0.001, Age = 0.564). Firmicutes and Bacteroidetes dominated the relative phylum abundance, with each representing between 35% and 54% of the total microbiota across ethnicities (S1 Fig).
We next test for ethnicity signatures in the gut microbiota by analyzing alpha and beta diversity, abundance and ubiquity distributions, distinguishability, and classification accuracy [30]. Shannon’s Alpha Diversity Index [31], which weights both microbial community richness (observed operational taxonomic units [OTUs]) and evenness (Equitability), significantly varies across ethnicities in the AGP data set (Kruskal–Wallis, p = 2.8e-8) with the following ranks: Hispanics > Caucasians > Asian-Pacific Islanders > African Americans (Fig 2A). In the HMP, there is a significantly lower Shannon diversity for Asian-Pacific Islanders relative to Caucasians and a trend of lower Shannon diversity for Asian-Pacific Islanders relative to Hispanics; African Americans change position in diversity relative to other ethnicities, potentially as a result of undersampling bias. Five alpha diversity metrics, two rarefaction depths, and separate analyses of Observed OTUs and Equitability generally confirm the results (S3A Table).
If ethnicity impacts microbiota composition, pairwise beta diversity distances (ranging from 1/completely dissimilar to 0/identical) will be greater between ethnicities than within ethnicities. While average gut microbiota beta diversities across all individuals are high (Bray–Curtis = 0.808), beta diversities between individuals of the same ethnicity (intraethnic, Bray–Curtis = 0.806) are subtly but significantly lower than those between ethnicities in both the AGP (interethnic, Bray–Curtis = 0.814) and HMP data sets (intraethnic, Bray–Curtis = 0.870 versus interethnic, Bray–Curtis = 0.877) (Fig 2B). We confirm AGP results by subsampling individuals from over-represented ethnicities across beta metrics and rarefaction depths (S4A and S4B Table). Finally, we repeat analyses across beta metrics and rarefaction depths using only the average distance of each individual to all individuals from the ethnicity to which they are compared (S4C and S4D Table).
Next, we explore interethnic differences in the number of OTUs shared in at least 50% of individuals within an ethnicity, as the likelihood of detecting a biological signal is improved in more abundant organisms relative to noise that may predominate in lower abundance OTUs. Out of 5,591 OTUs in the total AGP data set, 101 (1.8%) OTUs meet this ubiquity cutoff in all ethnicities, and 293 (5.2%) OTUs meet the cutoff within at least one ethnicity. Hispanics share the most ubiquitous OTUs and have the lowest average abundance/ubiquity (A/U) ratio (Fig 2C), indicating stability, whereby stability represents a more consistent appearance of OTUs with lower abundance but higher ubiquity [32]. This result potentially explains their significantly lower intraethnic beta diversity distance and thus higher microbial community overlap relative to the other ethnicities (Fig 2B). Comparisons in the AGP between the higher sampled Hispanic, Caucasian, and Asian-Pacific Islander ethnicities also reveal a trend wherein higher intraethnic community overlap (Fig 2B) parallels higher numbers of ubiquitous OTUs (Fig 2C), higher Shannon alpha diversity (Fig 2A), and higher stability of ubiquitous OTUs as measured by the A/U ratio (Fig 2C).
We next assess whether a single ethnicity disproportionately impacts total gut microbiota distinguishability in the AGP by comparing ANOSIM results from the consensus beta diversity distance matrix when each ethnicity is sequentially removed from the analysis (Fig 3A and S2E Table). Distinguishability remains unchanged when the few African Americans are removed but is lost upon removal of Asian-Pacific Islanders or Caucasians, likely reflecting their higher beta diversity distance from other ethnicities (Fig 3A). Notably, removal of Hispanics increases distinguishability among the remaining ethnicities, which may be due to a higher degree of beta diversity overlap observed between Hispanics and other ethnicities (S4B Table). Results conform across rarefaction depths and beta diversity metrics (S2F Table), and pairwise combinations show strong distinguishability between African Americans and Hispanics (ANOSIM, R = 0.234, p = 0.005) and Asian-Pacific Islanders and Caucasians (ANOSIM, R = 0.157, p < 0.001).
Finally, to complement evaluation with ecological alpha and beta diversity, we implement a random forest (RF) supervised learning algorithm to classify gut microbiota from genus-level community profiles into their respective ethnicity. We build four one-versus-all binary classifiers to classify samples from each ethnicity compared to the rest and use two different sampling approaches to train the models synthetic minority oversampling technique (SMOTE) [33] and downsampling for overcoming uneven representation of ethnicities in both the data sets (see Materials and methods). Given that the area under the receiver operating characteristic (ROC) curve (or AUC) of a random guessing classifier is 0.5, the models classify each ethnicity fairly well (Fig 3B), with average AUCs across sampling techniques and data sets of 0.78 for Asian-Pacific Islanders, 0.76 for African Americans, 0.69 for Hispanics, and 0.70 for Caucasians. Ethnicity distinguishing RF taxa and out-of-bag error percentages appear in (S2 Fig).
Subtle to moderate ethnicity-associated differences in microbial communities may in part be driven by differential abundance of certain microbial taxa. 16.2% (130/802) of the AGP taxa and 20.6% (45/218) of HMP taxa across all classification levels (i.e., phylum to genus, S5 Table) significantly vary in abundance across ethnicities (Kruskal–Wallis, pFDR < 0.05). Between data sets, 19.2% (25/130) of the AGP and 55.6% (25/45) of the HMP varying taxa replicate in the other data set, representing a significantly greater degree of overlap than would be expected by chance (ethnic permutation analysis of overlap, p < 0.001 each taxonomic level and all taxonomic levels combined). The highest replication of taxa varying by abundance occurs with 22.0% of families (nine significant in both data sets / 41 significantly varying families in either data set), followed by genus with 13.4% (nine significant in both data sets / 67 significantly varying genera in either data set).
Among 18 reproducible taxa, we categorize 12 as taxonomically distinct (Fig 4) and exclude six in which nearly identical abundance profiles between family/genus taxonomy overlap. Comparing relative abundance differences between pairs of ethnicities for these 12 taxa in the AGP reveals 30 significant differences, of which 20 replicate in the HMP (p < 0.05, Mann–Whitney U). Intriguingly, all reproducible pairwise differences are a result of decreases in Asian-Pacific Islanders (Fig 4). We also test taxon abundance and presence/absence associations with ethnicity separately in the AGP using linear and logistic regression models, respectively, and we repeat the analysis while incorporating categorical sex and continuous age and BMI as covariates (S6 Table). Clustering microbial families based on their abundance correlation reveals two co-occurrence clusters: (i) a distinct cluster of six Firmicutes and Tenericutes families in the HMP and (ii) an overlapping but more diverse cluster of 20 families in the AGP (S3 Fig). Nine of the 12 taxa found to recurrently vary in abundance across ethnicities are represented in these clusters (Fig 4), with four appearing in both clusters and the other five appearing either in or closely correlated with members of both clusters (S3 Fig). Furthermore, 90% (18/20) of families in the AGP cluster and 66% (4/6) of taxa in the HMP cluster significantly vary in abundance across ethnicities. We also found overlap for AGP and HMP data sets between taxa significantly varying in abundance across ethnicities (with false discovery rate [FDR] < 0.05) and taxa in RF models with percentage importance greater than 50% for an ethnicity (S2B Fig). Taken together, these results establish general overlap of the most significant ethnicity-associated taxa between these methods, reproducibility of microbial abundances that vary between ethnicities across data sets, and patterns of co-occurrence among these taxa, which could suggest they are functionally linked.
Identified as the most heritable taxon in the human gut [34, 35], the family Christensenellaceae exhibits the second strongest significant difference in abundance across ethnicities in both AGP and HMP data sets (S5 Table, Family: AGP, Kruskal–Wallis, pFDR = 1.55e-9; HMP, Kruskal–Wallis, pFDR = 0.0019). Additionally, Christensenellaceae is variable by sex and BMI (AGP: Sex, Kruskal–Wallis, pFDR = 1.22e-12; BMI, Kruskal–Wallis, pFDR = 0.0020) and represents some of the strongest pairwise correlations with other taxa in both co-occurrence clusters (S3 Fig). There is at least an eight-fold and two-fold reduction in average Christensenellaceae abundance in Asian-Pacific Islanders relative to the other ethnicities in the AGP and HMP, respectively (S5 Table), and significance of all pairwise comparisons in both data sets show reduced abundance in Asian-Pacific Islanders (Fig 4). Christensenellaceae also occurs among the top 10 most influential taxa for distinguishing Asian-Pacific Islanders from other ethnicities using RF models for both AGP and HMP data sets (S2A Fig). Abundance in individuals possessing Christensenellaceae and presence/absence across all individuals significantly associate with ethnicity (S6 Table, Abundance, Linear Regression, pBonferroni = 0.006; Presence/Absence, Logistic Regression, pBonferroni = 8.802e-6), but there was only a slight correlation between the taxon’s relative abundance and BMI (S4 Fig). Confirming previous associations with lower BMI [36], we observe that AGP individuals with Christensenellaceae also have a lower BMI (Mean BMI, 23.7 ± 4.3) than individuals without it (Mean BMI, 25.0 ± 5.9; Mann–Whitney U, p < 0.001). This pattern is separately reflected in African Americans, Asian-Pacific Islanders, and Caucasians but not Hispanics (Fig 5), suggesting that each ethnicity may have different equilibria between the taxon’s abundance and body weight.
Many factors associate with human ethnicity, including a small subset of population specific genetic variants (estimated approximately 0.5% genome wide) that vary by biogeographical ancestry [37, 38]; self-declared ethnicity in the HMP is delineated by population genetic structure [20]. Here, we investigate whether ethnicity-associated taxa overlap with (i) taxa that have a significant population genetic heritability in humans [34, 35, 39, 40] and (ii) taxa linked with human genetic variants in two large Genome-Wide Association Studies (GWAS)-microbiota analyses [35, 40]. All recurrent ethnicity-associated taxa except one were heritable in at least one study, with seven replicating in three or more studies (Table 1). Likewise, abundance differences in seven recurrent ethnicity-associated taxa demonstrate significant GWAS associations with at least one variant in the human genome. Therefore, we assess whether any genetic variants associated with differences in microbial abundance exhibit significant rates of differentiation (fixation index [FST]) between 1,000 genome superpopulations [38]. Out of 49 variants associated with ethnically varying taxa, 21 have higher FST values between at least one pair of populations than that of 95% of other variants on the same chromosome and across the genome; the FST values of five variants associated with Clostridiaceae abundance rank above the top 99% (S7 Table). Since taxa that vary across ethnicities exhibit lower abundance in Asian-Pacific Islanders, it is notable that the FST values of 18 and 11 variant comparisons for East Asian and South Asian populations, respectively, are above that of the 95% rate of differentiation threshold from African, American, or European populations. Cautiously, the microbiota and 1,000 genomes data sets are not drawn from the same individuals, and disentangling the role of genetic from social and environmental factors will still require more controlled studies.
Many common diseases associate with microbiota composition and ethnicity, raising the central hypothesis that microbiota differences between ethnicities can occasionally serve as a mediator of health disparities. Self-declared ethnicity in the US can capture socioeconomic, cultural, geographic, dietary, and genetic diversity, and a similarly complex array of interindividual and environmental factors influence total microbiota composition. This complexity may result in challenges when attempting to recover consistent trends in total gut microbiota differences between ethnicities. The challenges in turn emphasize the importance of reproducibility, both through confirmation across analytical methods and replication across study populations [15–17, 20, 27, 42]. In order to robustly substantiate the ethnicity–microbiota hypothesis, we evaluated recurrent associations between self-declared ethnicity and variation in both total gut microbiota and specific taxa in healthy individuals. Results provide hypotheses for examining specific members of the gut microbiota as mediators of health disparities.
Our findings from two American data sets demonstrate that (i) ethnicity consistently captures gut microbiota with a slightly stronger effect size than other variables such as BMI, age, and sex; (ii) ethnicity is moderately predictable from total gut microbiota differences; and (iii) 12 taxa recurrently vary in abundance between the ethnicities, of which the majority have been previously shown to be heritable and associated with human genetic variation. Whether shaped through socioeconomic, dietary, healthcare, genetic, or other ethnicity-related factors, reproducibly varying taxa represent sources for novel hypotheses addressing health disparities. For instance, the family Odoribacteriaceae and genus Odoribacter are primary butyrate producers in the gut, and they have been negatively associated to severe forms of Crohns disease and Ulcerative Colitis in association with reduced butyrate metabolism [43–45]. Asian-Pacific Islanders possess significantly less Odoribacteriaceae and Odoribacter than Hispanics and Caucasians in both data sets, and severity of Ulcerative Colitis upon hospital admission has been shown to be significantly higher in Asian Americans [46]. Considering broader physiological roles, several ethnicity-associated taxa are primary gut anaerobic fermenters and methanogens [47, 48] and associate with lower BMI and blood triglyceride levels [36, 49]. Indeed, Christensenellaceae, Odoribacteriaceae, Odoribacter, and the class Mollicutes containing RF39 negatively associate with metabolic syndrome and demonstrate significant population genetic heritability in twins [39]. Implications for health outcomes warrant further investigation but could be reflected by positive correlations of Odoribacteriaceae, Odoribacter, Coriobacteriaceae, Christensenellaceae, and the dominant Verrucomicrobiaceae lineage Akkermansia with old age [50, 51]. Akkermansia associations with health and ethnicity in Western populations may reflect recently arising dietary and lifestyle effects on community composition, as this mucus-consuming taxon is rarely observed in more traditional cultures globally [23]. Moreover, these findings raise the importance of controlling for ethnicity in studies linking microbiota differences to disease because associations between specific microbes and a disease could be confounded by ethnicity of the study participants.
Based on correlations in individual taxon’s abundance, a similar pattern of co-occurrence previously identified as the “Christensenellaceae Consortium” includes 11 of the 12 recurrent ethnically varying taxa [34], and members of this consortium associate with genetic variation in the human formate oxidation gene, aldehyde dehydrogenase 1 family member 1 (ALDH1L1), which is a genetic risk factor for stroke [35, 52, 53]. Formate metabolism is a key step in the pathway reducing carbon dioxide to methane [54, 55], and increased methane associates with increased Rikenellaceae, Christensenellaceae, Odoribacteriaceae, and Odoribacter [56]. Products of methanogenic fermentation pathways include short chain fatty acids such as butyrate, which, through reduction of proinflammatory cytokines, is linked to cancer cell apoptosis and reduced risk of colorectal cancer [57, 58]. Asian Americans are the only ethnic group where cancer surpasses heart disease as the leading cause of death, and over 70% of Asian Americans were born overseas, which can affect assimilation into Western lifestyles, leading to reduced access to healthcare and screening and proper medical education [57, 59–61]. Preliminary results from other groups suggest that the gut microbiome of Southeast Asian immigrants changes after migration to the US [62]. Indeed, as countries in Asia shift toward a more Western lifestyle, the incidence of cancers, particularly gastrointestinal and colorectal cancers, are increasing rapidly, possibly indicating incompatibilities between traditionally harbored microbiota and Western lifestyles [63–66]. Asian Americans have higher rates of type 2 diabetes and pathogenic infections than Caucasians [67], and two metagenomic functions enriched in control versus type 2 diabetes cases appear to be largely conferred by cluster-associated butyrate-producing and motility-inducing Verrucomicrobiaceae and Clostridia taxa reduced in abundance among AGP and HMP Asian-Pacific Islanders [11]. Both induction of cell motility and butyrate promotion of mucin integrity can protect against pathogenic colonization and associate with microbial community changes [11, 58, 68]. Levels of cell motility and butyrate are key factors suspected to underlie a range of health disparities including inflammatory bowel disease, arthritis, and type 2 diabetes [11, 69–71]. Patterns of ethnically varying taxa across ethnicities could result from many factors including varying diets, environmental exposures, sociocultural influences, human genetic variation, and others. However, regardless of the mechanisms dictating assembly, these results suggest that there is a reproducible, co-occurring group of taxa linked by similar metabolic processes known to promote homeostasis.
The utility of this work is establishing a framework for studying ethnicity-associated taxa and hypotheses of how changes in abundance or presence of these taxa may or may not shape health disparities, many of which also have genetic components. Differing in allele frequency across three population comparisons and associated with the abundance of Clostridiales, the genetic variant rs7587067 has a significantly higher frequency in African (minor allele frequency [MAF] = 0.802) versus East Asian (MAF = 0.190, FST = 0.54, Chromosome = 98.7%, Genome-Wide = 98.9%), admixed American (MAF = 0.278, FST = 0.44, Chromosome = 99.0%, Genome-Wide = 99.1%), and European populations (MAF = 0.267, FST = 0.45, Chromosome = 98.7.3%, Genome-Wide = 98.7%). This intronic variant for the gene HECW2 is a known expression quantitative trait locus (eQTL) (GTEx, eQTL Effect Size = -0.18, p = 7.4e-5) [72, 73], and HECW2 encodes a ubiquitin ligase linked to enteric gastrointestinal nervous system function through maintenance of endothelial lining of blood vessels [74, 75]. Knockout of HECW2 in mice reduced enteric neuron networks and gut motility, and patients with Hirschsprungs disease have diminished localization of HECW2 to regions affected by loss of neurons and colon blockage when compared to other regions of their own colon and healthy individuals [76]. Hirschsprungs disease presenting as full colon blockage is rare and has not undergone targeted examination as a health disparity; however, a possible hypothesis is that lower penetrance of the disease in individuals with the risk allele at rs7587067 could lead to subtler effects on gut motility resulting in Clostridiales abundance differences.
Despite the intrigue of connecting the human genome, microbiota, and disease phenotypes, evaluating such hypotheses will require more holistic approaches including incorporating metagenomics and metabolomics to identify whether enzymes or metabolic functions reproducibly vary across ethnicities, as well as direct functional studies in model systems to understand if correlation is truly driven by causation. Further limitations should also be considered, including recruitment biases for the AGP versus HMP, variation in sample processing and OTU clustering, and uneven sampling, which could only be addressed with downsampling of over-represented ethnicities. Still, despite these confounders, care was taken to demonstrate the reproducibility of results across statistical methods, ecological metrics, rarefaction depths, and study populations. Summarily, this work suggests that abundance differences of specific taxa, rather than whole communities, may represent the most reliable ethnic signatures in the gut microbiota. A reproducible co-occurring subset of these taxa link to a variety of overlapping metabolic processes and health disparities and contain the most reproducibly heritable taxon, Christensenellaceae. Moreover, a majority of the microbial taxa associated with ethnicity are also heritable and genetically associated taxa, suggesting that there is a possible connection between ethnicity and genetic patterns of biogeographical ancestry that may play a role in shaping these taxa. Our results emphasize the importance of sampling ethnically diverse populations of healthy individuals in order to discover and replicate ethnicity signatures in the human gut microbiota, and they highlight a need to account for ethnic variation as a potential confounding factor in studies linking microbiota differences to disease. Further reinforcement of these results may lead to generalizations about microbiota assembly and even consideration of specific taxa as potential mediators or treatments of health disparities.
Access to HMP data was obtained through dbGaP approval granted to SRB and AWB. Institutional Review Board approval was granted with nonhuman subjects determination IRB161231 by Vanderbilt University.
AGP data was obtained from the project FTP repository located at ftp://ftp.microbio.me/AmericanGut/. AGP data generation and processing prior to analysis can be found at https://github.com/biocore/American-Gut/tree/master/ipynb/primary-processing. All analyses utilized the rounds-1–25 data set, which was released on March 4, 2016. Throughout all analyses, QIIME v1.9.0 was used in an Anaconda environment (https://continuum.io) for all script calls, and custom scripts and notebooks were run in the QIIME 2 Anaconda environment with python version 3.5.2, and plots were postprocessed using Inkscape (https://inkscape.org/en/) [77]. Ethnicity used in this study was self-declared by AGP study participants as one of four groups: African American, Asian or Pacific Islander (Asian-Pacific Islander), Caucasian, or Hispanic. Sex was self-declared as either male, female, or other. Age was self-declared as a continuous integer of years old, and age categories defined by the AGP by decade (i.e., 20’s, 30’s, etc.) were used in this study. BMI was self-declared as an integer, and BMI categories defined by AGP of underweight, healthy, overweight, and obese were utilized. A total of 31 categorical metadata factors were assessed for structuring across ethnicities with a two proportion Z test between pairs of ethnicities using a custom python script (S1 Table additional sheets). The p-values were Bonferroni corrected within each metadata factor for the number of pairwise ethnic comparisons. 97% OTUs generated for each data set are utilized throughout to maintain consistency with other published literature; however, microbial taxonomy of the HMP is reassigned using the Greengenes reference database [78]. Communities characterized with 16S rDNA sequencing of variable region four followed an identical processing pipeline for all samples, which was developed and optimized for the Earth Microbiome Project [79]. HMP 16S rDNA data processed using QIIME for variable regions 3–5 was obtained from http://hmpdacc.org/HMQCP/. Demographic information for individual HMP participants was obtained through dbGaP restricted access to study phs000228.v2.p1, with dbGaP approval granted to SRB and nonhuman subjects determination IRB161231 granted by Vanderbilt University. Ethnicity and sex were assigned to subjects based on self-declared values, with individuals selecting multiple ethnicities being removed unless they primarily responded as Hispanic, while categorical age and BMI were established from continuous values using the same criteria for assignment as in the AGP. The HMP Amerindian population was removed due to severe under-representation. This filtered HMP table was used for community level analyses (ANOSIM, alpha diversity, beta intra-inter); however, to allow comparison with the AGP data set, community subset analyses (co-occurrence, abundance correlation, etc.) were performed with taxonomic assignments in QIIME using the UCLUST method with the GreenGenes_13_5 reference.
AGP quality control was performed in Stata v12 (StataCorp, 2011) using available metadata to remove samples (Raw N = 9,475) with BMI more than 60 (−988 [8,487]) or less than 10 (−68 [8,419]); missing age (−661 [7,758]), with age greater than 55 years old (−2,777 [4,981]) or less than 18 years old (−582 [4,399]); and blank samples or those not appearing in the mapping file (−482 [3,917]), with unknown ethnicity or declared as other (−131 [3786]), not declared as a fecal origin (−2,002 [1784]), with unknown sex or declared as other (−98 [1686]) or located outside of the US (−209 [1477]). No HMP individuals were missing key metadata or had other reasons for exclusion (−0[298]). Final community quality control for both the AGP and HMP was performed by filtering OTUs with less than 10 sequences and removing samples with less than 1,000 sequences (AGP, −102 [1375]; HMP, −0 [298]). All analyses used 97% OTUs generated by the AGP or HMP, and unless otherwise noted, results represent Bray–Curtis beta diversity and Shannon alpha diversity at a rarefaction depth of 1,000 counts per sample.
The ANOSIM test was performed with 9,999 repetitions on each rarefied table within a respective rarefaction depth and beta diversity metric (Fig 1 and S2A–S2B Table), with R values and p-values averaged across the rarefactions. Consensus beta diversity matrices were calculated as the average distances across the 100 rarefied matrices for each beta diversity metric and depth. Consensus distance matrices were randomly subsampled 10 times for subset number of individuals from each ethnic group with more than that subset number prior to ANOSIM analysis with 9,999 repetitions, and the results were averaged evaluating the effects of more even representations for each ethnicity (S2C Table). Consensus distance matrices had each ethnicity and pair of ethnicities removed prior to ANOSIM analysis with 9,999 repetitions, evaluating the distinguishability conferred by inclusion of each ethnicity (Fig 3A, S2F Table). Significance was not corrected for the number of tests to allow comparisons between results of different analyses, metrics, and depths. PERMANOVA analyses were run using the R language implementation in the Vegan package [80], with data handled in a custom R script using the Phyloseq package [81]. Categorical variables were used to evaluate the PERMANOVA equation (Beta Diversity Distance Matrix ~ Ethnicity + Age + Sex + BMI) using 999 permutations to evaluate significance, and the R and p-values were averaged across 10 rarefactions (S2D Table). The BioEnv test, or BEST test, was adapted to allow evaluation of the correlation and significance between beta diversity distance matrices and age, sex, BMI, and ethnicity simultaneously (S2E Table) [29]. At each rarefaction depth and beta diversity metric, the consensus distance matrix was evaluated for its correlation with the centered and scaled Euclidian distance matrix of individuals continuous age and BMI, and categorical ethnicity and sex encoded using patsy (same methodology as original test) (https://patsy.readthedocs.io/en/latest/#). The test was adapted to calculate significance for a variable of interest by comparing how often the degree of correlation with all metadata variables (age, sex, BMI, ethnicity) was higher than the correlation when the variable of interest was randomly shuffled between samples 1,000 times.
Alpha diversity metrics (Shannon, Simpson, Equitability, Chao1, Observed OTUs) were computed for each rarefied table (QIIME: alpha_diversity.py), and results were collated and averaged for each sample across the tables (QIIME: collate_alpha.py). Pairwise nonparametric t tests using Monte Carlo permutations evaluated alpha diversity differences between the ethnicities with Bonferroni correction for the number of comparisons (Fig 2A, S3 Table, QIIME: compare_alpha_diversity.py). A Kruskal–Wallis test implemented in python was used to detect significant differences across all ethnicities.
Each consensus beta diversity distance matrix had distances organized based on whether they represented individuals of the same ethnic group or were between individuals of different ethnic groups. All values indicate that all pairwise distances between all individuals were used (Fig 2B, S4A and S4B Table), and mean values indicate that for each individual, their average distance to all individuals in the comparison group was used as a single point to assess pseudo-inflation (S4C and S4D Table). A Kruskal–Wallis test was used to calculate significant differences in intraethnic distances across all ethnicities. Pairwise Mann–Whitney U tests were calculated between each pair of intraethnic distance comparisons, along with intra-versus-interethnic distance comparisons. Significance was Bonferroni corrected within the number of intra-intraethnic and intra-interethnic distance groups compared, with violin plots of intra- and interethnic beta diversity distances generated for each comparison.
RF models were implemented using taxa summarized at the genus level, which performed better compared to RF models using OTUs as features, both in terms of classification accuracy and computational time. We first rarefied OTU tables at a sequence depth of 10,000 (using R v3.3.3 package vegan’s rrarefy() function) and then summarized rarefied OTUs at the genus level (or a higher characterized level if genus was uncharacterized for an OTU). We filtered for rare taxa by removing taxa present in fewer than half of the number of samples in rarest ethnicity (i.e., fewer than 10/2 = 5 samples in HMP and 13/2 = 6 [rounded down] in AGP), retaining 85 distinct taxa in the HMP data set and 322 distinct taxa in the AGP data set at the genus level. The resulting taxa were normalized to relative abundance and arcsin-sqrt transformed before being used as features for the RF models. We initially built a multiclass RF model, but since the RF model is highly sensitive to the uneven representation of classes, all samples were identified as the majority class, i.e., Caucasian. In order to even out the class imbalance, we considered some sampling approaches, but most existing techniques for improving classification performance on imbalanced data sets are designed for binary class imbalanced data sets and are not effective on data sets with multiple under-represented classes. Hence, we adopted the binary classification approach and built four one-versus-all binary RF classifiers to classify samples from each ethnicity compared to the rest. 10-fold cross-validation (using R package caret [82]) was performed using ROC as the metric for selecting the optimal model. The performance metrics and ROC curves were averaged across the 10 folds (Fig 3B). Without any sampling during training the classifiers, most samples were identified as the majority class, i.e., Caucasian, by all four one-versus-all RF classifiers. In order to overcome this imbalance in class representation, we applied two sampling techniques inside cross-validation: i) downsampling and ii) SMOTE [33]. In the downsampling approach, the majority class is downsampled by random removal of instances from the majority class. In the SMOTE approach, the majority class is downsampled, and synthetic samples from the minority class are generated based on the k-nearest neighbors technique [33]. Note, the sampling was performed inside cross-validation on training set, while the test was performed on unbalanced held-out test set in each fold. In comparison to a no-sampling approach, which classified most samples as the majority class, i.e., Caucasians, our sampling-based approach leads to improved sensitivity for classification of minority classes on unbalanced test sets. Nevertheless, the most accurate prediction remains for the inclusion in the majority class. The ROC curves and performance metrics table in Fig 3B show the sensitivity–specificity tradeoff and classification performance for one-versus-all classifier for each ethnicity for both the sampling techniques applied on both of the data sets. For both of the data sets, downsampling shows higher sensitivity and lower specificity and precision for minority classes (i.e., African Americans, Asian-Pacific Islanders, and Hispanics) compared to SMOTE. However, for the majority class (i.e., Caucasian), downsampling lowers the sensitivity and increases the specificity and precision compared to SMOTE. The sensitivity–specificity tradeoff, denoted by the AUC, is reduced for Hispanics in both the data sets. The most important taxa with >50% importance for predicting an ethnicity using RF model with SMOTE sampling approach are shown in S2A Fig. Among the 10 most important taxa for each ethnicity, there are nine taxa that overlap between the AGP and HMP data sets (highlighted by the blue rectangular box); however, which ethnicity, they best distinguish varies between the two data sets. Within each data set we highlighted taxa that are distinguishing in RF models and have distinguishing differential abundance in S2B Fig, reporting both the FDR corrected significance for Kruskal–Wallis tests of differential abundance, and the percent importance for the most distinguished ethnicity of each in RF models. We also report out-of-bag errors for the final RF classifier that was built using the optimal model parameters obtained from cross-validation approach corresponding to each ethnicity and sampling procedure for both AGP and HMP data sets in S2C Fig.
Taxon differential abundance across categorical metadata groups was performed in QIIME (QIIME: group_significance.py, S5 Table) to examine whether observation counts (i.e., OTUs and microbial taxon) are significantly different between groups within a metadata category (i.e., ethnicity, sex, BMI, and age). The OTU table prior to final community quality control was collapsed at each taxonomic level (i.e., Phylum–Genus; QIIME: collapse_taxonomy.py), with counts representing the relative abundance of each microbial taxon. Differences in the mean abundance of taxa between ethnicities were calculated using Kruskal–Wallis nonparametric statistical tests. p-values are provided alongside false discovery rate and Bonferroni corrected p-values, and taxon were ranked from most to least significant. Results were collated into excel tables by taxonomic level and metadata category being examined, with significant (FDR and Bonferroni p-value < 0.05) highlighted in orange, and taxa that were false discovery rate significant in both data sets were colored red. The Fisher’s exact test for the overlap of number of significant taxa between data sets was run at the online portal (http://vassarstats.net/tab2x2.html), with the expected overlap calculated as 5% of the number of significant taxa at all levels within the respective data set, and the observed 25 taxa that overlapped in our analysis. The permutation analysis was performed by comparing the number of significant taxa (S5 Table, pFDR < 0.05) overlapping between the AGP and HMP to the number overlapping when the Kruskal–Wallis test was performed 1,000 times with ethnicity randomly permuted. In 1/1,000 runs, there was one significant taxon overlapping at the family level and one in 3/1,000 permutations at the genus level, with no significant taxa overlapping in any repetitions at higher taxonomic levels. The 12 families and genera that were significantly different were evaluated to not be taxonomically distinct if their abundances across ethnicities at each level represented at least 82%–100% (nearly all >95%) of the overlapping taxonomic level, and the genera was used if classified and family level used if genera was unclassified (g__). Average relative abundances on a log10 scale among individuals possessing the taxon were extracted for each taxon within each ethnicity, and the abundance for 12 families and genera were made into bar chart figures (Fig 4). The external whisker (AGP above, HMP below) depicts the 75th percentile of abundance, and the internal whisker depicts the 25th percentile. Pairwise Mann–Whitney U tests were performed between each pair of ethnicities using microbial abundances among all individuals and were Bonferroni corrected for the six comparisons within each taxon and data set. Bonferroni significant p-values are shown in the figure and shown in bold if significance and direction of change replicate in both data sets. Ubiquity shown above or below each bar was calculated as the number of individuals in which that taxon was detected within the respective ethnicity. Additional confirmation of ethnically varying abundance was also performed at each taxonomic level (S6 Table), at which the correlation of continuous age and BMI along with categorically coded sex and ethnicity were simultaneously measured against the log10 transformed relative abundance of each taxon among individuals possessing it using linear regression (S6 Table, Abundance) and against the presence or absence of the taxon in all individuals with logistic regression (S6 Table, Presence Absence). Significance is presented for the models each with ethnicity alone and with all metadata factors included (age, sex, BMI), alongside Bonferroni corrected p-values and individual effects of each metadata factor.
Bacterial taxonomy was collapsed at the family level, Spearman correlation was calculated between each pair of families using SciPy [83], cluster maps were generated using seaborn (S3 Fig), and ethnic associations were drawn from S5 Table. Correlations were masked where Bonferroni corrected Spearman p-values were >0.05, and clusters were identified as the most prominent (strongest correlations) and abundance enriched. Enrichment of ethnic association was evaluated by measuring the Mann–Whitney U of cluster families’ ethnic associations (p-values, S5 Table) compared to the ethnic associations of noncluster taxa. Cluster-associated families were identified as having at least three significant correlations with families within the cluster.
The abundance of the family Christensenellaceae was input as relative abundance across all individuals from the family level taxonomic table. Individuals were subset based on the presence/absence of Christensenellaceae, and BMIs were compared using a one-tailed Mann–Whitney U test, then each was further subset by ethnicity and BMI compared using one-tailed Mann–Whitney U tests and boxplots within each ethnicity (Fig 5).
Genetically associated taxa from population heritability studies [34, 35, 39, 40] with a minimum heritability (A in ACE models or H2r) >0.1 and from GWAS studies [35, 40] were examined for exact taxonomic overlap with our 12 ethnically-associated taxa. The 42 genetic variants associated with Unclassified Clostridiales are rs16845116, rs586749, rs7527642, rs10221827, rs5754822, rs4968435, rs17170765, rs1760889, rs6933411, rs2830259, rs7318523, rs17763551, rs2248020, rs1278911, rs185902, rs2505338, rs6999713, rs5997791, rs7236263, rs10484857, rs9938742, rs1125819, rs4699323, rs641527, rs7302174, rs2007084, rs2293702, rs9350764, rs2170226, rs2273623, rs9321334, rs6542797, rs9397927, rs2269706, rs4717021, rs7499858, rs10148020, rs7524581, rs11733214, and rs7587067 from [35]. These 40 variants along with variants in Table 1 except for chr7:96414393 (total = 49) were then assessed in 1,000 Genomes individuals for significant differentiation across superpopulations [38]. The 1,000 Genomes VCF files were downloaded (ftp://ftp.1000genomes.ebi.ac.uk/vol1/ftp/release/20130502/), and variants with a minor allele frequency less than 0.01 were removed, with FST calculated between each pair of superpopulations using vcftools [84]. The East Asian versus South Asian FST rates were not used in the analysis. A custom script was used to examine the FST for each of the 49 variants and was compared to the FST of all variants on the same chromosome and all variants genome-wide for that pair of populations, with percentile calculated and the number of variants with a higher FST divided by the total number of variants. The eQTL value and significance for rs7587067 were drawn from the GTEx database [73].
|
10.1371/journal.pntd.0007281 | Vector competence of Australian Aedes aegypti and Aedes albopictus for an epidemic strain of Zika virus | Recent epidemics of Zika virus (ZIKV) in the Pacific and the Americas have highlighted its potential as an emerging pathogen of global importance. Both Aedes (Ae.) aegypti and Ae. albopictus are known to transmit ZIKV but variable vector competence has been observed between mosquito populations from different geographical regions and different virus strains. Since Australia remains at risk of ZIKV introduction, we evaluated the vector competence of local Ae. aegypti and Ae. albopictus for a Brazilian epidemic ZIKV strain. In addition, we evaluated the impact of daily temperature fluctuations around a mean of 28°C on ZIKV transmission and extrinsic incubation period.
Mosquitoes were orally challenged with a Brazilian ZIKV strain (8.8 log CCID50/ml) and maintained at either 28°C constant or fluctuating temperature conditions. At 3, 7 and 14 days post-infection (dpi), ZIKV RNA copies were quantified in mosquito bodies, as well as wings and legs, using qRT-PCR, while virus antigen in saliva (a proxy for transmission) was detected using a cell culture ELISA. Despite high body and disseminated infection rates in both vectors, the transmission rates of ZIKV in saliva of Ae. aegypti (50–60%) were significantly higher than in Ae. albopictus (10%) at 14 dpi. Both species supported a high viral load in bodies, with no significant differences between constant and fluctuating temperature conditions. However, a significant difference in viral load in wings and legs between species was observed, with higher titres in Ae. aegypti maintained at constant temperature conditions. For ZIKV transmission to occur in Ae. aegypti, a disseminated virus load threshold of 7.59 log10 copies had to be reached.
Australian Ae. aegypti are better able to transmit a Brazilian ZIKV strain than Ae. albopictus. The results are in agreement with the global consensus that Ae. aegypti is the major vector of ZIKV.
| Zika virus (ZIKV) is a mosquito-borne pathogen that generally causes a mild febrile illness but mostly remains asymptomatic in 50–80% of infections. Infection during pregnancy can cause congenital malformations, notably microcephaly. In adults, it can cause Guillain-Barré syndrome. The recent ZIKV epidemic in the Americas has been linked to the urban vector Aedes aegypti. The presence of the species in Australia makes the region vulnerable to emerging mosquito-borne viruses. A mosquito’s competence to transmit a pathogen will depend on both the virus and vector strains. Here, we determine the vector competence of Australian Ae. aegypti and Ae. albopictus mosquitoes for a ZIKV epidemic strain, originating from the epicentre of the Brazilian outbreak, under constant and fluctuating temperatures that simulate field environments in Australia. Our results demonstrate that, although both species were susceptible to ZIKV infection, Ae. aegypti is more likely to transmit virus. Our results may aid in the formulation of public health strategies to mitigate the threat of ZIKV.
| Over the past decade, Zika virus (ZIKV) has caused unprecedented epidemics in the Western Pacific and the Americas. ZIKV is a mosquito-borne, single-stranded RNA virus that belongs to the Flavivirus genus within the Flaviviridae family [1]. First discovered in Uganda in 1947 [2], ZIKV spread from equatorial Africa into Asia in 1960, producing two main genotypes, the African and Asian lineages [3, 4]. Major epidemics of ZIKV have occurred on Yap Island, Federated State of Micronesia [5, 6], French Polynesia [7], some islands in the south and south-west Pacific region [8–11], and most recently Latin America [12–15]. Although 80% of ZIKV infections remain asymptomatic or cause a mild febrile illness [5, 16], recent epidemics have seen more severe disease manifestations, such as microcephaly and central nervous malformations in neonates [17, 18], and Guillain-Barré syndrome in adults [19, 20].
Although ZIKV can be transmitted sexually [21], through blood transfusion [22], and from mother-to-child [23], humans are primarily infected through the bite of infected Aedes (Ae.) mosquito species [24–28]. In Africa, where it was first isolated from Ae. africanus [29], ZIKV is mainly transmitted by sylvatic Aedes mosquitoes (Ae. furcifer, Ae. luteocephalus, Ae. taylori, Ae. opok, Ae dalzieli) [24, 30]. Initial evidence for human infections implicated Ae. aegypti in the urban transmission of ZIKV in Africa [26, 31, 32]. In Asia [4, 33, 34] and the Americas [35–37], Ae. aegypti is considered the main vector for human ZIKV transmission. Although Ae. hensilli was suspected to be responsible for ZIKV transmission during the Yap outbreak [27], Ae. aegypti and Ae. polynesiensis were the main vectors in the French Polynesian outbreak [28]. Recent evidence of vertical transmission of ZIKV in field-collected eggs of Ae. aegypti from Brazil suggests that, in endemic areas, virus may also be maintained in drought resistant eggs [38]. Ae. albopictus is an invasive vector which has colonized most of the tropics and subtropics, as well as more temperate regions of the United States and Europe [39]. The high vectorial capacity of Ae. albopictus for various arboviruses [40–42] places any area colonized by this species at risk of local ZIKV transmission [43, 44]. Considerable variation in ZIKV vector competence, similar to that reported for DENV [45–47], has been observed in both Ae. aegypti and Ae. albopictus from across the globe [25, 48–54]. The transmission efficiency of ZIKV is governed by interactions between mosquito strain [25, 53] and virus genotype/strain [45, 53, 55–57]. This variability underscores the importance of evaluating the vector competence of local mosquito populations for ZIKV.
Australia remains at risk of ZIKV introduction due to its close proximity to the Western Pacific, the presence of competent strains of Ae. aegypti in Queensland [58, 59] and Ae. albopictus in the Torres Strait [48, 60], and favourable climatic conditions for transmission [61]. Despite 51 reports of imported cases of ZIKV since 2014 (Queensland Government, Australia, accessed 8 October 2018), Australia has not yet reported autochthonous transmission. Previous studies have reported the vector competence of Australian Ae. aegypti for African, Cambodian and Western Pacific strains [48, 58, 59] and Ae. albopictus (Torres Strait islands) for Cambodian ZIKV [48, 58, 59]. These studies demonstrated that Australian mosquito strains can be infected and transmit ZIKV; however, large heterogeneity has been observed in the susceptibility of mosquitoes to infection, which may be associated with the origin of the virus strains. There have been no investigations of the vector competence of Australian strains to isolates of ZIKV from South America, despite the continent recording the largest epidemics with a high prevalence of the most severe ZIKV disease manifestations [62]. To assess the public health risk imposed by ZIKV to Australia, we determined the vector competence of local populations of Ae. aegypti and Ae. albopictus (Torres Strait Islands) for a strain of ZIKV isolated from a febrile patient from Paraiba state, at the centre of the 2015/2016 Brazil epidemic. In addition to maintaining infected mosquitoes under a standard constant temperature regime, we also used a fluctuating diurnal temperature range (DTR). Our study indicates that Ae. aegypti has higher relative vector competence than Ae. albopictus, which may be mediated by a salivary gland barrier to virus transmission in Ae. albopictus for this ZIKV strain.
Ae. aegypti eggs were obtained from a colony established from Wolbachia-free eggs collected from Innisfail, north Queensland, in April 2016 and subsequently maintained in the QIMR Berghofer insectary at 27°C, 70% relative humidity [RH] and 12:12 h lighting with 30 min crepuscular periods. Ae. albopictus eggs were obtained from a colony established from eggs collected on Hammond Island, Torres Strait, Australia, in July 2014 and subsequently maintained in the QIMR Berghofer insectary. Eggs of both colonies were hatched and larvae were reared at a density of 400 individuals in 3 L of rainwater. Larvae were provided ground TetraMin Tropical Flakes fish food (Tetra, Melle, Germany) ad libitum. Pupae were transferred to a container of rainwater inside a 30 × 30 × 30 cm cage (BugDorm, MegaView Science Education Services Co., Taichung, Taiwan) for adult emergence. Adult mosquitoes were provided with 10% sucrose solution on cotton wool pledgets.
The Brazilian ZIKV strain KU365780 [63] used in this study was isolated from a febrile patient in Joao Pessoa City, Paraiba State, Brazil, 18-05-2015 (provided by the Evandro Chagas Institute, Brazil). Viruses were propagated in C6/36 Ae. albopictus cells, maintained at 28°C in RPMI-1640 medium (Sigma Life Sciences, USA) supplemented with 10% fetal bovine serum. Following three passages in C6/36 cells, virus stocks were concentrated using Ultracel-100k filters (Amicon, Tullagreen, Cork Ireland) and frozen once at -80°C until further use. Virus stocks were titrated using a Cell Culture Enzyme-linked Immunosorbant Assay (CCELISA) based on the method of Broom et al. [64]. Ten-fold serial dilutions of virus stocks were inoculated on C6/36 (Ae. albopictus) cells grown in RPMI 1640 supplemented with L-glutamine, 5% heat denatured FBS, 1% penicillin/streptomycin (Gibco Life Technologies, USA) and maintained at 28°C, 5% CO2 for 5 days. Monolayers were incubated at 28°C, 5% CO2 for 5 days, and cells fixed at -20°C for 1 h in 80% acetone/20% phosphate-buffered saline (PBS). Plates were air-dried and antigen was detected using a 4G4 anti-Flavivirus NS1 monoclonal antibody hybridoma supernatant (1:40 in PBS-Tween), Horseradish peroxidase (HRP-) conjugated goat anti-mouse polyclonal antibody (DAKO, Carpinteria, CA, USA) (1:2000 in PBS-Tween), and 3,3′,5,5′-Tetramethylbenzidine (TMB) Liquid Substrate System for Membranes (Sigma-Aldrich, St. Louis, MO, USA). Staining was observed using an inverted microscope, and cell monolayers that stained blue were scored positive for infection. The 50% cell culture infectious dose (CCID50) was determined from titration end points as previously described [65] and expressed as the CCID50/ml in C6/36 cells.
Mosquito infection with ZIKV occurred in a Biosafety Level 3 insectary at QIMR Berghofer. An artificial membrane feeding apparatus, fitted with a porcine intestinal membrane, was used to orally challenge adult females (3–5 day old) with a mixture of defibrinated sheep blood (Life Technologies, Mulgrave, VIC, Australia) and virus at a final concentration of 8.8 log CCID50/ml (C6/36 cells) for 1 h. Following ZIKV inoculation, mosquitoes were maintained in environmental growth chambers (Panasonic), with either a constant temperature program set to 28°C or a fluctuating (cyclical) temperature program (24.5–32°C) around a mean of 28°C [66] (Fig 1). The temperature treatments are referred to here as “constant” and “fluctuating”, respectively. For both treatments, RH was set to 75% and a 12:12 h day:night lighting program was used.
Twenty mosquitoes were harvested at each of three time points (3, 7 and 14 days) post blood feeding by anaesthetizing the insects with CO2 on ice before removing legs and wings. Mosquitoes were gently transferred by their antennae to a glass plate and immobilized on double-sided sticky tape. Saliva was collected by placing the proboscis of each mosquito into a 200 μl pipette tip containing 10 μl of 10% FBS and 10% sugar solution. The insertion of the proboscis into the salivation solution was performed under a dissecting scope and peristaltic movement of the abdomen observed to indicate salivation. Mosquitoes were allowed to salivate for 20 min, after which the contents of the pipette tip were then expelled into a microtube and preserved at -80°C.
Nucleic acids were extracted from individual mosquito bodies or body parts using the High Pure Viral Nucleic Acid Kit (Roche Diagnostics, Mannheim, Germany) according to the manufacturer’s protocol. Briefly, 200 μl Binding Buffer/poly A solution was added to each 2 ml screw cap vial containing the individual mosquito body or body parts. The samples were homogenized by shaking the tubes, containing zirconium silica glass beads (Daintree Scientific, St Helens, TAS, Australia), using a MiniBeadbeater-96 (Biospec Products, Bartlesville, Oklahoma, USA) for 90 s. Following the addition of 50 μl of Proteinase K, nucleic acids were extracted as per the manufacturer’s instructions, and eluted in 50 μl of RNAse/DNase-free Ultrapure water (Invitrogen). Samples were frozen at -80°C until further analysis.
To determine the absolute number of ZIKV copies in each mosquito body or body part, a control plasmid, containing a cloned copy of the targeted ZIKV gene fragment (nucleotides 835 to 911, Genbank accession number EU545988), was constructed. Viral RNA was extracted using the QIAamp Viral RNA Mini Kit (Qiagen, Germany), and cDNA synthesized using the SuperScript III Reverse Transcriptase kit (Invitrogen, Thermo Fisher Scientific, USA) according to the manufacturer’s protocol. The targeted ZIKV fragment was amplified using CloneAmp HiFi PCR Premix (Takara, Clontech Laboratories, USA), and cloned into the pUC19 plasmid vector (Genscript, New Jersey, United States) using the In-Fusion Cloning Kit (Takara, Clontech Laboratories, USA) as described by the manufacturer. The presence of the insert DNA was confirmed by nucleotide sequencing. For quantitative RT-PCR analysis, the plasmid was linearized by EcoRI (Promega, USA) and purified using the Nucleospin Gel and PCR clean-up kit (Macherey-Nagel, Germany). The concentration and purity of the linearized plasmid DNA was determined using the NanoDrop Lite spectrophotometer (Thermo Fisher Scientific, USA). The plasmid copy number was calculated based on the measured DNA concentration and its molecular weight. Plasmid DNA concentrations were confirmed prior to the preparation of a 10-fold serial dilution from 3×107 to 3×102 copies/μl and run in parallel with the samples in all qRT-PCRs.
ZIKV RNA from mosquitoes was amplified by one-step qRT-PCR using the TaqMan Fast Virus 1-Step Master Mix (Applied Biosystems, USA) according to the manufacturer’s protocol, in a Rotor-Gene 6000 Version 1.7.87 system (Corbett Research, NSW, Australia). Primers and probe used in this study have previously been described [6] and were synthesized at Macrogen, Korea. The probe was labelled with FAM and BHQ1 at the 5′ and 3′ ends, respectively. The 20 μl reaction mixture consisted of 1 μl extracted sample, 4 × TaqMan Fast Virus 1-Step Master Mix, 400 nM of each primer, 250 nM of probe and Ultrapure water (Invitrogen, Thermo Fisher Scientific, USA). Reactions were run in triplicate, and a 10-fold serial dilution of linearized control plasmid DNA (3×107 to 3×102 copies/μl), as well as negative controls (without template), were included in each run. The following thermal profile was used: a single cycle of reverse transcription for 5 min at 50°C, reverse transcriptase inactivation and DNA polymerase activation at 95°C for 20 s followed by 40 cycles of 95°C for 3 s and 60°C for 30 s (annealing-extension step). Data were analysed and quantified using the Rotor-Gene 6000 Version 1.7.87 software (Corbett Research, NSW, Australia). To calculate the total number of ZIKV RNA copies present in each mosquito body or body part, the measured ZIKV RNA copy numbers in 1μl were multiplied by the elution volume (i.e., 50 μl). Samples were scored positive for virus if ZIKV amplification occurred in at least two technical replicates and the number of copies was above the limit of detection of the standard curve.
Samples in which ZIKV failed to amplify were classified as negative. The presence of mosquito nucleic acid in negative samples was verified by amplification of the housekeeping genes RpS17 (Ae. aegypti; Genbank accession number AY927787.2) or RpS7 (Ae. albopictus; Genbank accession number XM_019671546). qRT-PCR for house-keeping genes were performed using the SuperScript III SYBR Green One-Step qRT-PCR kit (Invitrogen, Life Technologies, USA) as per manufacturer’s recommendations. The reactions were performed in a 10 μL total volume containing SuperScript III RT/Platinum Taq Mix, 2 × SYBR Green Reaction Mix, 200 nM of each RpS17/RpS7 primer (RpS17 F: 5′-TCCGTGGTATCTCCATCAAGCT-3′, R: 5′-CACTTCCGGCACGTAGTTGTC-3′; RpS7 F: 5’-CTCTGACCGCTGTGTACGAT-3’, R: 5’-CAATGGTGGTCTGCTGGTTC-3’), 1 μL of extracted sample and Ultrapure water (Invitrogen, Life Technologies, USA). Cycling was performed using the Rotor-Gene 6000 system (Corbett Research, NSW, Australia) with 1 cycle at 50°C for 5 min and 95°C for 2 min, followed by 40 amplification cycles of 95°C for 15 s, 60°C for 30 s and 72°C for 20 s. Melt curve analysis was performed to analyse the specificity of the reaction.
The presence of infectious virus in blood meals and in mosquito saliva samples was determined using CCELISA as described above, with the following modifications. Blood meals were titrated by 10-fold serial dilution on C6/36 (Ae. albopictus) cells grown in RPMI 1640 supplemented with L-glutamine, 5% heat denatured FBS, 1% penicillin/streptomycin (Gibco Life Technologies, USA) and maintained at 28°C, 5% CO2 for 5 days. Mosquito saliva samples were diluted 1:25 in the media described above, supplemented with 0.1% Gibco Amphotericin B (ThermoFisher Scientific, Waltham, MA USA), and used to inoculate duplicate monolayers of C6/36 cells (~90% confluent). Samples were then fixed and stained as described above.
The legs and wings were removed from mosquitoes and the remaining body was fixed in 4% paraformaldehyde and 0.5% Triton X overnight before mosquitoes were transferred to 70% ethanol. Mosquitoes were dehydrated and embedded in paraffin using standard procedures. Paraffin sections (3–4 μM) were fixed to Menzel Superfrost Plus glass histology slides (Menzel-Gläser, Braunschweig, Germany) and air-dried overnight at 37°C. The sections were dewaxed and rehydrated in a descending alcohol series to water, and antigen retrieval was performed in Dako Target Retrieval solution (pH 9.0) at 100°C for 20 min using a Biocare Medical decloaking chamber. On completion of the cooling cycle, slides were cooled for a further 20 min before being washed in three changes of Tris-buffered saline containing 0.025% Tween 20 (TBS-Tween). The sections were incubated in Background Sniper solution (Biocare Medical, Walnut Creek, CA, USA) plus 1% BSA for 15 min to inhibit nonspecific antibody binding. Excess Background Sniper was removed and slides transferred to an opaque humidified chamber for subsequent incubation steps. Sections were incubated in 4G4 anti-Flavivirus NS1 monoclonal antibody hybridoma supernatant (undiluted) overnight at room temperature in a humidified chamber, washed three times in TBS-Tween, and incubated for 1 h in AlexaFluor-488 conjugated donkey anti-mouse antibody, diluted 1:300 in TBS-Tween. After washing three times in TBS-Tween, sections were counterstained with the fluorescent DNA stain 4',6-diamidino-2-phenylindole (DAPI) for 5 min, washed as above and mounted with coverslips using Dako fluorescent mounting medium. Slides were scanned using an Aperio ScanScope Fl slide scanner (Aperio Techologies, Vista, CA, USA) at a magnification of 20×. Quantitative image analysis was performed as previously described [67].
Percentage infection (number of positive bodies/total tested), dissemination (number of positive leg/wing samples per total tested), and transmission (number of positive saliva samples/total tested) were calculated at each dpi for each species under fluctuating and constant temperature regimes. Significant differences between percentages were detected using Chi-Square tests. The median and interquartile range (IQR) values were calculated using GraphPad Prism Version 7.00 (GraphPad Software, La Jolla, California USA, 2008). Log-transformed virus titres in mosquitoes with infected bodies and wings and legs were analysed using two-way Analysis of Variance (ANOVAs) as a function of temperature, species, and their interactions, in IBM SPSS Statistics software version 23.0. Differences were considered statistically significant at p < 0.05. Receiver Operator Characteristic (ROC) curve analysis was performed for both species to predict threshold disseminated titre at which saliva infection was likely to occur. ROC curve analyses were performed using the pROC package in R version 1.12.1 (May 2018) [68], with samples pooled across days and temperatures for each mosquito species to ensure maximum predictive power. The ZIKV staining density (ratio of ZIKV/DNA positive pixel area) within defined tissues in histological sections was analysed by two-way ANOVA as a function of temperature, days post infection and their interaction using GraphPad Prism Version 8.02. Post hoc comparisons of the main effects of days post infection were performed using Sidak’s multiple comparison test.
High body infection percentages (number of positive bodies/total mosquitoes tested) were observed for both mosquito species under constant and fluctuating temperature conditions, at all the time points tested (Table 1). The body infection percentage in Ae. aegypti were 80% (constant) and 75% (fluctuating) at 3 dpi, 65% (constant) and 70% (fluctuating) at 7 dpi, and 70% (constant and fluctuating) at 14 dpi (Table 1). Compared to Ae. aegypti, higher body infection percentages were observed in the Ae. albopictus temperature groups at all time points (Table 1). Infection percentages in Ae. albopictus bodies reached 95% (constant) and 85% (fluctuating) at 3 dpi, 90% (constant and fluctuating) at 7 dpi, and 80% (constant) and 100% (fluctuating) at 14 dpi (Table 1).
Disseminated infection percentages in Ae. aegypti increased from 10% (constant and fluctuating) at 3 dpi, to 60% (constant) and 70% (fluctuating) at 7 dpi, and remained at 70% (constant and fluctuating) thereafter (Table 1). Disseminated infection percentages in Ae. albopictus were 15% (constant) and 0% (fluctuating) at 3 dpi, 70% (constant) and 60% (fluctuating) at 7dpi, and 45% (constant) and 100% (fluctuating) at 14 dpi, with significant differences at this time interval (Table 1). We also found a significant difference in dissemination percentages between the vector species for the fluctuating temperature regime (Table 1).
At early time points, ZIKV was either generally not detectable in saliva, or transmission percentages were too low to be detected with our sample size. None of the Ae. aegypti in the fluctuating temperature group were infectious at 3 dpi; however, in the constant temperature group, ZIKV was detected in the saliva of a single Ae. aegypti mosquito (5% transmission) (Table 1). ZIKV was not detected in the saliva of Ae. albopictus at 3 dpi (Table 1). At day 7 dpi, no Ae. aegypti saliva samples were found to contain infectious ZIKV. At the same time point, ZIKV was first detected in the saliva of Ae. albopictus mosquitoes maintained at constant temperature (10% transmission), but not in the fluctuating temperature group. The ZIKV transmission percentages of Ae. aegypti were significantly higher than in Ae. albopictus at 14 dpi, for both temperature conditions (Table 1). Whereas transmission percentages of 60% (constant) and 50% (fluctuating) were observed for Ae. aegypti at 14 dpi, only 10% (constant and fluctuating) of Ae. albopictus had infectious saliva at this time point (Table 1).
Both species exhibited high viral loads (>107 copies/body) in bodies from 7 dpi in constant and fluctuating temperature groups, which remained at high levels until 14 dpi (Fig 2, S1 Table). No significant differences were observed in viral copy number in bodies between constant and fluctuating temperature regimes (p > 0.05). Overall, we found no significant effect of temperature (p = 0.718), species (p = 0.107), or an interaction between these two factors (p = 0.411) on viral load in mosquito bodies. We did find a significant effect of day post-infection (p < 0.0005) on virus loads, consistent with the observed increase in body titre across the time points in both species (Fig 2, S1 Table).
ZIKV RNA was detected in the wings and legs of Ae. aegypti constant and fluctuating temperature groups at 3 dpi, albeit in only a very few mosquitoes (Fig 3, S2 Table). The median number of RNA genome copies in the wings and legs of both Ae. aegypti temperature groups increased from 3 dpi and reached its highest level at 14 dpi (>107 copies/mosquito wings and legs) (Fig 3, S2 Table). At early time points (3 dpi), levels of ZIKV RNA were ~104 copies/ wings and legs in Ae. albopictus mosquitoes maintained at constant temperature. Thereafter, ZIKV RNA levels in Ae. albopictus marginally increased in both the constant and fluctuating temperature groups until 14 dpi (Fig 3, 7 and 14 dpi, S2 Table). In contrast to Ae. aegypti, a significantly lower disseminated viral load (p < 0.05) was observed in the wings and legs of Ae. albopictus mosquitoes at day 14 (Fig 3, 14 dpi, S2 Table). Overall, a significant difference in the disseminated viral load was observed between species (Fig 3, S2 Table). Significant effects in disseminated titre due to day (p < 0.001), species (p = 0.001) and temperature (p = 0.032) were identified. The results suggested that exposure to constant versus fluctuating temperature does influence viral disseminated titre, although these effects were only observed at days 3 and 14 post-infection (Fig 3, S2 Table). Furthermore, there was a statistically significant interaction between species and dpi (p < 0.001), indicating that disseminated titres differed between species within each of the days post-infection sampled here.
To visualize ZIKV distribution in Ae. aegypti tissues over time, we performed immunofluorescent antibody staining using a monoclonal antibody recognising Flavivirus NS1 proteins (Figs 4A and 5). We quantified ZIKV staining density (Fig 4B) through image analysis of the relative staining area of ZIKV to DNA for individual organs/tissues (Fig 4C–4E). The ZIKV staining density in mosquito midguts was visible from 3 dpi (Fig 4B). ZIKV staining was detectable in tissue/organs surrounding midguts (“body” samples) from 7 dpi. It was detected in the heads of a majority of mosquitoes from 10 dpi. Fewer salivary glands than other organs/tissues could be observed within the mid-sagittal mosquito sections, however, staining of the salivary glands that were observed indicated that a small proportion were infected by 7 dpi. By 10 dpi, all salivary glands had detectable ZIKV staining. An analysis of the staining density within the tissue/organs as a function of dpi and temperature regime found that, for all tissues, the effect of time post infection on ZIKV staining density was highly significant, whereas the effect of temperature regime was not significant (S3 Table). Interactions between time and temperature were non-significant in all cases. Post hoc comparisons revealed that significant increases in staining density occurred between 7 and 14 dpi for all tissues (Fig 4B). ZIKV was also detected and quantified within the ovaries of Ae. aegypti mosquitoes, which showed a significant increase in staining density between 5 and 14 d (Fig 6A and S3 Table). Staining was limited to the follicular epithelium surrounding oocytes (Fig 6B).
We found that a disseminated titre of 7.50 log10 genome copies per mosquito wings/ legs (sensitivity of 0.943; 95% CI: 0.857–1.000) was required to predict successful infection of mosquito saliva in Ae. aegypti. Surprisingly, a lower threshold titre of 6.52 log10 (sensitivity of 0.922; 95% CI: 0.843–0.980) was necessary in Ae. albopictus to obtain ZIKV infection in saliva, in the proportions (2/20 each at 7 and 14 dpi in constant and 2/20 at 14 dpi in fluctuating temperature regimes) of mosquitoes that were able to transmit the virus.
Our study demonstrates that Ae. aegypti populations from north Queensland are susceptible to a Brazilian epidemic ZIKV strain from Asian lineage, and able to transmit ZIKV from 10 dpi. We also show that Torres Strait Ae. albopictus could be infected in high percentages, but only 10% could transmit virus by 14 days. Our results suggest that a high threshold titre of disseminated infection in Ae. aegypti was required to overcome the salivary barrier and allow transmission. A recent report suggested that a threshold viral load of at least 105.1 TCID50 equivalents/mL in the legs and wings of Australian Ae. aegypti mosquitoes had to be reached for transmission of the prototype African strain of ZIKV to occur [58]. The infectious titre of disseminated virus could therefore be a significant predictor of virus detection in saliva. A similar correlation between disseminated virus titre and transmission rate has been reported for ZIKV [69] and DENV-1 [70], with high dissemination titres resulting in increased transmissibility by Ae. aegypti.
In Australia, variable vector competence of Ae. aegypti populations from north Queensland for Zika has been reported [48, 58, 59]. Those populations were shown to be competent vectors for the African lineage of ZIKV [58], but relatively inefficient vectors of a Western Pacific ZIKV strain belonging to the Asian ZIKV lineage [59]. Our data suggest that Ae. aegypti from northern Queensland in Australia may be less susceptible to Asian ZIKV strains than to the prototype African strain [58]. Our findings are supported by the results of oral challenges of Australian Ae. aegypti with a strain of ZIKV from the Western Pacific [59] in which infection and transmission rates were 40 and 37% respectively, using a similar virus titre to that employed here (8.5 and 8.8 log CCID50/ml, for the Western Pacific and Brazilian strains, respectively). It should be noted that the titres used in both studies are higher than those expected in typical human viremias [6, 36]. However, oral challenge of Australian Ae. aegypti with a lower titre of the Western Pacific ZIKV strain (5.6 log CCID50) resulted in only 3% of mosquitoes becoming infected [59]. Our study suggests that both high viremias and high disseminated threshold titres are required in order to obtain successful infection of Ae. aegypti and allow viral transmission to occur. Although Ae. aegypti could transmit ZIKV at moderate efficiency following challenge with a high titre, we have shown that under similar conditions, the transmission capability of Torres Strait Ae. albopictus was only 10%. Ae. albopictus is therefore less likely to participate in local transmission cycles than Ae. Aegypti in Australia.
A higher transmission rate (87%) of a Cambodian ZIKV strain (Asian lineage) has been reported for Ae. aegypti from Cairns in north Queensland [48]. Our data suggest the vector competence of Australian Ae. aegypti mosquitoes could depend on the geographical origin of populations and the virus strain/genotype, although differences between experiments will also contribute to the variation. The importance of investigating vector/virus strain interactions was recently demonstrated for a strain of Ae. aegypti from New Caledonia [69]. Infection, dissemination and transmission rates were significantly lower for recently isolated ZIKV strains from Africa and Asian lineages, compared with older African lineage isolates. In compatible combinations, ZIKV transmission occurred as early as 6 dpi [69]. Such genotype × genotype interactions have also been reported for DENV transmission [71]. Our study is in agreement with proportions of mosquitoes able to transmit ZIKV at 14 dpi reported for American Ae. aegypti challenged with Brazilian (75%), Puerto Rican (65%), and Malaysian (53%) ZIKV strains [72]. Similar to a study of French Polynesian Ae. aegypti [28], we found a significant increase in ZIKV transmission percentages from early time points (3 and 7 dpi) to 14 dpi. Similar transmission patterns have also been observed for other commonly investigated flaviviruses, i.e. dengue [73]. Our results from immunofluorescence analysis indicate that ZIKV transmission in Ae. aegypti potentially occurs from 10 dpi, similar to populations from the Island of Madeira in Portugal that were infectious at 9 days following an oral challenge with a New Caledonian ZIKV strain [49]. In contrast, Ae. aegypti mosquitoes from Singapore were able to transmit an Ugandan ZIKV strain as early as 5 dpi [74].
Compared to Ae. aegypti, Ae. albopictus mosquitoes were poor vectors for the Brazilian strain of ZIKV. The ZIKV transmission percentages observed in our study are similar to those reported for French and Italian Ae. albopictus mosquitoes challenged with ZIKV from the Asian genotype [49, 54]. However, the infection (10–18%) and disseminated infection (10–29%) rates reported in these studies were much lower than those observed in our study. Our results are strikingly different from a vector competence study in Singapore reporting that all Ae. albopictus mosquitoes challenged with an Ugandan ZIKV strain were infectious by 14 dpi [51]. Ae. albopictus populations from the Australian Torres Strait Islands have previously been shown to be highly susceptible to a Cambodian ZIKV strain, with a high prevalence (>75%) of virus in saliva at day 14 post-infection [48]. This suggests that the transmission of ZIKV in this population of Ae. albopictus is highly virus strain-dependent, as previously reported for American Ae. albopictus populations [57]. A specific vector/virus combination may therefore be more efficient at transmitting ZIKV than another.
The extrinsic incubation period, which is the time between oral infection and presence of virus in the saliva of vectors, is a major determinant of transmission efficiency [75]. We established the kinetics of ZIKV infection, dissemination and transmission in Ae. aegypti by measuring viral RNA in mosquito tissues and live virus in saliva and mosquito organs and tissues and measured viral RNA in Ae. albopictus tissues and live virus in saliva. Our findings support an extrinsic incubation period (EIP) of approximately 10 days in Ae. aegypti under the conditions tested. We found that there were dose-dependent thresholds for infection of salivary glands in both species. Surprisingly, despite the lower transmission percentages observed for Ae. albopictus compared to Ae. aegypti, the estimated threshold for transmission was also lower. The result suggests factors other than disseminated viral titre may be responsible for the transmission percentages observed in Ae. albopictus. Possible explanations for the lower ZIKV transmission percentages at 14 dpi for Ae. albopictus, compared to Ae. aegypti, is that EIP is longer in Ae. albopictus, and/or may be modulated significantly by temperature. This has important public health implications for preparedness, and efficient implementation of mosquito control efforts. A recent study reported that the administration of a second, non-infectious blood meal significantly shortened the EIP of ZIKV-infected Ae. aegypti and Ae. albopictus by enhancing virus escape from the mosquito midgut [76]. Ae. albopictus may therefore be more competent for ZIKV transmission under field conditions of frequent feeding, suggesting the risk of an outbreak mediated by this vector may be higher than is indicated by our data. Whether this holds true for Australian Ae. aegypti and Ae. albopictus remains to be determined. Last, we observed ZIKV staining in mosquito ovarian tissue, limited to the follicular epithelium surrounding developing eggs. This may indicate a potential route of infection leading to vertical transmission, which has been observed recently from field specimens collected as larvae [38].
Although most vector competence studies only take mean temperature values into account, recent evidence for DENV shows that diurnal temperature range (DTR) plays an important role in influencing the behaviour of Ae. aegypti [73, 77]. The DTR mimics more realistic field conditions, which could provide more accurate predictive disease outbreak models [73, 77, 78]. Taking into account the daily temperature fluctuation recorded during the summer months in Cairns Australia, we tested the effect of temperature fluctuations on Ae. aegypti and Ae. albopictus vector competence for ZIKV. Fluctuating temperature significantly affected viral dissemination to wings and legs rather than viral titre in bodies. Our findings suggest using a DTR that mimics field conditions is needed to better understand infection dynamics within mosquito hosts.
This study has demonstrated that north Queensland Ae. aegypti are more competent for a Brazilian strain of ZIKV than Ae. albopictus, confirming that Ae. aegypti is the primary vector of Asian lineage ZIKV. The risk of emergence of ZIKV in Australia is potentially high due to the presence of competent mosquito vectors, climatic conditions suitable for transmission, imported cases, and a large naïve population for ZIKV. However, our data were obtained under high-titre challenge conditions and should be viewed in the context of a recent study that shows low competence of north Queensland Ae. aegypti under more typical viremic titres [59]. We also need to add the caveat that our estimates of vector competence were derived from a single experimental replicate. Additional replicates may yield different estimates due to stochastic variance inherent in vector competence experiments.
In the absence of an effective vaccine and as ZIKV transmission is primarily vector-borne, mosquito control is likely to be the most effective preventative measure. In this regard, the use of the endosymbiotic bacterium Wolbachia pipientis has shown potential for the biocontrol of ZIKV [79] and other human pathogenic flaviviruses and alphaviruses [80, 81]. Large field releases in north Queensland of novel Wolbachia-transinfected Ae. aegypti mosquitoes, refractory to infection by a range of arboviruses [79–81], have shown the ability to drive Wolbachia into wild populations [82]. Our data could be beneficial for modelling likely ZIKV transmission dynamics in north Queensland and addressing emerging ZIKV threats to Australia.
|
10.1371/journal.pgen.1002458 | Reduction of NADPH-Oxidase Activity Ameliorates the Cardiovascular Phenotype in a Mouse Model of Williams-Beuren Syndrome | A hallmark feature of Williams-Beuren Syndrome (WBS) is a generalized arteriopathy due to elastin deficiency, presenting as stenoses of medium and large arteries and leading to hypertension and other cardiovascular complications. Deletion of a functional NCF1 gene copy has been shown to protect a proportion of WBS patients against hypertension, likely through reduced NADPH-oxidase (NOX)–mediated oxidative stress. DD mice, carrying a 0.67 Mb heterozygous deletion including the Eln gene, presented with a generalized arteriopathy, hypertension, and cardiac hypertrophy, associated with elevated angiotensin II (angII), oxidative stress parameters, and Ncf1 expression. Genetic (by crossing with Ncf1 mutant) and/or pharmacological (with ang II type 1 receptor blocker, losartan, or NOX inhibitor apocynin) reduction of NOX activity controlled hormonal and biochemical parameters in DD mice, resulting in normalized blood pressure and improved cardiovascular histology. We provide strong evidence for implication of the redox system in the pathophysiology of the cardiovascular disease in a mouse model of WBS. The phenotype of these mice can be ameliorated by either genetic or pharmacological intervention reducing NOX activity, likely through reduced angII–mediated oxidative stress. Therefore, anti-NOX therapy merits evaluation to prevent the potentially serious cardiovascular complications of WBS, as well as in other cardiovascular disorders mediated by similar pathogenic mechanism.
| Williams-Beuren Syndrome (WBS) is a rare developmental disorder characterized by distinctive facial, neurobehavioral, and cardiovascular features, caused by a heterozygous loss of genetic material (deletion) at the 7q11.23 chromosomal band. Elastin protein deficiency, due to deletion of one copy of the ELN gene, is responsible for developmental anomalies in arterial wall remodeling, predisposing WBS patients to high blood pressure and other serious cardiovascular complications. We have previously shown that a fraction of WBS patients who lack a copy of the NCF1 gene, which codes for p47phox, a subunit of NADPH-oxidase (NOX), have lower cardiovascular risk associated with decreased oxidative stress. Here, we used a mouse model of elastin deficiency to better define the contribution of NOX–mediated oxidative stress to the cardiovascular phenotype of WBS and to confirm the role of Ncf1 as a major modulator. In addition, pharmacological inhibition of NOX activation or synthesis with either losartan or apocynin significantly rescued the cardiovascular phenotype of these mice, suggesting that these drugs should also be evaluated in human patients.
| Williams-Beuren syndrome (WBS [MIM 194050]) is a developmental disorder with multisystemic manifestations and a prevalence of ∼1/10,000 newborns, caused by a segmental aneusomy of 1.55–1.83 Mb at chromosomal band 7q11.23, which includes ELN (coding for elastin [MIM 130160]) and 25–27 additional genes [1], [2]. The recurrent WBS deletion common to most patients is mediated by nonallelic homologous recombination between regional segmental duplications that flank the WBS critical region [3]. In addition to distinctive craniofacial characteristics and mild mental retardation with social disinhibition and hyperacusis, a hallmark feature of WBS is a generalized arteriopathy presenting as narrowing of the large elastic arteries [4]. Histological characterization of arterial vessel walls of WBS patients showed increased number and disorganized lamellar structures, fragmented elastic fibers, and hypertrophy of smooth muscle cells [5]. This large arterial vessel remodeling which is a consequence of abnormalities in vascular development, is thought to be responsible for the cardiovascular disease manifested in 84% of WBS patients [4], [6]. Identical vascular features, most prominently supravalvular aortic stenosis, are also found in patients with heterozygous deletions or disruptions of the ELN gene, implicating elastin haploinsufficiency in this phenotype [5], [7]. The arteriopathy is the main cause of serious morbidity in WBS, including systemic hypertension and possible complications such as stroke, cardiac ischemia, and sudden death [8], [9].
Animal models provide further evidence for elastin deficiency as the main cause of cardiovascular disease in WBS, underscoring the prominent role of the elastic matrix in the morphogenesis and homeostasis of the vessel wall [10]. Heterozygous knockout mice with only one copy of the Eln gene reproduce many of the alterations observed in the WBS vascular phenotype [11], [12]. Hypertension is a consistent feature of Eln+/− mice, associated with elevated plasma renin activity (PRA) and angiotensin II (angII) levels, that can be blocked by the administration of angII type 1 receptor (AT1R) antagonists [13]. In addition to direct effects on the vasculature, many of the cellular actions of angII are mediated by the activation of the NADPH-oxidase (NOX), thus stimulating the formation of reactive oxygen species (ROS). Evidence is accumulating that increased oxidative stress has a relevant pathophysiological role in cardiovascular disease, including hypertension, atherosclerosis, and heart failure [14].
In WBS, the dosage of the NCF1 gene, encoding the p47phox subunit of NOX, is a strong modifier of the risk of hypertension. Hypertension was significantly less prevalent in patients whose deletion included NCF1, indicating that hemizygosity for NCF1 was a protective factor against hypertension in WBS. Decreased p47phox protein, superoxide anion production, and protein nitrosylation levels, were all observed in cell lines from patients hemizygous at NCF1 [15]. Reduced angII-mediated oxidative stress in the vasculature was the proposed mechanism behind this protective effect. Indeed, studies performed in Ncf1 knockout mice have revealed that p47phox is one of the major effectors of angII action. The administration of angII did not lead to increased superoxide production or blood pressure elevation in homozygous knockout animals, as it did in wild-type mice [16].
The aim of the present study was to evaluate whether oxidative stress significantly contributes to the cardiovascular phenotype of a mouse model for WBS, and whether reduction of NOX activity by genetic modification and/or by pharmacological inhibition might have a potential benefit in the rescue of this phenotype. By using non-invasive blood pressure measurements, histological, biochemical and molecular analyses, we have documented a negative correlation between NOX activity and the cardiovascular phenotype in a mouse model of WBS, as well as prevention of many of the manifestations by using anti-NOX therapies.
Previously reported mice bearing a heterozygous deletion of half of the orthologous region of the WBS locus (0.67 Mb, from Limk1 to Trim50, including Eln), called DD, were used as a model for the WBS cardiovascular phenotype [11], [17].
We confirmed the elevated systolic, diastolic and mean blood pressures of 16-weeks old DD mice, ∼40% higher than their wild-type littermates on average, without increased heart rate (Table 1 and Table S1). Hypertension was already present at 8 weeks of age and persisted throughout life (Table S2) without reducing life-expectancy, since these animals have been kept alive for more than 2 years with no instances of early death or unexpected morbidity [11]. As previously reported [17], body weight was significantly reduced for DD mice at all ages when compared to wild-type (P<0.001) (Table 1).
Post-mortem evaluation at 16 and 32 weeks revealed significantly larger hearts in DD mice, measured as the heart wet-weight relative to the body weight (P<0.001). The cardiac hypertrophy was associated with increased cardiomyocyte size both in left and right ventricles (P<0.001) (Table 1, Tables S3 and S4). Vascular histology and morphologic examination provided insight into the structure of the aorta. DD mice showed fragmented, disorganized and jagged elastin sheets when compared to wild-type vessels in sections of the ascending aortic wall stained with elastic VVG, as previously reported [11]. We also observed a significantly increased arterial wall thickness (P<0.001), with small changes in the number of lamellar units (Table 1).
The expression of three genes encoding components of the angII biosynthetic pathway, angII precursor (angiotensinogen, Agt), renin (Ren) and angII converting enzyme (Ace), was increased 2 to 5 fold by qRT-PCR on mRNA of several tissues (heart, aorta, lung and kidney) (Figure S1). Accordingly, DD mice showed significantly elevated angII peptide plasma levels (P<0.001) (Table 1). Plasma renin activity levels were, however, highly variable even within groups, thus preventing inter-group comparisons.
The level of oxidative stress in ascending aortas was determined using two experimental avenues, quantifying the levels of superoxide anion and protein nitrosylation (Table 1). These assays demonstrated higher levels of oxidative stress in DD mice when compared to those in wild-type littermates. These results confirm the relationship between hypertension, elevated angII and increased oxidative stress in these mice.
NCF1 gene dosage had been shown to modify the risk of hypertension in WBS patients [15]. Interestingly, while DD mice were consistently hypertensive, PD mice (heterozygous 0.45 Mb deletion, from Gtf2i to Limk1) were normotensive and mean blood pressure in D/P mice (harboring both deletions in trans) was only slightly increased by ∼10% [11]. Although the Ncf1 gene is located outside the PD deletion (Figure 1A), we investigated whether the expression levels of Ncf1 could be affected in these mice, by qRT-PCR in three different tissues and using Eln (hemizygously deleted in DD and D/P and not deleted in PD) as control. DD mice showed a ∼3 fold increase of Ncf1 mRNA, while the expression was reduced in PD animals, and elevated but only ∼2 fold in D/P mice, correlating with the blood pressure (Figure 1B). The low basal Ncf1 expression in PD and the relatively lower (compared to DD) in D/P strongly suggest that there may be a cis regulator element controlling Ncf1 expression in the PD deletion. It also indicates that Ncf1 is a strong modifier for the cardiovascular phenotype secondary to elastin deficiency.
We then investigated the expression of other genes related to the NOX system. On average, transcript levels of all genes but Cyba were significantly increased in DD animals with respect to wild-type. In contrast, they were not significantly different in PD mice, and D/P mice showed elevated expression of Nox2 along with the ∼2-fold increase of Ncf1 levels (Figure 1C and Figure S2). The elevated transcriptional NOX levels observed in DD mice could be the basis for the excessive ROS and protein nitrosylation documented in the aortic wall.
We then crossed mice homozygous for a spontaneous loss of function mutation in Ncf1 [18] with DD mice, in order to generate double heterozygotes in trans (DD/Ncf1−), resembling the genotype of WBS patients with lower risk of hypertension and deletions that include the NCF1 locus. At 16 weeks of age, DD/Ncf1− animals had normal blood pressure similar to wild-type littermates (Figure 2A). AngII plasma levels were reduced with respect to DD mice (P = 0.036), although still elevated compared to wild-type values (P = 0.002) (Figure 2B), and were accompanied by a significant reduction of mRNA expression of the angII biosynthetic pathway genes (Figure 2C).
The hearts of DD/Ncf1− mice were 20% smaller than those of DD mice (P = 0.039), although they still were slightly larger than those of wild-type animals (P = 0.046) (Figure 2D and Table S3). Heart size reduction in DD/Ncf1− mice was accompanied with a decrease in the size of the cardiomyocytes of the left and right ventricles (P = 0.025 and 0.039 respectively). We also observed a reduction of the aortic wall thickness (P = 0.007), with a slight improvement in the organization of the elastin sheets (Figure 2E).
Expression of NOX-related genes was down regulated in DD/Ncf1− as compared to DD animals, reaching values similar to wild-type littermates. Consistently with these data, DD/Ncf1− mice showed significantly reduced levels of protein nitrosylation (P<0.001) and superoxide anion (P<0.001) in their ascending aortas compared to DD mice (Figure 2F).
The evidence for genetic complementation prompted us to investigate whether the treatment of DD mice with either losartan (an AT1R antagonist) or apocynin (a NOX inhibitor) could rescue the abnormal cardiovascular parameters. Both pre- and postnatal-onset treatments with losartan or apocynin corrected the elevated blood pressure levels seen in 16 week-old DD mice (Figure 3A). Blood pressure control was associated with a significant reduction of angII plasma levels in all treated with respect to untreated DD mice, although the levels still remained higher than those of wild-type mice (Figure 3B). Both drugs acted synergistically with the genetic reduction of Ncf1 gene dosage, as shown by the below normal values of angII in treated DD/Ncf1− mice (Table S5). The therapeutic effect was evident at the gene transcription level, since a significant reduction of transcripts encoding three angII biosynthetic pathway proteins was observed both in DD (Figure 3C) and DD/Ncf1− mice (Table S6). A reduction in ROS production was also noted in the ascending aortas of treated DD mice (Figure 4A and Table S7), as well as down regulation of several oxidative stress genes, including Ncf1, Ncf2, Nox2 and Nox4 (Figure 4B and Table S8).
Either apocynin or losartan therapy also completely prevented the cardiac hypertrophy of DD mice. Treated DD animals displayed heart weights and cardiomyocyte sizes similar to those of wild-type counterparts, significantly smaller than those of the untreated DD group (Figure 5A and 5B). A mild improvement of the arterial wall thickness was also evident in all animals treated with both medications, along with reduced elastic fiber fragmentation during histological observation (Figure 5C and 5D). All evaluated parameters of cardiovascular phenotypic rescue (blood pressure, heart size, vascular morphology) persisted at 32 weeks of age in treated mice.
We found a high proportion of fetal deaths (∼32%) associated with the prenatal administration of losartan (Table S9). No difference in the expected Mendelian proportions was observed in the offspring of heterozygous crosses (DD x Ncf1+/−). We also observed premature postnatal deaths in ∼15% of the treated mice with prenatal-onset losartan (Table S9), mostly due to renal failure before the age of sacrifice, with similar frequencies among genotypes. Our data are in agreement with previous reports contraindicating losartan in pregnancy due to its potential teratogenicity [19], [20]. No specific genotype was associated with increased susceptibility to losartan toxicity. On the other hand, no instances of prenatal death or early postnatal complications were observed in apocynin treated animals, and no other complications were observed in any of the groups treated with postnatal onset.
Note: The full dataset of clinical, morphological, biochemical and molecular parameters at the different time-points, including the effects of treatment on wild-type and DD/Ncf1− animals, is provided as supplementary information (Tables S1, S2, S3, S4, S5, S6, S7, S8, S9).
The majority of patients with WBS (84%) manifest cardiovascular problems throughout their lives, particularly an arteriopathy consisting of stenosis of medium and large size arteries that can be present at birth [4]. Hypertension is found in 40%–70% of patients, even during childhood, and there is a significant risk of other cardiovascular complications, such as stroke, cardiac ischemia, and sudden death [4], [6], [8], [21], [22], [23], [24], [25]. Surgical treatment of focal vascular lesions is required in ∼20% of cases and frequently relies on vascular grafts or balloon dilatation angioplasty [4]. Although β-adrenergic blocker and calcium channel blocker drugs have been utilized, there is insufficient evidence to recommend a specific drug therapy for hypertension [8], [22], [26], [27]. Molecules that can either promote elastin biosynthesis or suppress vascular smooth muscle cell proliferation and migration, such as minoxidil [28], glucocorticoids [29], and retinoids [30] have been proposed as possible approaches to the treatment of cardiovascular disease, but none of them have shown to be clinically effective yet. Therefore, additional insight into the pathophysiology is needed to define better-targeted therapies as alternatives to current protocols to prevent the common complications of WBS arteriopathy.
A recently developed mouse model with elastin deficiency (DD) has provided further insight into the cardiovascular disease of WBS [17]. DD mice develop morphological changes in the aortic wall (thickening with disorganized elastin fibers) leading to chronic hypertension and cardiac hypertrophy. As in the case of Eln+/− mice, hypertension of DD mice is related to elevated angII plasma levels, and we have also shown over-expression of several NOX-related genes (Ncf1, Ncf2, Nox2/Cybb and Nox4) and significantly increased oxidative stress in these mice. AngII is an important physiological regulator of blood pressure and cardiac function, with hypertensive, growth, and remodeling effects mediated through AT1R. AngII acting through AT1R is also known to stimulate NOX generating ROS in a variety of cells [31]. Chronic infusion of angII in rats increases vascular NOX-derived ROS preceded by a prominent expression of the p47phox subunit of NOX in the vasculature and kidney [32]. Although some of the ROS serve as signaling molecules in the cells, excessive production is damaging and has been implicated in the progression of many disease processes.
In WBS patients, deletions are almost identical in size and mediated by non allelic homologous recombination, but the deletion breakpoints determine whether a functional copy of the NCF1 gene is included or not in the deleted interval [33]. Patients with ELN deletion and only one functional NCF1 allele have a 4-fold decreased risk of hypertension compared with those with more than one copy of NCF1 [15]. Interestingly, mean blood pressure in adult D/P mice, combining DD and PD deletions, was only slightly increased by ∼10%, suggesting a modifying effect of gene(s) within or near the PD region on blood pressure in these mice [11]. In addition, the presence of the PD deletion somehow decreased Ncf1 expression, being likely the main modifier for the non-significant blood pressure increase in D/P mice. Similarly, by genetic crossing, we have demonstrated that the loss of a functional copy of Ncf1 in DD mice completely restored oxidative stress and plasma angII levels preventing development of serious cardiovascular anomalies. These data reinforced the idea that pharmacological NOX inhibition could be efficient in the treatment of DD mice. Genetic ablation or pharmacological inhibition of Nox4 has proven to have a remarkable neuroprotective role in a mouse model for cerebral ischemia [34]. Antioxidants could also decrease blood pressure in several models of hypertension with proven implication of angII and the redox system, acting to scavenge the ROS produced by NOX, but their clinical effectiveness is limited [35].
The AT1R blocker losartan is known to lower blood pressure and rescue vascular wall alterations in other connective tissue defects by inhibiting the TGF-β signaling [14], [36]. Losartan inhibits the growth and remodeling effects of angII but also the NOX generated ROS, all mediated through AT1R.
On the other hand, apocynin is a naturally occurring methoxy-substituted catechol, experimentally used as a more specific inhibitor of NOX with anti-inflammatory activity demonstrated in a variety of cell and animal models. In resting cells, p47phox is folded in on itself through intramolecular interactions between the autoinhibitory region and the bis-SH3 and PX domains. These interactions are destabilized by phosphorylated serine residues within the autoinhibitory region, allowing p47phox to adopt an activated open conformation. Apocynin is thought to inhibit NADPH-oxidase assembly by preventing phosphorylation of the autoinhibitory region of p47phox, along with some scavenger activity of hydrogen peroxide [37]. Despite the controversy about the specific mode of action to decrease NOX activity, apocynin has been successfully used in a mouse model to treat hypertension and faster arterial thrombosis [38].
We have evaluated the efficacy and safety of both drugs, losartan and apocynin, in our mouse model of WBS cardiovascular pathology using previously titrated dosages with prenatal and postnatal onset of the therapies [36], [38]. Both treatments were highly effective in the prevention of the development of cardiovascular anomalies in DD mice. Similar effects were manifested by reducing the consequences of NOX activity, with almost complete control of hormonal and biochemical parameters in plasma and tissues, and resulting in normalized blood pressure and improved cardiovascular histology. There was an improvement in aortic wall thickness and architecture without complete reversion of the developmental anomalies secondary to elastin deficiency. Apocynin was as efficient as losartan in prenatal onset, with excellent tolerance and without secondary effects. However, a high proportion of fetal and premature postnatal deaths were associated with the prenatal administration of losartan, supporting its contraindication in pregnancy [19], [20]. Both drugs had significant beneficial effects after postnatal onset of the intervention, with excellent tolerance and no secondary effects.
In conclusion, both, losartan and apocynin, have significant efficacy in the treatment of the cardiovascular phenotype of a mouse model for WBS. Losartan is already approved for human use, while apocynin has been used by inhaler in some clinical trials [39]. The validation of apocynin for human use and the development of additional specific inhibitors of NOX are of great interest, given their potential therapeutic utility in some forms of cardiovascular disease [40]. We believe that these drugs merit evaluation as potential therapeutic agents to prevent the serious cardiovascular problems in human patients with WBS.
The study has been performed in accordance with the ARRIVE guidelines, reporting of in vivo experiments (http://www.nc3rs.org/ARRIVE).
Animal procedures were conducted in strict accordance with the guidelines of the European Communities Directive 86/609/ EEC regulating animal research and were approved by the local Committee of Ethical Animal Experimentation (CEEA-PRBB).
Previously reported mice bearing a heterozygous deletion of half of the orthologous region of the WBS locus on chromosomal band 5G1 (0.67 Mb from Limk1 to Trim50, including Eln), called DD for distal deletion, were used as a model for the WBS cardiovascular phenotype [11], [17]. Mice with the proximal half-deletion of the orthologous WBS locus (0.45 Mb from Limk1 to Gtf1i), called PD, and the double mutants in trans (with homozygous Limk1 deletion), D/P [17] were also used for some studies. Heterozygous DD animals were crossed with mice bearing a homozygous loss of function mutation of the Ncf1 gene (B6 (Cg)-Ncf1m1J)[18] to obtain double mutants in the first generation (DD/Ncf1−), harbouring then a mutant allele (DD deletion and Ncf1 mutation) in each chromosome (Figure 1A). All mice were bred on a majority C57BL/6J background (97%). Tail clipping was performed within 4 weeks of birth to determine the genotype of each mouse using PCR and appropriate primers (See primer sequences in Table S10).
Fifteen different groups of mice (7–15 littermate animals per group, 5 groups per genotype: wild-type, DD or DD/Ncf1−), were used in this study for a total of n = 208. The 5 groups per genotype corresponded to untreated animals (NT), treated with losartan (Coozar, MSD) with prenatal (LP) or postnatal onset (LN), and treated with apocynin (Sigma) with prenatal (AP) or postnatal onset (AN). As previously described, drugs were administered in the drinking water with final concentrations of 0.002 g/day for losartan [36] and 2.5×10−4 g/day for apocynin [41]. In the groups of prenatal initiation, pregnant females started treatment at 14.5 dpc and therapy was continued throughout lactation. Postnatal treatments started at 7 weeks of age. In both cases, mice continued on oral therapy until 16 or 32 weeks of age, when they were sacrificed. Drinking water with drugs were refreshed every 3 days and protected from light by wrapping the drinking water container with aluminum foil. We recorded drinking volumes for untreated and treated mice in order to avoid any interference in the drinking water because of drugs supplement (Table S11).
Systolic, mean, and diastolic blood pressure were measured in conscious male mice on three separate occasions by using a tail cuff system (Non-Invasive Blood Pressure System, PanLab), while holding the mice in a black box on a heated stage. In order to improve measurement consistency, multiple sessions were performed to train each mouse. At least 12 readings (4 per session) were made for each mouse (n = 7–15 per group).
Animals were sacrificed at two time points (16 or 32-week-old). Immediately following sacrifice, all the organs in the thoracic cage (thymus, lung, heart and aorta) were removed in block and fixed in 10% buffered formalin at 4°C for 16 hours. Hearts and aorta were dissected, washed, and weighed (wet weight). Hearts and vessels were processed for paraffin embedding. Wall thickness and lamellar units were analyzed using 5 µm cross-sections of the ascending aorta (transected immediately below the level of the brachiocephalic artery) stained with Verhoeff-van Gieson (VVG) to visualize elastic lamina. Wall thickness at 10 different representative locations was measured and averaged by an observer blinded to genotype and treatment arm for each mouse. The number of medial lamellar units (MLUs) at 4 sites was assessed and averaged by 2 separate blinded observers. These axial cross-sections were imaged with an Olympus BXS1 microscope with epifluorescence and phase-contrast optics equipped with the Olympus DP71 camera, and images were captured with the CellB Digital Imaging system software. MLUs counting and wall thickness were quantified using Adobe Photoshop CS (Adobe Systems).
RNA was extracted from the visceral organs of the thorax by using TRIZOL reagent (Invitrogen) according to the manufacturer's instructions, followed by a second spin columns (Qiagen) purification. To avoid possible contamination of gDNA, all samples were analyzed before conversion to cDNA using standard PCR. In addition, primers were designed in different exons to avoid undesired amplification. cDNA was prepared from 1 µg total RNA using random hexamers and SuperScript II RNase H- reverse transcriptase (Invitrogen). The expression of genes involved in the angII biosynthetic pathway (Agt, Ren and Ace) and NOX-related oxidative stress (Ncf1, Ncf2, Nox2/Cybb, Nox4, Cyba and Rac2) were evaluated by quantitative real-time PCR (qRT-PCR). After diluting the cDNA (from 1∶10 to 1∶100, depending on the tissue), 5 µl were used as template for qRT-PCR using an ABI5700 thermocycler (Applied Biosystems) with the FastStart DNA Master SYBR Green Kit (Roche) and gene specific primers. Characteristics of primers are given in Table S10. Amplification of the Rps28 transcript served as RNA control for relative quantification. Each sample and the corresponding negative controls for each pair of primers were analyzed in triplicate at least in two independent experiments. Threshold cycle values were set manually and analyzed using the comparative method [42].
Blood was collected from the mouse heart into EDTA tubes immediately after sacrifice. Plasma was collected after centrifugation at 1,500 g for 10 minutes and stored at −80°C until use. AngII levels were determined with the Renin Fluorometric Assay Kit Sensolyte 520 following the manufacturer's instructions.
Formalin-fixed, paraffin-embedded transverse sections (5 µm in thickness) were mounted on polylysine-coated glass slides. After blockade with 5% bovine serum albumin plus 0.1% Triton X-100 in phosphate-buffered saline overnight at 4°C, the sections were incubated for 90 min at 37°C with the fluorescent probe DHE (Calbiochem, Darmstadt, Germany). In the presence of O2-, DHE is oxidized to ethidium, which intercalates with DNA, and yields bright red fluorescence. After washing with PBS plus 0.1% Triton X-100, sections were mounted and visualized by fluorescence microscopy (Olympus BX51, Japan). DHE fluorescence intensity was analyzed with NIH ImageJ software (v1.43, April 2010;U.S. National Institutes of Health, Bethesda, MD) as previously described [43]. The fluorescence intensity is proportional to the amount of superoxide anion. Thereafter, sections were incubated with 40,6-diamindino-2-fenilindol (DAPI) (300 nM) for 5 min at 37°C, reactive with fluorescent blue, marking the interlayer between DNA base pairs of cell nuclei. DAPI staining of cell nuclei helps detect true DHE staining (present in the nucleus) versus nonspecific staining. The specificity of the immunostaining was evaluated by the omission of the dye (negative controls). For the quantification of fluorescence, we also subtracted the background present in the negative control, in an attempt to eliminate any autofluorescence. All comparisons were made on cuts prepared with the same experimental conditions and the same day.
The distribution of 3-nitrotyrosine residues, as an indirect marker of peroxynitrite (ONOO-) production, was evaluated by indirect immunofluorescence. In brief, arterial sections were blocked for 2 h at 37°C and incubated overnight at 4°C with a polyclonal anti-nitrotyrosine antibody (dilution 1∶100; Chemicon International, Temecula, CA, USA).
All data are presented as means ± SD. Statistical analysis was performed using ANOVA with a post hoc Bonferroni comparison between multiple groups. In specific cases of two-group comparisons we performed t-test. Values of p<0.05 were considered significant.
|
10.1371/journal.ppat.1000805 | Dynamic Imaging of Experimental Leishmania donovani-Induced Hepatic Granulomas Detects Kupffer Cell-Restricted Antigen Presentation to Antigen-Specific CD8+ T Cells | Kupffer cells (KCs) represent the major phagocytic population within the liver and provide an intracellular niche for the survival of a number of important human pathogens. Although KCs have been extensively studied in vitro, little is known of their in vivo response to infection and their capacity to directly interact with antigen-specific CD8+ T cells. Here, using a combination of approaches including whole mount and thin section confocal microscopy, adoptive cell transfer and intra-vital 2-photon microscopy, we demonstrate that KCs represent the only detectable population of mononuclear phagocytes within granulomas induced by Leishmania donovani infection that are capable of presenting parasite-derived peptide to effector CD8+ T cells. This restriction of antigen presentation to KCs within the Leishmania granuloma has important implications for the identification of new candidate vaccine antigens and for the design of novel immuno-therapeutic interventions.
| Leishmania donovani is a protozoan parasite that causes severe disease in humans with associated pathology in the spleen and liver. In experimental models of L. donovani infection, the hepatic response to infection is characterised by the presence of a focal mononuclear cell-rich inflammatory response (a granuloma) surrounding cells infected with intracellular amastigotes. Granulomas provide focus to the ensuing immune response, helping to contain parasite dissemination and providing the major effector site responsible for parasites elimination from the liver. Although granulomas are believed to form around infected resident liver macrophages (Kupffer cells), the role of these cells in intra-granuloma antigen presentation is currently unknown. As CD8+ T cells have been shown to play an important role in hepatic resistance to L. donovani following natural infection, vaccination and during immunotherapy, we asked which cells within the granuloma microenvironment serve as targets for antigen recognition by effector CD8+ T cells. Here we provide evidence that the heavily infected mononuclear cell core of the granuloma is composed almost entirely of Kupffer cells, many having migrated from the surrounding sinusoids. Furthermore, by intra-vital 2-photon microscopy, we show that only Kupffer cells laden with intracellular amastigotes are able to form long-lasting antigen-specific interactions with CD8+ T cells within the granuloma microenvironment. These data have important implications for the understanding of how granulomas function to limit infection and may have important implications for the development of vaccines to Leishmania that are designed to induce CD8+ T cell responses.
| Kupffer cells (KCs), first identified in 1876, are now recognised as the major population of mononuclear phagocytes to inhabit the resting liver. Lining the sinusoids, KCs express a wide range of phagocytic and innate recognition receptors, including CD32 [1], lectin receptors [2] and TLRs (notably TLR2, 3, 4 and 9) [3], and their avid phagocytic activity has been associated with the clearance of blood borne pathogens and the maintenance of immune homeostasis [4]. Although for many years regarded as a homogenous population, recent data suggest that KCs may be divided into two sub-populations, one sessile and radiation resistant, the other motile and bone marrow derived and expressing higher levels of the costimulatory molecule CD80 [5], reminiscent of the CX3CR1+ subset of monocytes that were recently shown to patrol healthy tissues including blood vessels and the skin [6]. In spite of the importance for KCs in the uptake of pathogens, data on their role in the presentation of pathogen-derived antigens is scarce, with most studies focusing on the role of sinusoidal endothelial cells [7] and hepatocytes [8] in the induction of CD8+ T cell tolerance, or the ability of hepatic stellate cells and dendritic cells (DCs) to prime CD4+, CD8+ and NKT cells [9],[10].
In addition to providing a first line of defense against pathogens, KCs are also believed to be involved in downstream events associated with chronic disease, notably in granulomatous inflammation. Granulomas are well-defined mononuclear cell-rich aggregates that ideally serve to ‘contain and control’ pathogen spread [11],[12], but when unregulated may also contribute to disease pathology [13]. Experimental infection with visceralising species of Leishmania provides, along with experimental mycobacterial infection, some of the best characterised models for evaluating granuloma form and function [14],[15], particularly within the hepatic microenvironment. In experimental visceral leishmaniasis (VL), current models of hepatic granuloma formation, based largely upon data obtained using static imaging approaches, suggest that infected KCs create the central nidus of the granuloma, fusing with other mononuclear phagocytes of less well-defined origin, and ultimately attracting lymphocytes and monocytes [16] through chemokine secretion [17],[18]. More recent studies using BCG infection have provided some additional information on macrophage dynamics and T cell motility within hepatic granulomas during this infection [19] but fail to directly address KC function. In spite of the fact that granuloma macrophages harbour much of the hepatic pathogen load during experimental VL, and there have been numerous reports of intracellular infection with Leishmania parasites affecting macrophage APC function [20],[21],[22] the role of KCs as antigen presenting cells in these models has yet to be directly addressed.
In experimental VL, CD8+ T cell responses are required for the effective clearance of parasites [23], provide one of the best correlates of protection following vaccination [24] and can be used effectively in adoptive immunotherapy [25]. These and other data [26],[27],[28] have fuelled interest in the potential for immuno-prophylactic or immuno-therapeutic expansion of CD8+ T cells as a means of disease control. In the present study, therefore, we have directly addressed the question of whether KCs laden with intracellular Leishmania can be directly recognized by antigen-specific effector CD8+ T cells. Our data demonstrate that the majority of amastigote-infected cells within the core of a granuloma represent KCs that have migrated from neighbouring sinusoids, and by flow cytometry, only this population of KCs expresses detectable Kb-SIINFEKL complexes after infection of mice with OVA-transgenic L. donovani. To determine whether KCs engage in cognate interactions with CD8+ T cell in situ, we used intra-vital 2-photon microscopy to quantify T cell recruitment into and behaviour within individual granulomas. These studies show that effector CD8+ T cells accumulate in granulomas in an antigen-specific manner, as a result of having prolonged interactions with amastigote-laden KCs. Thus, we provide the first evidence that KCs undergo cognate interactions with CD8+ T cells in the context of Leishmania infection, a result which has important implications for the development of immunotherapy against this intracellular pathogen.
L. donovani amastigotes are usually identified in tissue based on their characteristic staining pattern after H&E staining of thin sections [14], with the sensitivity of detection, particularly for individual parasites being improved by immuno-histology using polyclonal or monoclonal antibodies [29]. To more readily observe parasites by fluorescent microscopy, we generated stable infective clones of L. donovani expressing tdTomato (tdTom; [30]), a fluorochrome amenable to both confocal and multi-photon imaging. We first infected mice with tdTom-L. donovani and examined their distribution in the liver at day 14 p.i. (Figure 1) in conjunction with staining for F4/80, a marker of mature KCs [31] and CD11c, a marker characteristically associated with DCs [32]. L. donovani amastigotes were readily apparent both at low magnification, where individual amastigotes within heavily-infected cells could not be resolved (Figure 1A), and at higher magnification, where individual parasites were easily distinguished (Figure 1B). Parasites were observed in two main anatomical locations: within granulomas, where they were predominantly associated with the core, and within the parenchyma, where by DAPI staining they appeared to be within isolated cells in areas largely devoid of local inflammatory reactions (Figure 1C). Almost invariably, amastigotes in either location were found within F4/80+ cells (Figure 1A–C). The close apposition and membrane interdigitation of F4/80+ cells made it difficult to score individual cells, so we did not attempt to calculate the percentage of F4/80+ cells that were infected within the core of the granuloma. Reminiscent of the pattern of staining with NLDC-145, a DEC 205-specific antibody [33], a diffuse but detectable level of CD11c expression was also observed on cells at the core of many, but not all, granulomas. These CD11c+ cells also expressed somewhat lower levels of F4/80, compared to the F4/80+ CD11c− cells that occupied the granuloma mantle (Figure 1D–F). Heterogeneity of expression of CD11c within granulomas did not correlate with the presence or absence of amastigotes. In contrast, CD11b+ cells were usually found in the granuloma mantle, with some clearly identifiable as neutrophils based on nuclear morphology. Importantly, the large amastigote-laden cells at the granuloma core that co-stained for F4/80 and CD11c were almost uniformly CD11b− (Figure 1G–I). These data, together with previously published studies [33] suggest that the majority of intra-granuloma amastigotes are found within F4/80+ cells, and some of these cells acquire markers in this local micro-environment that are often associated with DC.
Although flow cytometry might be expected to provide a means for further phenotypic analysis of tdTom-L. donovani infected macrophages, separation of tdTom-L donovani positive cells by cell sorting (Figure S1A and B), followed by cytospin and Giemsa staining (Figure S1C) indicated that parasites became associated with a range of different cell types, including macrophages, monocytes, lymphocytes and polymorphonuclear cells. In many cases, parasites were bound rather than internalised by these cells. Similarly, co-preparation of cells after mixing of liver tissue from C57BL/6 (CD45.2) mice that were infected with WT-L. donovani and from B6.CD45.1 mice that were infected with tdTom-L. donovani clearly demonstrated transfer of tdTom-L.donovani from CD45.1 to CD45.2 cells. Hence, flow cytometry does not provide a reliable means to further characterise the phenotype of cells infected in situ.
Although macrophages are acknowledged to be a central feature of granulomatous inflammation, the precise origin of these cells has not been directly determined. To address this issue, we first studied the distribution of liver resident and inflammatory phagocytes in naïve and L. donovani-infected mice. KCs in the liver of uninfected mice show a characteristically uniform distribution, lining the sinusoids and forming a reticular surveillance network [34]. To more fully determine the spatial context in which KCs line the sinusoids, we performed whole mount immuno-histochemistry, using F4/80 as a marker of mature KCs (Figure 2). In naïve mice, large KCs with extensive projections were readily apparent within sinusoidal spaces (Figure 2A, and Video S1) forming a regular uniformly distributed phagocytic network. In contrast, in mice infected for 14 days with L. donovani, many KCs were aggregated within granulomas, leaving large areas of the sinusoidal network devoid of detectable KCs (Figure 2B and Video S1). Strikingly, although not participating in the granulomatous inflammatory response, KCs that remained isolated within the sinusoidal network nevertheless displayed morphological changes, which could be quantified as a reduced total cell volume compared to KCs in uninfected mice (Figure 2C, D). Although losing the spatial information provided by whole mount immunohistochemistry, we isolated hepatic mononuclear cells and labeled with F4/80 and CD11c to identify four populations of cells in both naive (Figure 2E) and L. donovani infected (Figure 2F) livers. While all four populations were present in both naïve and infected mice, the proportions changed with infection. CD11c−F4/80− cells (Figure 2E and F, R1) accounted for 51.7+/− 5.13% of F4/80+ cells in naïve mice and 38.88 +/− 4.34% in infected mice. CD11chiF4/80int cells (Figure 2E and F, R2) accounted for 17.11 +/− 3.12% in naïve mice and 19.29 +/− 3.31% in infected mice. CD11chiF4/80hi cells (Figure 2E and F, R3) accounted for 13.39 +/− 2.51% in naïve mice and 13.5 +/− 2.96% in infected mice. Finally, CD11cintF4/80int cells (Figure 2E and F, R4) accounted for 7.83 +/− 0.87% in naïve mice and 14.67 +/− 4.82% in infected mice. MHCII expression, used as a surrogate marker for macrophage activation, was shown to be upregulated on all four populations upon infection (Figure 2G–J). These data suggest that most KCs in the infected liver, even if not recruited into granulomas, had responded to the developing inflammatory environment.
To determine if the aggregation of KCs in granulomas was due to a re-distribution of liver-resident KCs, or whether this reflected the recruitment/differentiation of blood or BM-derived precursors after infection had been established, we used fluorescent nanobeads (NBs) to label KCs (and other potential liver-resident phagocytic cells) prior to infection. Such cells could then be subsequently discriminated from inflammatory phagocytes recruited after infection (Figure 3A–F). We first analysed the distribution of these NBs after intravenous injection into naïve mice. As shown in Figure 3A, NBs were readily ingested by liver-resident F4/80+ KC in uninfected mice, providing a readily detectable measure of their phagocytic activity. Most KCs were phagocytic (∼74%, n = 42), with a variable phagocytic load of NBs. Within individual KCs, multiple ‘patches’ of NB labeling could often be observed, presumably reflecting uptake of NBs into discrete phagosomes. These patches also varied in size, a result that might reflect either aggregation of NBs during injection and/or coalescence of multiple phagosomes each containing small numbers of NBs. NBs were also phagocytosed by desmin+ hepatic stellate cells in naïve mice (∼66% of desmin+ cells contained NBs, n = 90), but large aggregates were rarely observed in these cells (Figure 3B). CD11b+ cells are rare in the resting liver as determined by immuno-histochemistry [35], and when observed, these cells did not contain NBs (Figure 3C). We then injected mice with NBs and 4–12 h later, infected them with L. donovani. The distribution of NB+ cells was then observed at both day 14 p.i. (Figure 3D–F) and at d28 p.i. (data not shown), with similar results being obtained at each time point. NBs were readily observed in L. donovani- infected mice, confirming their value as a long-term cell tracer. NBs were highly concentrated in granulomas, largely at the core, and almost exclusively within F4/80+ KCs (Figure 3D). In contrast, although occasionally present within granulomas, hepatic stellate cells were normally excluded from the core of the granuloma and usually did not contain readily distinguishable NBs (Figure 3E). Strikingly, NBs were also not observed in CD11b+ cells (presumptive monocytes, DC and neutrophils) either at the core of the granuloma or when more peripherally dispersed at the granuloma mantle (Figure 3F). To confirm that the distribution of NBs in granulomas was not the result of rapidly recruited inflammatory cells, NBs were injected and the mice infected with L. donovani 12 hours later as described above. No significant infiltration of inflammatory cells was observed 6 hours after infection, with the proportions of CD11b−, CD11bint and CD11bhi cells being similar between mice that received NBs only or mice that received NBs and L. donovani, whether measured in terms of either the frequency or absolute number of cells (Figure S2). These data suggest that NB distribution after infection reflects KC redistribution and is not influenced by rapidly recruited inflammatory cells. Collectively, these data therefore strongly support the contention that the core of the granuloma is derived almost exclusively from resident KCs recruited from the sinusoids early during the inflammatory process.
As a first step to determining whether cells within hepatic granulomas could present MHC class I-restricted antigens derived from L. donovani amastigotes, we infected mice with double transgenic L. donovani made by transfecting an OVA-expressing L. donovani clone (PINK; [25]) with tdTom. OVA expressed by PINK is localised to the parasite plasma membrane by virtue of the HASPB N-terminal dual acylation sequence [36], and is available for in vivo recognition by Kb- restricted OVA257–263 (SIINFEKL) -specific TCR transgenic CD8+ T cells [25],[37]. To determine which cells could process and present SIINFEKL derived from these transgenic parasites, we first used 25-D1.16, a mAb specific for this MHC-peptide complex [38]. By immunohistochemistry, however, we were unable to detect expression of this complex in any cells within the infected liver (data not shown), probably reflecting the very low levels of complex expressed in this physiological setting. Although loosing the spatial information provided by immuno-histochemistry, we next used flow cytometery as a more sensitive assay to detect whether this complex was expressed and on which cells (Figure 4), comparing the expression of 25-D1.16 on hepatic mononuclear cells isolated from mice infected with either PINK or WT L. donovani. Four discrete populations were identified on the basis of CD11c and F4/80 expression (Figure 4A and B, gates 1–4). In comparison to ‘control’ staining determined from analysis of mice infected with WT L. donovani, no expression of 25-D1.16 was observed in CD11c−F4/80− cells ((Figure 4C, R1) nor in CD11chiF4/80int cells (Figure 4D, R2). These CD11chiF4/80int most likely equate to the small number of intra-granuloma DCs observed by histology (Figure 1). We also could not detect specific staining in CD11chiF4/80hi cells (Figure 4E, R3) though the high autofluorescence of these cells may have precluded detection of low levels of 25-D1.16 expression. In contrast, CD11cintF4/80int cells from mice infected with PINK, which represented 11.47±1.4% of total hepatic leucocytes, contained two populations of cells with differing intensity of expression complexes recognised by 25-D1.16 (Figure 4F, R4). Importantly, CD11cintF4/80int cells mice infected with WT L. donovani (a genetic control for non-specific mAb binding) were not stained with 25-D1.16. CD31+ liver sinusoidal endothelial cells account for approximately 35% of the total hepatic mononuclear cells, but are negative for F4/80 and CD11c. Similarly, hepatic stellate cells, noted for their strong autofluorescence and high side scatter properties [39] make up approximately 3% of the hepatic mononuclear cells in these preparations and are likely located within the R3 population, based on expression of F4/80 and CD11c expression (data not shown). Neither the F4/80− nor the R3 population however expressed MHCI-peptide complexes as determined by 25-D1.16 staining. These data argue, therefore, for expression of the Kb-SIINFEKL epitope on restricted population(s) of L. donovani-infected hepatic cells whose phenotype as determined by flow cytometry closely resembles that of infected F4/80+CD11clo KCs at the core of the granuloma (Figure 1).
Although we detected MHC-peptide complex on presumptive intra-granuloma KCs, the inherent loss of spatial information associated with flow cytometry prompted us to seek alternate approaches to identify antigen recognition by CD8+ T cells in situ. As real-time imaging of T cell dynamics has been shown to be a valuable tool for analysing T cell-APC [40] and T cell-target [41] interactions, and we had already established an adoptive transfer model that provided indirect evidence for cognate antigen recognition by CD8+ T cells, we combined these approaches to study the dynamics of CD8+ T cells in the liver of L. donovani-infected mice.
First, to establish the nature of the T cell environment into which adoptively transferred cells would be imaged, we used hCD2.GFP reporter mice [42] to visualise the entire T cell (and NK cell) content of the L. donovani granuloma. In mice infected with either wild type L. donovani or tdTom-L. donovani, prominent accumulations of T cells were observed from d14 onwards (Figure 5 and data not shown). These accumulations were heterogeneous in nature with the T cells demarcating a structure that varied from being a large flat accumulation of cells close to the collagenous liver capsule (Figure 5A, B and Video S2) to more compact, rounded accumulations of cells that protruded further into the parenchyma (Figure 5C and Video S2). Examination of the total volume of the T cell accumulations at d14 (Figure 5D) and d25 (Figure 5E) showed that while the response was heterogeneous in nature throughout the time course of infection studied, smaller granulomas were more frequent in early infection, while larger accumulations were seen later in the response. Most T cell accumulations had readily detectable parasites (Figure 5F and G and Video S3), confirming that these accumulations were indeed granulomas, though as shown earlier using DAPI staining (Figure 1A), infected macrophages could also be found in the parenchyma in the absence of local T cell recruitment (Figure 5H and I and Video S3). Second harmonic imaging of collagen (Figure 5A–C and F–I) also confirmed earlier reports indicating that the L. donovani granuloma is not highly fibrotic in mice [43],[44] and in some instances migration of T cells along collagen fibres within individual granulomas was observed (data not shown).
As the tracking of T cells inside BCG-induced granulomas has suggested that the granuloma microenvironment inhibits the motility of T cells in a non-antigen specific manner [19], we compared the dynamics of OT-I T cells found within WT L. donovani- induced granulomas with those found in the liver parenchyma of the same infected mice. As shown in Figure 5J-L, we found no significant difference in cell velocity or track length whether cells were moving in the parenchyma or within granulomas. Although the meandering index was higher for cells outside of granulomas, this might reflect the influence of the sinusoidal network on the path of the cell movement. Additional analysis of the instantaneous velocities of cells shown in Video S5, also failed to show any obvious difference in the pattern of instantaneous velocity for OT-I cells inside compared to outside of granulomas (Figure S3). Hence, unlike the BCG granuloma, the L. donovani induced granuloma does not appear to pose a major physical barrier to CD8+ T cells motility.
To study the antigen-specific behaviour of CD8+ T cells in granulomas, we labelled effector memory-like CD62Llo OT-I T cells [45] with CMTMR and adoptively transferred these cells into hCD2.GFP mice infected 21d earlier with either WT L. donovani or PINK. The fate of these OT-I T cells in the liver was then followed for up to12 h post transfer. Within 4 h of transfer, transferred OT-I T cells were detected in the liver and found to be primarily within sinusoids (Figure 6A and Video S4) but by 12 h post-transfer, large numbers were fully embedded within granulomas (Figure 6B and Video S4). From full 3D-reconstructions of granulomas, we scored the number of OT-I T cells embedded within granulomas in mice infected with either WT L. donovani or PINK at either 4 h or 12 h post transfer. Although antigen-independent accumulation of OT-I T cells was observed at 4 h (Figure 6C), by 12 h, antigen-specific accumulation of OT-I T cells was evident (Figure 6D). Importantly, granuloma volume, a surrogate measure of the number of T cells, was not significantly different in mice infected with these two parasite lines, ruling this out as one possible explanation for the effect observed (Figure 6E). As an alternate means to confirm the antigen specificity of intra-granuloma CD8+ T cell accumulation, we also transferred effector memory-like influenza-specific F5 CD8+ T cells into WT L. donovani and PINK infected mice. No difference in F5 T cell accumulation was observed in the granulomas in these mice (Figure 6F).
Altered accumulation of CD8+ T cells within granulomas could be the result of altered rates of immigration or emigration. To distinguish between these possibilities, we examined the dynamics of OT-I T cell movement within individual granulomas in WT L. donovani and PINK-infected mice 5–14 h post-transfer of OT-I T cells. We calculated the rate at which OT-I T cells entered granulomas by dividing the number of cells entering or exiting the granuloma in each imaging period by the length of the imaging period in minutes. No significant differences were seen in the rate at which OT-I T cells entered granulomas in WT L. donovani- and PINK-infected mice (Figure 7A). In contrast, the rate at which OT-I T cells left granulomas in PINK-infected mice was slower than in WT L. donovani-infected mice (Figure 7B). The finding that exit rate, but not entrance rate, was influenced by the presence or absence of cognate antigen suggested that OT-I T cells behaved differently if antigen was available. To determine whether this was reflected in altered velocity, we calculated the average velocity of OT-I cells (n = 311 cells from 43 imaging fields) in PINK-infected and OT-I cells (n = 266 cells from 48 imaging fields) in WT L. donovani-infected hCD2.VaDs Red mice (here used to identify the border of the granuloma by endogenous labelling of all other T cells). The results of this analysis demonstrated that OT-I T cells moved significantly more slowly in the presence of cognate antigen (Figure 7C, D, E and Video S5). The meandering index (calculated by diving the displacement of the cell from its original starting point by the total track length of that cell) was significantly higher for OT-I T cells transferred into PINK-infected mice than those transferred into WT L. donovani- infected mice (Figure 7E). This was reflected by significantly lower track lengths for OT-I cells transferred into PINK infected mice and therefore in the presence of cognate antigen (Figure 7F).
As further independent confirmation that the difference in dynamics of OT-I in the presence and absence of antigen was due to antigen recognition and not due to other differences in the granulomas formed following infection with PINK and WT L. donovani, we labelled OT-I T cells with hoescsht-33342 and F5 T cells with CFSE and co-transferred equal numbers into PINK-infected mice. In these experiments, granulomas were visualised by pre-injection of fluorescent NBs to mark the core of the granuloma (Figure 2). In agreement with the data generated using OT-I cells transferred into mice infected with WT L. donovani or PINK parasites, OT-I cells had slower average velocity than F5 T cells imaged simultaneously in granulomas of PINK-infected mice (Figure 7H, I and Video S6). Thus, the presence or absence of cognate antigen determines the dynamics of CD8+ T cell motility in hepatic granulomas.
To determine whether this antigen-dependent reduction in CD8+ T cell motility was due to more extensive or more prolonged interactions with granuloma-resident cells presenting MHCI-peptide complexes, we first asked whether transferred OT-I cells interacted with the granuloma-associated KCs, by labelling the latter at the onset of infection with NBs, as described above. Transferred OT-I cells were observed to make frequent contacts with NB-labelled KCs (defined by large aggregates of NBs; Figure 8A and B and Video S7). However, the presence of cognate antigen did not influence either the percentage of OT-I T cells interacting with NB+ cells (Figure 8C) or in the duration of these contacts (Figure 8D). On the other hand, as shown above, not all granuloma-associated KCs contained amastigotes (Figure 1) and similarly not all cells with this phenotype expressed detectable Kb-SIINFEKL complexes (Figure 4). Many NB+ KC would be expected, therefore, to be devoid of antigen/parasites, and represent KCs recruited during the process of granuloma development (Figure 2), with the net effect of diluting out any the effect of any antigen-specific interactions between KCs and OT-I cells. Therefore, to more directly assess the potential of infected KCs to present OVA peptide, we infected mice with tdTom-PINK or tdTom-WT L. donovani and evaluated the interaction of these cells with CFSE-labelled transferred OT-I T cells (Figure 8E and F and Video S8). As with NB-labelled cells, OT-I T cells made multiple contacts with amastigote-infected KCs within granulomas containing both PINK and WT L. donovani. However, both the frequency of intra-granuloma OT-I T cells that engaged in this behaviour (Figure 8G) and the subsequent duration of these contacts (Figure 8H) was clearly influenced by the presence of cognate antigen. These studies provide the first direct evidence of intra-granuloma antigen recognition by CD8+ T cells and for in situ presentation of MHCI-restricted peptides by KCs.
Granulomas are well-recognised as a central feature of the pathogenesis of human [46],[47], canine [48] and experimental [49] VL, and most if not all perturbations of immune function made under experimental conditions can be related to alterations in granuloma form and function [14],[50]. Nevertheless, the processes by which these structures form around initially infected KCs and how the microenvironment they create serves to guide and focus host effector function remain poorly understood. Here, we provide the first direct in situ evidence that KCs serve as targets for antigen recognition by granuloma-infiltrating CD8+ T cells. In addition, our study, together with that of Egen and colleagues using experimental BCG infection [19], help dispel the notion of the granuloma as being a static tissue structure and reveal the intricate dynamics of lymphocytes within this unique microenvironment.
Historically, granulomatous inflammation during L. donovani infection has been classified on the basis of the histological response that occurs around each infected KC, providing both a quantitative means to score granuloma ‘maturation’ and a surrogate measure of the quality of the host protective response [14],[50],[51]. Our studies using fluorochrome-reporter transgenic parasites and mice, whole mount confocal and 2-photon microscopy, performed here as a prelude to the analysis of antigen presentation within granulomas, also provide new insight into some of the basic features of granuloma formation. For example, our data shows that a significant proportion of the sinusoidal KC network becomes incorporated within developing granulomas, yet at the same time even those KCs not directly engaged in the process undergo profound morphological changes indicative of activation. Such changes in morphology have been used previously in vitro [52] and ex vivo [53] as correlates of macrophage activation, but cell volume has not previously been measured in situ. The correlation of cell volume with increased expression of cell surface MHCII suggests that it is a true indication of macrophage activation, opening new avenues for the use of whole mount microscopy in the study of KC activation in the study of diseases such as liver injury [54] or liver regeneration following resection or transplantation [55]. Our results also confirm earlier observations [56] that some infected KCs fail, at least for many days or even weeks, to form a focus for inflammation. This marked asynchrony in granuloma development has been the subject of debate [50] and has recently been subjected to systems biology-based approaches [57],[58], but the key determinants of this response remain to be identified. In a recent study in the model organism zebrafish, macrophages infected with BCG were able to migrate out of granulomas [56]. Although migration of L. donovani-infected KCs might also give rise to a population of infected cells apparently uninvolved in the granuloma formation, we do not believe that this scenario is likely in the intact mammalian host, as in neither our studies nor in those of Egen et. al [19] has KC exit from granulomas been observed.
The main focus of this study, however, was on identifying the nature of the cells which engaged with effector CD8+ T cells within the granuloma microenvironment, and in this regard, we provide the first in vivo evidence of a cognate interaction between KCs and antigen-specific CD8+ T cells. Whereas KCs were abundant in granulomas, CD11chi F4/80− DCs were notable by their relative paucity, a finding also reflected in the low frequency of CD11chiF4/80−/int DC observed in mononuclear cell preparations made from infected mice. Although CD11c+ cells were detectable, co-labelling with F4/80, the presence of high numbers of intracellular amastigotes and labelling with NBs confirmed that most of these cells were KCs on which CD11c expression had been aberrantly induced (as is also the case for DEC-205 [33]). Likewise, we observed few CD11b+CD11c+ cells in granulomas, and CD11b+ cells rarely contained intracellular amastigotes. These later data are in stark contrast to the situation observed in the lesions of mice infected with L. major, where the bulk of the amastigote load has been reported to reside within CD11b+CD11c+ ‘inflammatory monocytes’ or ‘TipDC’ [59],[60]. Our data are, however, consistent with earlier reports that indicated both a preference by L. donovani for infection of ‘resident’ compared to inflammatory macrophages and the greater capacity of L. major to stimulate CD11b+ cell recruitment even to hepatic sites of infection [61].
To determine the capacity of these infected KCs to interact in a cognate manner with effector CD8+ T cells, we first tried using immunohistochemical approaches and flow cytomtery to identify which cells could process SIINFEKL from OVA-transgenic PINK parasites and then form complexes recognised by mAb 25-D1.16 [38]. We were unable to detect expression by any immunohistochemical approach we tested (including teramide labelling). By flow cytometry, however, we could detect specific staining on CD11cintF4/80int cells that we believe represent intra-granuloma KCs. Such staining was notably absent on CD11chiF4/80−/int DCs. These ex vivo analyses should however be viewed with some caution. First, granulomas cannot be specifically isolated for analysis, and as a consequence cells analysed by flow cytometry may originate from any anatomical compartment within the infected liver. Second, we cannot exclude the possibility that MHCI-peptide complexes and/or whole parasites are either shed or transferred to other cells during the isolation procedure. Such transfer of MHCI-peptide complexes has been noted under in vitro culture conditions [62],[63] and indeed transfer of parasites between populations of cells during tissue disruption and subsequent cell isolation has been noted by us (Figure S1) and by others [59]. It is, however unlikely that processing of antigen into MHCI is able to occur within the 30 min collagenase digestion step, or the 10 min density gradient centrifugation steps both of which were performed at room temperature, as detection of MHC-I-peptide complexes takes >1 hr following virus infection [64] and is likely to follow similar kinetics following Leishmania infection. All other processing steps were performed on ice. Whilst analysis of the expression of MHCI-peptide complex expression might, therefore, also suffer from the same technical difficulties, we believe this is unlikely. Third, only low numbers of MHCI-peptide complex are required for productive engagement with CD8+ T cells [65], well below that detectable by mAb staining. These caveats notwithstanding, our data suggested that KCs and not DCs expressed such complexes in most abundance. It should also be noted that sessile KCs are not readily isolated by the methods we used [5] and as such have been largely excluded from this ex vivo analysis. We cannot therefore comment on whether such KCs do or do not express complexes recognised by 25-D1.16.
To more definitively identify the sites of antigen presentation, we therefore turned to intra-vital imaging of adoptively transferred CD8+ effector cells. CD62Llo effector CD8+ T cells generated in vitro using antigen expansion and IL-2, were chosen for analysis, as these cells have previously been shown to bring about a rapid and antigen-specific reduction in hepatic parasite burden [25]. Furthermore, analysis of the fate of such cells may provide clues as to how similar effector cells induced by vaccination may behave. We used similar methods to those of others working in lymphoid tissue [66],[67], in tumor microenvironments [68],[69], in BCG granulomas [19] and in the brains of Toxoplasma gondii infected mice [70],[71] to define the dynamic behaviour of CD8+ T cells in hepatic granulomas caused by L. donovani and our results not surprisingly showed marked similarities in T cell behaviour. The dynamic nature of the T cell compartment within the L. donovani granuloma was also clearly evident in all the imaging that we performed. Though superficially similar to that reported for BCG infection, contrasts between L. donovani and BCG granulomas can be noted. For example, whilst T cells were reported to stay within the granuloma structure following BCG injection [19], we found that endogenous T cells, as well as adoptively transferred antigen-specific and non-antigen specific CD8+ T cells, could readily migrate out of granulomas, indicating a net flux through this ‘compartment’. Additionally, while a marked difference in the velocity of cells tracked within and outside of BCG induced granulomas showed that the granuloma per se was capable of inducing a change in cell movement [19], CD8+ T cells migrated without apparent constraint into, within and out of L. donovani granulomas, and non-antigen specific cells showed the same speed of cell movement whether located inside or outside of granulomas. Similarly, analysis of the instantaneous velocity of cells within and outside of granulomas showed no obvious differences, confirming the presence of antigen as the only factor that induces a change in cellular behaviour. The differences in behaviour of T cells in these two types of granuloma may be attributable to differences in composition of the mononuclear cell mantle or reflect differences in other environmental factors e.g. the level of fibrosis [43],[44]. The traffic of antigen-non-specific T cells through granulomas also provides a timely reminder that the histological identification of T cells within granulomas, in the absence of dynamic measurement, is neither an indicator of antigen-specificity nor a good marker for the effector capacity of these structures.
In most cases, we validated our approach by cross-over experiments in which on the one hand we used adoptive transfer of OT-I T cells into mice infected with WT L. donovani or PINK parasites, and on the other hand, we used co-transfer of OT-I and F5 T cells into PINK-infected mice. While labelling with NBs was sufficient to allow the identification of the core of the granuloma, it does not delimit the extent of the granuloma. Hence, CD8+ T cells often appear distant from the core of the granuloma, they were still maintained within its boundaries. Cells frequently migrated near to and, in fact, through the NB-labelled core, but the interactions with NB-labelled cells were not sufficient to demonstrate antigen-specific interactions. This result is not surprising, given that not all cells present in the granuloma core contained parasites (Figure 1) and the effect of any antigen-specific contacts with infected NB+ cells would likely be diluted out by interactions with non-infected NB+ cells. Meandering of antigen-specific CD8+ T cells was extensive, as might be predicted from the dense packing of lymphocytes within granulomas and whereas migration upon collagen fibres was noted, this was not seen in all instances. Numerous contacts were also made between CD8+ T cells and amastigote-infected KCs. Importantly, as measured by all these parameters, the intra-granuloma behaviour of effector CD8+ T cells was markedly influenced by the presence of cognate antigen.
In spite of clear data supporting intra-granuloma antigen recognition by CD8+ T cells, we have not to date observed evidence of direct effector activity of these transferred CD8+ T cells, such as dispersal of amastigotes or their loss of apparent viability. The failure to do so may be due to the length of time taken for CTL to lyse targets in vivo, with target cell lysis in vivo reported to take as little as 17 min in the case of target cells pulsed with high doses of peptide [72] or as long as 6 hours for tumour targets [73]. Additionally, there are technical limitations to these methods as the maximum imaging window we can achieve is 10 h, and this may be insufficient to observe degradation of amastigotes (and/or tdTom protein) subsequent to cytokine-mediated macrophage activation. Similarly, as KC integrity was not directly imaged in these experiments, it is possible that host cell lysis occurs but amastigotes were rapidly re-engulfed by neighbouring KCs. Furthermore, we cannot rule out that after recognition, CD8+ T cells exert their leishmanicidal effect indirectly and over a longer time frame than examined here. Additional developments in imaging technology and new tools to study macrophage responsiveness to activation signals in real time will be required to conclusively address this issue. Although our data are the first to directly demonstrate KC interactions with effector CD8+ T cells, KC-mediated priming of CD8+ T cells was recently demonstrated using cell lines in vitro [74] and also with freshly isolated KCs ex vivo [75], suggesting that further studies into the role of KC in presenting Leishmania-derived antigens to naïve CD8+ T cells at the initiation of infection are also now warranted.
In conclusion, we have shown that KCs laden with amastigotes serve as the principal target for antigen recognition by effector CD8+ T cells within the granuloma microenvironment. Our data suggest that if CD8+ T cell recognition is to form the basis for prophylactic or therapeutic vaccination, then it will be essential to understand the rules which govern MHC class I epitope selection within infected KCs, as well as within those APC (e.g. DCs) that are responsible for induction of CD8+ T cell responses. Furthermore, chemotherapeutic or immunotherapeutic interventions that enhance antigen presentation by KCs may prove highly beneficial.
C57BL6 mice were obtained from Charles River (UK). hCD2.GFP [42] and VaDS Red B6 and Rag-1−/− F5 mice, originally a kind gift from Dimitris Kioussis (NIMR, Mill Hill, UK), and Rag-1−/− OT-I mice were bred and housed under specific pathogen-free conditions and used at 6–12 weeks of age. The Ethiopian strain of Leishmania donovani (LV9) and OVA expressing LV9 (PINK) [25] were maintained by serial passage in Rag-1−/− mice. Amastigotes were isolated from infected spleens, as previously described [24], and mice were infected with 2×107 L. donovani amastigotes intravenously (i.v.) via the tail vein in 200 µl of RPMI 1640 (GIBCO, Paisley, UK). For pre-labelling of liver-resident macrophages, PD nanobeads (545 marked) (Sigma) were pre-injected into mice i.v. 5–24 hours prior to injection of L. donovani amastigotes. All experiments were approved by the University of York Animal Procedures and Ethics Committee and performed under UK Home Office license (‘Immunity and Immunopathology of Leishmaniasis’ Ref # PPL 60/3708).
Tandem Tomato fluorescent protein (tdTom) gene [30] was cloned into the plasmid pSSU-Neo-Infantum to give pSSU-Neo-Infantum-tdTom [76] [Oyola et. al. manuscript in prep]. WT L. donovani and L. donovani HASPB::OVA (PINK) promastigotes [25],[36] were transfected with this construct (which targets genes into the ribosomal locus of L. donovani) and clones selected by serial dilution in the presence of neomycin. Clones were checked for correct integration of the tdTom gene by PCR and Southern Blotting of BamHI and ScaI digested genomic DNA with a 586 bp probe against the neomycin phosphotransferase gene.
Confocal microscopy was performed on 8–10 µm frozen sections. For tissue containing tdTom expressing parasites, tissue was fixed in 4% paraformaldehyde (PFA) for two hours before overnight incubation in 30% sucrose and embedding in Optimal Cutting Temperature (OCT) medium (Sakura). For all other labelling, tissue was snap-frozen in OCT and sections fixed in ice cold acetone for 8 min. F4/80, CD11c and CD11b antibodies were conjugated to Alexa488 or Alexa647 (eBioscience, UK) and Rabbit anti-desmin (Abcam) was detected with goat-anti Rabbit-647 (Invitrogen).
For whole mount confocal microscopy, thick tissue sections were cut with a scalpel blade and labelled as previously described [77]. Briefly, sections were fixed in 4% PFA for 15 min at room temperature (RT), washed in PBS-Triton (0.15%) and blocked for 2 hours at RT. All subsequent antibody labelling steps were performed for 8 hours or overnight at 4°C followed by final fixing in 4% PFA for 15 min at RT followed by dehydration in methanol. Samples were optically cleared in BABB (sigma) and imaged using a Zeiss LSM510 axioplan microscope (Carl Zeiss Microimaging). Data were rendered and analysed using Volocity software (Improvision). Cell volumes were calculated by generating a measurement item based on RGB and exclusion of objects <300 µm3 and >15 0000 µm3. All objects were manually checked for accuracy before data were plotted and analysed in Prism v5.1 (Graphpad).
Hepatic mononuclear cells were prepared from the livers of wild type and PINK L. donovani infected livers, or livers from C57BL/6 mice injected with 100 µg SIINFEKL peptide I.V, following collagenase digestion as previously described [78]. Briefly, livers were perfused with PBS containing 2%FCS and digested in 350 µg/mL collagenase D (Worthington, UK) for 30 min at RT. Digested livers were passed through a 100 µm cell strainer, washed twice in 2%FCS.PBS and hepatocytes removed by centrifugation on a 33% percoll density gradient for 12 min at 693 g. The remaining cell pellet was kept for further analysis. Isolated cells were labelled with 25-D1.16-biotin [38] and streptavidin-Alexa488 as well as CD11c-PeCy7, F4/80-Alexa647 and CD11bPE or pacific blue (eBioscience, UK). Cells were sorted based on expression of tdTom, with approximately 3000 sorted cells spun onto glass slides, fixed in methanol and stained with Giemsa for morphological analysis.
CD8+ T cells resembling effector memory cells were derived in vitro as described previously [45]. Briefly, splenocytes from naive OT-1 transgenic mice were incubated with 10 µg/ml OVA257–264 (Cambridge Bioscience) for 1 h at 37°C, washed, and cultured for a further 48 h. Cells were then washed and incubated for a further 5–9 days with 20 ng/ml recombinant hIL-2. CD62-L low cells were enriched to >95% purity by negative selection using anti-CD62-L microbeads (Miltenyi). Enriched cells were labeled with 5 µM CFSE or CMTMR (invitrogen) or 6 µM Hoescht 33342 (Sigma) before transfer of 2×107 cells to recipient mice by intravenous injection.
Freshly removed liver tissue was placed in 35 mm coverslip bottom Petri dishes (MatTek corporation), kept moist with PBS and imaged on an inverted LSM 510 multiphoton microscope (Carl Zeiss Microimaging). Images were acquired with a 40×1.1 water immersion objective and fluorescence excitation provided by a Chameleon XR Ti:sapphire laser (Coherent) tuned to 872 nm. Data were rendered and analysed using Volocity software (Improvision). Granuloma volumes were calculated by drawing regions of interest in Volocity to get a 3D volume measurement in µm3. Exported videos were arranged in After Effects software (Adobe).
Mice were anaesthetised with a combination of ketamine (100 mg/kg), xylazine (10 mg/kg) and acepromazine (1.7 mg/kg) given intraperitoneally. After 60 min, anaesthesia was maintained by subcutaneous injections of half doses approximately every 45 min. The abdomen of the animal was shaved, and a ∼1.5 cm midline incision made to expose the xiphoid process which was retracted to allow dissection of the falciform ligament. The left lobe of the liver was then gently exteriorised and the animal inverted onto a glass coverslip mounted within a custom made imaging platform. The liver was covered with sterile saline-soaked gauze to prevent dehydration and the mouse stabilised with micropore tape (3 M). Images were acquired on an inverted LSM 510 multiphoton microscope (Carl Zeiss Microimaging) (as above) which was maintained at 36°C by a blacked-out environmental chamber (Solent Scientific, UK). For 4D analysis, 20–35 µM Z stacks were acquired with a Z distance of 2–3 µM approximately every 15–30 sec. Data were rendered and analysed using Volocity software (Improvision) and cell tracking performed manually, or automatically with manual checking. Entrance and exit rates were calculated by monitoring the number of OT-I cells entering or exiting granulomas, as defined by the endogenous T cell border or the presence of nanocrystals within each imaging window and dividing this number by the time of each imaging window in minutes, to get a rate of OT-I entrance and exit per minute. All granulomas imaged were included in this analysis, irrespective of whether they had associated OT-I cells.
|
10.1371/journal.pgen.1006503 | Mitochondria and Caspases Tune Nmnat-Mediated Stabilization to Promote Axon Regeneration | Axon injury can lead to several cell survival responses including increased stability and axon regeneration. Using an accessible Drosophila model system, we investigated the regulation of injury responses and their relationship. Axon injury stabilizes the rest of the cell, including the entire dendrite arbor. After axon injury we found mitochondrial fission in dendrites was upregulated, and that reducing fission increased stabilization or neuroprotection (NP). Thus axon injury seems to both turn on NP, but also dampen it by activating mitochondrial fission. We also identified caspases as negative regulators of axon injury-mediated NP, so mitochondrial fission could control NP through caspase activation. In addition to negative regulators of NP, we found that nicotinamide mononucleotide adenylyltransferase (Nmnat) is absolutely required for this type of NP. Increased microtubule dynamics, which has previously been associated with NP, required Nmnat. Indeed Nmnat overexpression was sufficient to induce NP and increase microtubule dynamics in the absence of axon injury. DLK, JNK and fos were also required for NP. Because NP occurs before axon regeneration, and NP seems to be actively downregulated, we tested whether excessive NP might inhibit regeneration. Indeed both Nmnat overexpression and caspase reduction reduced regeneration. In addition, overexpression of fos or JNK extended the timecourse of NP and dampened regeneration in a Nmnat-dependent manner. These data suggest that NP and regeneration are conflicting responses to axon injury, and that therapeutic strategies that boost NP may reduce regeneration.
| Unlike many other cell types, most neurons last a lifetime. When injured, these cells often activate survival and repair strategies rather than dying. One such response is regeneration of the axon after it is injured. Axon regeneration is a conserved process activated by the same signaling cascade in worms, flies and mammals. Surprisingly we find that this signaling cascade first initiates a different response. This first response stabilizes the cell, and its downregulation by mitochondrial fission and caspases allows for maximum regeneration at later times. We propose that neurons respond to axon injury in a multi-step process with an early lock-down phase in which the cell is stabilized, followed by a more plastic state in which regeneration is maximized.
| The ability of neurons to survive injury, misfolded proteins, hypoxic stress and other deleterious conditions allows the nervous system to function for a lifetime without large-scale production of new neurons. Neuronal survival strategies buy the cells time to maintain or regain function. For example, neurons may remain non-functional for weeks, months or years after axonal trauma. Their survival allows axon regeneration to take place, and eventually, if an appropriate target is reached, the cells may again function.
Preconditioning is a transient survival strategy triggered by a stressful, but sublethal, event. For example, when blood flow to a region of the brain is transiently reduced, the effects of a subsequent ischemic stroke are not as severe [1, 2]. Tissue-level preconditioning seems to have an immediate phase, and then a longer-term transcription-dependent phase [2, 3] and is proposed to be a very general stress response mechanism.
Preconditioning has also been described at a single cell level. In Dorsal Root Ganglion (DRG) neurons, severing the peripheral axon enables the central axon for regeneration [4]. The initial peripheral lesion triggers transcriptional changes in the cell body that are proposed to facilitate subsequent regeneration of the central axon [5, 6]. In Drosophila models of conditioning lesion in sensory and motor neurons, axon severing turns on a stabilization pathway that is measured by resistance to degeneration after a subsequent injury [7, 8]. This single cell neuroprotection (NP) requires dual leucine zipper kinase (DLK) [7] and c-Jun N-terminal Kinase (JNK) [8]. DLK is a MAP kinase kinase kinase, and JNK is the downstream MAP kinase, which play central roles in the regulatory cascade that initiates axon regeneration in nematodes, flies and mammals [9–12]. DLK/JNK are therefore implicated in regulation of both axon regeneration and preconditioning or NP in response to axon injury.
Using the Drosophila sensory neuron model for preconditioning, we investigate the effectors mediating NP downstream of DLK/JNK, and the relationship between NP and axon regeneration. One hallmark of NP is a dramatic increase in microtubule dynamics [8], a response that has also been seen in mammalian neurons [13]. Mitochondria have been suggested to play a central role in brain preconditioning [14], and are important for axonal stability in C. elegans [15] and in many systems the Wallerian degeneration slow (Wlds) protein seems to act through mitochondria to stabilize axons [16–19]. We therefore started by investigating the role of mitochondria in NP. Surprisingly, we found that, rather than promoting NP, mitochondria have an inhibitory role in this process, and caspases share this negative regulatory role. Moreover, although regeneration and NP are downstream of the same kinase cascade, NP antagonizes regeneration. These results are unexpected, but fit together into a multi-step model of axon injury responses downstream of DLK/JNK.
In Drosophila sensory neurons, severing an axon with a pulsed UV laser stabilizes the cell such that if a dendrite is later removed its degeneration is delayed [8]. Dendrites normally degenerate completely within 18h (Fig 1A). However, when axons are damaged 8h prior to dendrite injury, the severed dendrites are stabilized and take more than 18h to fragment [8]. Stabilization is maximal at 8-24h and tapers off at 48h after axon injury [8] (Fig 1A). The timing of this stabilization correlates with a dramatic increase in the number of growing microtubules, and this increase in microtubule dynamics is required for stabilization [8].
To assess the role of mitochondria in axotomy-induced stabilization or NP, we depleted mitochondria from dendrites using RNAi-mediated knockdown of the mitochondrial Rho-GTPase Miro, which is required for mitochondrial transport in neurons [20, 21]. We have previously shown that Miro RNAi reduces the number of mitochondria in dendrites of ddaE neurons [22]. Because the NP assay uses the speed of dendrite degeneration to probe stability, we tested whether reduction of mitochondria would affect dendrite degeneration itself, without prior axon injury. We previously demonstrated that small regions of dendrites with no mitochondria degenerate with normal timing [22]. Here, we severed the whole dendrite with normal numbers of mitochondria (wild-type, WT) or reduced mitochondria (Miro RNAi) and assayed degeneration at different times after severing. The time course of degeneration in neurons expressing a control (Rtnl2) RNAi or Miro RNAi was similar (Fig 1E) with a few cells starting to degenerate at 7h after dendrite injury, and most cells degenerating by 11h after severing. In the standard NP assay we sever an axon, wait 8h, then sever a dendrite. Dendrite degeneration in this assay is scored 18h after dendrite severing, so the time course we used to assay dendrite degeneration alone (4h, 7h, 11h) was much finer and should have picked up any small differences in speed of degeneration without prior axon injury.
As Miro reduction did not change the timing of dendrite degeneration, we performed the NP assay in control and Miro RNAi neurons. This assay was performed as diagrammed in Fig 1A’. In control neurons, axon injury 8h before dendrite severing results in about 50% of dendrites remaining at 18h (Fig 1B and 1D and [8]). In Miro RNAi neurons, NP was increased and 100% of dendrites remained at the 18h timepoint (Fig 1C and 1D). RNAi targeting milton, which recruits Miro to mitochondria [20, 21], also increased NP (Fig 1D). These results suggest that normal mitochondrial trafficking or dynamics limits injury-induced NP.
To understand how mitochondria might regulate axotomy-induced NP, we compared mitochondrial shape in dendrites before and 8h post axon injury (hpa). Mitochondria were labeled with mito-GFP and membranes with mCD8-RFP. The average length of mitochondria decreased significantly after injury (Fig 2A and 2B). Specifically, more short (magenta arrows Fig 2A) and fewer long mitochondria (orange arrows, Fig 2A) were present at 8hpa (Fig 2C). The total number of mitochondria in dendrites also increased at 8hpa (Fig 2D).
To determine whether mitochondrial fission was responsible for the changes in mitochondria length after axon injury, we used RNAi to target Drp1, a dynamin-related GTPase that mediates mitochondrial fission [23]. Drp1 RNAi hairpins were expressed in ddaE neurons under control of 221-Gal4, together with Dicer2, mito-GFP and mCD8-RFP. One of the Drp1 RNAis (referred to as Drp1 RNAi #2) dramatically elongated mitochondria in dendrites and led to clustering of mitochondria in the cell body (S1A Fig). Injury-induced changes in mitochondria were not visible in this background (S1A Fig). However, mitochondrial shape was so different in these cells that we looked for an alternate Drp1 RNAi line that would not have such a strong effect in uninjured neurons. In neurons expressing a different Drp1 RNAi, mitochondrial length was fairly normal in uninjured cells (S1B Fig). The only significant difference was a decrease in the number of mitochondria under 0.5 μm, which is consistent with partial reduction of Drp1 protein levels. Although the effects in uninjured cells were subtle, this Drp1 RNAi also completely eliminated the axotomy-induced changes in length and number (Fig 2A’–2D’). Because this RNAi line eliminated injury-induced mitochondrial fission without dramatically altering baseline mitochondrial length we used it for the subsequent experiments. Mitochondrial motility was also upregulated in dendrites after axon injury, but this change was not related to Drp1-mediated fission (S1C Fig).
To determine whether the increase in mitochondrial fission induced by axon injury was related to downregulation of NP by mitochondria, we assayed NP in Drp1 RNAi neurons. As in Miro and milton RNAi neurons, Drp1 RNAi increased the level of protection, while overexpression of Drp1 had the opposite effect (Fig 2E and 2F). Drp1 RNAi did not influence the normal time course of dendrite degeneration (Fig 2G), thus Drp1 only affects degeneration after axotomy in the NP assay, but has no effect on the baseline rate of dendrite degeneration in the absence of axon injury.
Drp1-mediated mitochondrial fission occurs during apoptosis in Drosophila and other organisms [24, 25], and mitochondria and caspases have been linked in a neurodegenerative response triggered by glial signaling [26]. In mammals and in C. elegans fission is upstream of caspase activation [27, 28]. Because of this connection between mitochondrial fission and caspase activation, we hypothesized that caspases might also inhibit axotomy-induced NP.
To test this hypothesis we expressed large RNA hairpins to target the initiator caspase Dronc and assayed both dendrite degeneration and NP. Dendrite degeneration is normally complete by 18h after severing, and blocking caspases does not alter this [22]. To test whether caspase reduction might subtly alter the timing of dendrite degeneration, we assayed earlier timepoints, and again found that the dendrite degeneration proceeded was not influenced by caspase reduction (Fig 3A). However, Dronc RNAi did result in a significantly higher level of axotomy-induced NP compared to control cells (Fig 3B and 3C) consistent with a role for Dronc in negative regulation of NP.
We also tested whether Dcp-1 and Drice, two effector caspases, inhibit axotomy-induced NP. Indeed, both RNAi and a strong loss-of-function mutant of Dcp-1, Dcp-13 [29], increased NP (Fig 3C). Drice RNAi and heterozygous Drice17 [30] neurons also had higher NP than control, but the results were not significantly different. It was not possible to test homozygous Drice17 mutants as these animals die, so we introduced one copy of Drice17, into heterozygous Dcp-13 mutant animals, and this significantly enhanced protection (Fig 3C), indicating both effector caspases are likely to be involved in negative regulation of axon injury-induced protection.
Although in C. elegans and mammals, mitochondrial fission and Drp1 act upstream of caspases [27, 28], in Drosophila the effector caspase Dcp-1 can regulate mitochondrial shape and function [31]. To determine whether caspases were required for injury-induced mitochondrial fission, we assayed fission in Dronc RNAi neurons. Mitochondrial length still decreased in response to axon injury in Dronc RNAi neurons (Fig 3D), suggesting caspases do not act upstream of mitochondrial fission, but may be downstream as in other organisms. However, we cannot rule out that caspases and Drp1 act independently to dampen NP.
Thus far we have identified mitochondrial fission and caspases as negative regulators of NP. We also wished to identify positive regulators. As Nmnat, a conserved NAD+ biosynthetic enzyme, can protect neurons from degeneration induced by long poly-Q proteins [32] and tau [33], we tested whether it might also be involved in injury-induced NP. Indeed, Nmnat RNAi completely eliminated NP induced by axon injury (Fig 4B). To test the specificity of this effect, we also assayed dendrite degeneration without prior axon injury. We did not find any changes in the timing of degeneration in the absence of prior axon injury (Fig 4A) despite previous association of Nmnat with dendrite stability. The previous studies were done in class IV neurons, which have much larger and more complex dendrite arbors than the neurons used here, and exhibit gradual loss of complexity over time in Nmnat heterozygotes [34]. We did not observe any differences in arbor structure in ddaE neurons, and the previous degeneration was observed over days rather than hours.
To confirm a role for Nmnat in injury-induced degeneration, we assayed NP in both Nmnat heterozygous mutant animals (Fig 4B). The mutant is a previously characterized null allele of Nmnat [35]. NP was eliminated in this genetic background (Fig 4B), suggesting that normal levels of Nmnat are required for NP. As the phenotype in the Nmnat RNAi and heterozygous mutant animals was similar, we used an antibody to Nmnat to stain Nmnat RNAi neurons. We observed about 50% reduction in Nmnat signal in these neurons compared to control (S2A Fig), consistent with partial reduction of Nmnat protein in the RNAi experiment.
To determine whether the elimination of NP in animals with a partial reduction of Nmnat were due to a general inability to respond to injury, we tested whether Nmnat RNAi neurons could regenerate axons. When ddaE neurons are axotomized close to the cell body, axon regeneration proceeds by converting a dendrite into a growing axon [36]. Nmnat RNAi neurons were fully capable growing a new axon from a dendrite after a proximal axotomy, indicating the cell can mount at least one demanding injury response (S2B Fig). Together, these results suggest that a partial knockdown of Nmnat in ddaE neurons does not alter the ability of the cell to sense and respond to axon injury. Therefore the loss of NP in this background is most likely due to a specific role of Nmnat in injury-induced protection.
We next tested how Nmnat reduction impacts the increased stabilization that occurs when Miro, Drp1 and Dronc are reduced. We found that Nmnat RNAi completely eliminated the increase in NP caused by Miro, Drp1 and Dronc RNAi (Fig 4C). This result suggests that negative regulation of NP by caspases acts upstream of Nmnat.
To try to position Nmnat relative to other regulators of NP, we examined microtubule dynamics. We previously found that microtubule dynamics, specifically the number of growing plus ends, is dramatically upregulated in dendrites after axon injury in sensory neurons [29], and that this increase in dynamics acts to stabilize dendrites against degeneration [8]. To test whether microtubules and Nmnat protect dendrites in the same pathway or parallel pathways, we labeled the growing ends of microtubules using EB1-GFP in ddaE neurons. We then compared microtubule dynamics in control and Nmnat RNAi neurons. Nmnat reduction specifically abolished the increase in microtubule dynamics at 8hpa without influencing the base-line microtubule dynamics (Fig 4D). Thus the upregulation of microtubule dynamics after axon injury requires Nmnat. It seems unlikely, however, that microtubule dynamics is the sole effector of Nmnat as dampening microtubule dynamics does not block injury-induced protection as strongly or consistently as reducing Nmnat (Fig 4B and 4C and [8]). In summary, our results lead to a model in which Nmnat is a central effector of NP acting upstream of microtubule dynamics and downstream of negative regulation by Drp1 and Dronc (Fig 4E).
As Nmnat seemed so closely linked to NP and was required for both stabilization and increased microtubule dynamics induced by axon injury, we tested whether it was sufficient to induce these responses. We therefore expressed GFP-tagged Nmnat-B-delta-N. The delta-N refers only to a difference from the cDNA used to the annotated cDNA in flybase (see next section in results). In the background of Nmnat-B overexpression we severed a dendrite without prior axon injury and scored its presence 18h later. In control neurons almost no dendrites remained at this time, while in the Nmnat-expressing neurons almost all were intact (Fig 5A). Because Nmnat was sufficient to protect dendrites in the absence of axon injury, we also tested whether it was sufficient to increase microtubule dynamics in uninjured neurons. As a control we expressed a soluble fluorescent protein, Kaede. Expression of either GFP-Nmnat-B-delta-N and Wlds (mouse Nmnat1 with an additional stretch of amino acids at the N-terminus) increased the number of growing microtubule ends in dendrites of uninjured neurons (Fig 5B). Thus Nmnat is not only required for injury-induced NP, but is sufficient both for NP and the associated increase in microtubule dynamics.
While increased stability is likely to help neurons to survive after axon damage, the reason for limiting stability through caspase activity is not intuitive. However, the timing of events triggered by axon injury suggested a hypothesis. Axotomy-induced NP is maximal 8-24h after axon injury in ddaE neurons (Fig 1A and 1A’ and [8]), while axon regeneration typically begins 24-48h after injury in these cells [36]. We therefore hypothesized that turning down early NP might promote subsequent regeneration.
To test whether uncontrolled NP might inhibit regeneration, we compared regeneration in control and Dronc RNAi neurons. In control neurons the average amount of new axon growth 96h after injury was over 200 microns, but in Dronc RNAi neurons the average growth was less than half of that (Fig 6A). Dronc activity therefore promotes regeneration, perhaps by limiting axotomy-induced NP.
If reduced regeneration in Dronc RNAi neurons is due to overactive Nmnat, we predict that Nmnat overexpression should lead to a similar defect in axon regeneration. To test this idea, we generated transgenic flies that encode GFP-tagged Drosophila Nmnat. Two splice forms of Nmnat exist in Drosophila. Nmnat-A contains a nuclear localization signal (NLS) while Nmnat-B does not. The Nmnat-B described in flybase has 31 amino acids at its N-terminus that are not encoded in any existing cDNAs; our GFP-Nmnat-B does not have these 31 amino acids, so we refer to it as Nmnat-B-deltaN.
Consistent with our hypothesis, over-expression of either Nmnat isoform suppressed axon regeneration (Fig 6A). In addition, over-expression of Wlds [37], which includes mouse Nmnat1 and 70 additional amino acids [38], had the same effect (Fig 6A). HA-tagged Nmnat [35] had a similar, although not statistically significant, effect (S3A Fig). These results are consistent with previous studies showing that Wlds overexpression can lead to reduced axon regeneration in a variety of cell types and contexts [39–42].
To determine whether Nmnat might act to dampen regeneration downstream of Dronc, we paired Dronc RNAi with Nmnat RNAi to see if reducing Nmnat would rescue the Dronc RNAi phenotype. To control for potential Gal4 dilution effect when expressing many UAS-driven transgenes together, we paired Dronc RNAi with a control RNAi. Indeed, the addition of the control transgene reduced the effect of Dronc RNAi on regeneration (Fig 6). However, in Dronc plus Nmnat double RNAi neurons, regeneration was significantly enhanced compared to the matched control (Fig 6B). This result is consistent with Nmnat acting as a negative regulator of regeneration downstream of Dronc. Levels of regeneration in this experiment were higher than in other genetic backgrounds. It is possible that Dronc also targets positive regulators of regeneration that can increase outgrowth when Nmnat-mediated in inhibition of regeneration is reduced.
Although we found Nmnat was central to NP, we did not see large changes in amount or distribution of endogenous Nmnat in ddaE neurons after injury using immunofluorescence (S3D Fig). This may be because small or transient changes in levels or activity of Nmnat are sufficient to stabilize dendrites, and because endogenous Nmnat was difficult to detect. GFP-Nmnat-B-deltaN was evenly distributed in the nucleus and cytoplasm and did not change its localization in response to injury (S3B Fig). In uninjured ddaE neurons, GFP-Nmnat-A was detected primarily in nuclei (S3C Fig). At 8h post axon injury, the ratio of nuclear to cytoplasmic Nmnat-A signal was significantly decreased (S3C Fig). In contrast, the ratio did not change in response to axotomy in Dronc RNAi neurons (S3C Fig). Although we do not know the significance of the decrease in nuclear Nmnat-A relative to cytoplasmic, the fact that it is dependent on Dronc is consistent with Dronc regulating Nmnat after axon injury.
There are two ways excessive Nmnat could dampen regeneration: either by generating a persistent stump that blocks regeneration or more directly from within the regenerating cell. A persistent axon stump could block or repel new axon growth, as has been demonstrated in zebrafish [42]. In the regeneration assay used here, physical block by the stump cannot be important as the new axon grows from a dendrite on the opposite side of the cell. To test whether a persistent stump might influence regeneration in some other way, we expressed the Wlds protein in ddaC neurons, which are next to ddaE neurons. This approach enabled generation of a persistent stump near a cell body that did not itself express extra Wlds or Nmnat. When we severed axons of both the Wlds-expressing cell (ddaC) and wild-type ddaE, the ddaC axon persisted as expected, and regrowth of the axon from a dendrite occurred normally in the ddaE neuron (S3E Fig). Failure of a neighboring persistent stump to reduce regeneration is consistent with excessive Wlds or Nmnat acting cell-autonomously to dampen regeneration.
Thus far we have shown that Nmnat is required for NP, that caspases limit NP, and that overactivation of NP dampens regeneration. However, there must also be positive signals that turn NP on in response to axon injury. Indeed, we previously showed that JNK is required for NP mediated by increased microtubule dynamics [8]. JNK can act downstream of DLK in initiation of axon regeneration in an injury-induced cascade that results in fos-mediated transcription in Drosophila [10]. We therefore tested whether DLK and fos played a role in NP. A trans-heterozygous combination of wnd alleles (wnd is the name for Drosophila DLK), and a fos dominant negative (fosDN) transgene have been shown to block injury signaling [10] and regeneration [43], so we used these tools to test for a role in NP. In both genetic backgrounds induction of NP by axon injury was completely blocked (Fig 7A). In addition, fosDN blocked the increased microtubule dynamics in dendrites after axon injury (Fig 7B). We conclude that the DLK/JNK/fos pathway is required for NP and its associated upregulation of microtubule dynamics. The NP that protects dendrites in sensory neurons may therefore be similar to the DLK and fos-mediated axon stabilization induced by crushing motor axons [7].
We also tested whether the fos pathway might be upstream of mitochondrial fission induced by axon injury. Unlike control neurons (Fig 2A and 2B), no decrease in mitochondrial length was observed in fosDN neurons (Fig 7C). Thus fos activity is required for injury-induced mitochondrial fission. This suggests fos is required both to turn on NP and to induce mitochondrial fission that limits NP.
If fos is a critical regulator of NP, then its overexpression might alter the time course of dendrite stabilization. Indeed, in control neurons NP is low 48h after axon severing, but in fos overexpressing neurons it remained high (Fig 7D). This is consistent with previous studies showing that fos can stabilize axons in other situations [7, 44]. Overexpressing fos also blocked axon regeneration (Fig 7E). To determine whether the inhibition of regeneration by fos was due to excessive NP, we co-expressed Nmnat RNAi with fos. As in other experiments with multiple transgenes we paired fos with a control RNAi so that transgene number was matched. Nmnat RNAi completely rescued regeneration in the fos overexpression (Fig 7E). We conducted similar experiments with overexpressed bsk, the JNK homolog in Drosophila. Like fos, bsk overexpression extended protection (Fig 8A), and blocked regeneration in a Nmnat-dependent manner (Fig 8B). Together these results demonstrate that overexpression of fos or JNK extends the normal timing of NP and, in a Nmnat-dependent manner, reduces regeneration. Thus Nmnat is a both a positive regulator of NP and a negative regulator of regeneration that can act downstream of JNK and fos signaling.
Our results lead to a model (Fig 8C) in which axon injury triggers opposing responses downstream of the initial DLK/JNK/fos signaling cascade. One early output of this conserved injury response pathway is NP, a global stabilization of the parts of the neuron still connected to the cell body. The central mediator of NP is Nmnat. One Nmnat effector is the dramatic increase in microtubule dynamics observed after axon injury. As axon damage is likely to be accompanied by disturbances in the surrounding tissue, making the cell more resistant to degeneration by turning on NP may help the neuron survive the initial trauma.
Fos injury signaling also triggers Drp1-mediated mitochondrial fission in the first few hours after axon injury, and this leads to dampening of NP by caspases. We envision positive and negative regulation of NP balancing one another in different ways through time after injury. Eventually the negative pathway must outweigh the positive or regeneration is dampened by persistent NP (Fig 8D). It is possible that the timing of this balance shift is controlled by additional signals that report whether the environment is conducive for regeneration.
This model suggests that rather than DLK/JNK/fos directly regulating regeneration, this signaling pathway kicks off a multi-step response to axon injury that includes regeneration as a relatively late event. Indeed, although this pathway is known as the conserved axon regeneration pathway, we find that it first turns on a response that inhibits regeneration. Although this idea is surprising, this model does make sense in the overall picture of neuronal injury responses and stabilization. For example, in mammals [45, 46] and flies [10] the AP-1 transcription factor fos is activated soon after axon injury, but its role in regeneration is not as clear as that of some other transcription factors like jun. Our data suggests that this early activation could be because fos orchestrates the injury response that precedes regeneration.
Our results also touch on the role of caspases in axon regeneration. A study in C. elegans demonstrated that caspases are positive regulators of axon regeneration [47], which is surprising considering their involvement in self-destruct programs like apoptosis and dendrite pruning. We confirm that in Drosophila caspases are pro-regenerative. In addition, our data suggests that this effect is not through a direct role in regeneration, but because caspases down-regulate NP, which inhibits regeneration.
A negative role for mitochondria in NP is also intriguing. Mitochondria seem to promote axonal stability [15], and there are studies in several systems that suggest the neuroprotective effects of Nmnat or Wlds require mitochondria [17–19]. However, mitochondria can play prodegenerative roles in other contexts [26, 48]. More specifically mitochondrial fission can promote degeneration [49]. Here we demonstrate that mitochondria, Drp1 and caspases all counteract NP, suggesting that caspase activation may regulate NP downstream of mitochondrial fission. This does not mean that mitochondria are not also positive regulators of this type of NP. Indeed the data in this study combined with others suggests that mitochondria are critical nodes for control of neuronal stability and both positive and negative regulation likely converge on them.
Like mitochondria, the role of Nmnat in injury responses has been difficult to classify simply as either positive or negative. Its ability to prevent injury-induced Wallerian degeneration, as well as to act as an endogenous neuroprotective factor [50] has led to the idea that it has a purely positive influence on neuronal health. However, the myriad ways in which it can be regulated [50] suggest that it is useful only in exactly the right dose. Indeed we show that when its regulation is disrupted, Nmnat inhibits a different type of neuronal resilience: axon regeneration. Thus upregulation of Nmnat as a potential therapeutic strategy to counteract neurodegeneration could have negative outcomes due to dampened regeneration.
While our experiments support the idea that endogenous Nmnat is a central regulator of neuronal stability, the way it exerts this effect remains unclear. Nmnat is an enzyme that uses ATP and NMN (nicotinamide mononucleotide) to make NAD+. Protective effects of endogenous or overexpressed Nmnat have been proposed to be due to maintenance of high NAD levels [51–53], keeping levels of the precursor NMN low [54], acting as a chaperone [32, 35], and through maintaining mitochondrial integrity or function [17–19]. We now show that Nmnat also acts upstream of increased microtubule dynamics after axon injury. This Increased microtubule dynamics in response to axon injury is also seen in mammalian neurons, and so this part of the NP response is likely to be conserved [13]. Although we have previously shown increased microtubule dynamics plays a role in NP [8], and now show that Nmnat overexpression is sufficient to increase microtubule dynamics, it is possible that Nmnat has other effectors that can mediate NP.
In conclusion, we propose a model in which DLK signaling initiates key injury responses before axon regeneration begins. These responses include upregulation of Nmnat-mediated NP, microtubule dynamics and mitochondrial fission. Mitochondrial fission likely counteracts NP through caspase activation, although it is possible that mitochondria and caspases regulate NP independently. Although this early response is downstream of the core axon regeneration kinase cascade, it actually inhibits regeneration if unchecked. This multi-step model of injury responses downstream of DLK helps explain the function of caspases in promoting regeneration. We anticipate that understanding the transition between early injury responses and regeneration itself will suggest strategies for promoting axon regeneration without overactivating NP, which would, in turn, dampen regeneration. A more complete understanding of the relationship between NP and regeneration is essential to designing any therapeutic approach to either stabilize neurons or to enhance regeneration.
The following RNAi fly strains were used in this study: Rtnl2 (33320) [8], gammaTub37C (25271) [8], Miro (106683) [22], milton (41508) [55], Dronc (23035) [22], Dcp-1 (107560) Drice (28065) [56] and Drp1 (44156, referred to as #2) from the Vienna Drosophila RNAi Center, and Drp1 (27682), Nmnat (29402) [18] from the Bloomington Drosophila Stock Center (BDSC). All RNAi transgenes were coexpressed with UAS-Dcr2 to increase knockdown efficiency [57]. Other lines include 221-Gal4, ppk-Gal4, UAS-mito-GFP (BDSC 8443), UAS-EB1-GFP, UAS-mCD8-RFP, UAS-Drp1 [58], UAS-Wlds [37], Nmnatdelta4790-8 [35], Drice17 [30], Dcp-13 [29], UAS-fosWT (BDSC 7213), UAS-bsk-A.Y (BDSC 6407), wnd1, wnd3, and UAS-fosDN [59].
Fly embryos were collected at 20C overnight and aged at 25C for 2 or 3 days before imaging. Two days of aging was used for all axon regeneration assays because these extend 96h, and 3 days of aging were used for all other experiments. Larvae were mounted between a slide coated with a dry agarose patch and a coverslip, which was held in place with sticky tape. A MicroPoint pulsed UV laser (Andor Technology) was used to injure dendrites and axons of ddaE neurons expressing EB1-GFP or mCD8-RFP under the control of 221-Gal4. Confocal images were acquired using a Zeiss LSM510 with a 63X oil objective (NA1.4) right after injury. Larvae were then kept in individual food caps at 20C for the indicated time periods and were then reimaged using an Olympus FV1000 confocal microscope equipped with a 60X oil objective (NA1.42). Maximum intensity projections were generated using ImageJ software, and were aligned and processed using Adobe Photoshop software.
To measure microtubule dynamics in dendrites after injury, we imaged neurons for at least 100 frames (1 frame/2s) using an Olympus FV1000 microscope with a 60X objective at zoom 3, and counted the total number of EB1-GFP comets in a 10 μm dendrite segment close to the cell body from 3 in-focus frames. Only comets moving in 3 consecutive frames were included for quantification. The reslice tool of ImageJ was used to generate kymographs with 1 pixel spacing.
In uninjured neurons, EB1-GFP-expressing neurons were imaged with a Zeiss AxioImager M2 equipped with LED illumination and an AxioCam 506 camera. A 63x 1.4 NA objective was used to acquire images every second. After image capture, analysis was performed in ImageJ. In each movie, the length of the comb dendrite of the ddaE neuron that was in focus was measured. EB1-GFP comets that passed through this region during the 300 seconds of the movie were counted and this number was divided by the length to get comets per length. The time (300s, with 1 frame per second) was the same in all movies and so was not included in the normalization.
Live imaging of mitochondria was performed by expressing UAS-mCD8-RFP and UAS-mito-GFP under the control of 221-Gal4. Images were taken on a Zeiss LSM510 at 1 frame/s using a 63x objective and 2x zoom. Injury-induced mitochondrial behavior changes were analyzed in dendrites in a 66.8 μm2 region close to the cell body. The template matching plugin in ImageJ was used to minimize the effect of larval body movements. Mitochondrial length was measured along the longest dimension of mito-GFP shapes using the measure tool in ImageJ. The average length of mitochondria was calculated from at least 8 neurons, each of which contained 13–52 mitochondria in the imaging region. To further analyze changes in length, mitochondria were grouped according to the length of mito-GFP, and the percentage of each group was calculated before and after injury.
ddaE neurons expressing EB1-GFP were axotomized close to the cell body and reimaged after 96h. One dendrite usually extends and converts into a new axon by 96h in response to a proximal axon injury [36]. using the NeuronJ plugin in the ImageJ software, we measured the length of the specified dendrite at 0h (R0h) and 96h (R96h), and a nearby non-regenerating dendrite (NR0h and NR96h) so that normal dendrite expansion as the larva grows could be taken into account. The formula R96h-R0h*NR96h/NR0h was used to calculate growth of the new axon tip.
The comb dendrite of ddaE neurons was severed close to the cell body. Degeneration speed was measured by scoring morphology of the severed dendrite at 4, 7, and 11. Dendrites with no discontinuities were scored as intact, and any breaks resulted in them being scored as not intact.
ddaE neurons expressing EB1-GFP were axotomized close to the cell body 8h or 48h before the dorsal comb-like dendrite was severed. Dendrite status was determined 18h post dendrite injury. In all the neurons we examined, dendrites either completely degenerate and no remnants are left or remained intact. Example images of both types of result are shown in Fig 1B. Therefore, NP is measured using the percentage of neurons with intact dendrites at 18hpd.
GraphPad Prism 6 software was used to generate graphs and perform statistical analysis. A Fisher’s exact test was used to determine significance of neuroprotection assays and length distribution of mitochondria. Other types of data were tested for normal distribution using D’Agostino-Pearson normality test. If the data passed the normal distribution test, a t test or in Fig 6A, a one way ANOVA followed by Dunnett’s multiple comparison test, was used to determine statistical significance. Otherwise, a Mann-Whitney test was used. Details of the specific t test performed and sample size for each experiment are described in Fig legends. Data were plotted as mean ± standard deviation (SD). * p<0.05, ** p<0.01, *** p<0.001.
The coding sequence of Drosophila Nmnat isoform A was amplified from a cDNA clone using forward primer 5’-CCGGAATTCATGATTGTGAAAATCAGCTGGCCCAAG-3’ and reverse primer 5’-ATATGCGGCCGCCTAAAGTTGCACTTGGGAAATC-3’.
The coding sequence of Drosophila Nmnat isoform B was amplified from the UAS-Nmnat.HA construct [35] using forward primer: 5’-CCGGAATTCATGTCAGCATTCATCGAGGAAAC-3’ and reverse primer: ATATGCGGCCGCTCAAGAGTCGCATTCGGTCGGAG.
Both forward primers contain an EcoRI site and reverse primers contain a NotI site. The amplified sequences were cloned into a pUAST-GFP vector and the resulting constructs were injected into fly embryos to generate several transgenic flies. UAS-GFP-Nmnat-A4 and UAS-GFP-Nmnat-B-deltaN8 lines were used in this study.
|
10.1371/journal.ppat.1006182 | Boosting of HIV envelope CD4 binding site antibodies with long variable heavy third complementarity determining region in the randomized double blind RV305 HIV-1 vaccine trial | The canary pox vector and gp120 vaccine (ALVAC-HIV and AIDSVAX B/E gp120) in the RV144 HIV-1 vaccine trial conferred an estimated 31% vaccine efficacy. Although the vaccine Env AE.A244 gp120 is antigenic for the unmutated common ancestor of V1V2 broadly neutralizing antibody (bnAbs), no plasma bnAb activity was induced. The RV305 (NCT01435135) HIV-1 clinical trial was a placebo-controlled randomized double-blinded study that assessed the safety and efficacy of vaccine boosting on B cell repertoires. HIV-1-uninfected RV144 vaccine recipients were reimmunized 6–8 years later with AIDSVAX B/E gp120 alone, ALVAC-HIV alone, or a combination of ALVAC-HIV and AIDSVAX B/E gp120 in the RV305 trial. Env-specific post-RV144 and RV305 boost memory B cell VH mutation frequencies increased from 2.9% post-RV144 to 6.7% post-RV305. The vaccine was well tolerated with no adverse events reports. While post-boost plasma did not have bnAb activity, the vaccine boosts expanded a pool of envelope CD4 binding site (bs)-reactive memory B cells with long third heavy chain complementarity determining regions (HCDR3) whose germline precursors and affinity matured B cell clonal lineage members neutralized the HIV-1 CRF01 AE tier 2 (difficult to neutralize) primary isolate, CNE8. Electron microscopy of two of these antibodies bound with near-native gp140 trimers showed that they recognized an open conformation of the Env trimer. Although late boosting of RV144 vaccinees expanded a novel pool of neutralizing B cell clonal lineages, we hypothesize that boosts with stably closed trimers would be necessary to elicit antibodies with greater breadth of tier 2 HIV-1 strains.
Trial Registration: ClinicalTrials.gov NCT01435135
| Developing a successful HIV-1 vaccine remains a high global health priority. Several HIV-1 vaccine trials have been performed with only the RV144 vaccine trial showing vaccine efficacy, albeit modest. No broadly neutralizing antibody activity was identified in RV144 and inducing sterilizing immunity against a complex pathogen like HIV-1 remains a major challenge. Here we characterize the B cell responses after RV144 vaccine-recipients received two additional boosts severals years after the conclusion of the RV144 vaccine trial. Delayed and repetitive boosting of RV144 vaccine-recipients was capable of increasing somatic hypermutation of the Env-reactive antibodies and expanding subdominant pools of neutralizing B cell clonal lineages. These data are pertinent to HIV-1 vaccine-regimen design.
| Six HIV-1 vaccine efficacy trials have been performed [1–5], of which only one, the ALVAC-HIV and AIDSVAX B/E prime-boost RV144 trial, showed vaccine protection, with estimated vaccine efficacies of 60% at 12 months [6] and 31% at 42 months [7]. Plasma IgG antibodies binding to HIV-1 envelope variable region 2 (V2) and low Env IgA binding levels were immune correlates of decreased transmission risk [8]. V2-specific antibodies isolated from RV144 bound tier 2 HIV-1 infected CD4 T cells and mediated antibody dependent cellular cytotoxicity (ADCC) [9].
While no broadly neutralizing antibodies (bnAbs) were induced in RV144 [8,10] the induction of bnAbs remains a prime goal of HIV vaccine development, since passive administration of bnAbs has repeatedly shown to protect against simian HIV-1 (SHIV) chimeric virus challenge [11–15]. BnAbs develop in approximately 50% of HIV-1 infected individuals, but these arise only after several years of infection [16,17]. One hypothesis to explain why HIV-1 bnAbs have been difficult to induce by vaccination is that these antibodies have one or more unusual characteristic—long HCDR3 regions, autoreactivity with host antigens, and/or extensive somatic mutations—all traits of antibodies controlled by host tolerance control mechanisms [18–22]. A result of tolerance control of bnAbs is that bnAb precursors may be reduced in frequency in the pre-vaccination B cell repertoire; they may also be at a competitive disadvantage with other more dominant precursor B cell pools. For these reasons, inducing bnAbs may require an extensive vaccination-regimen.
Here we sought to determine if a pool of subdominant B cells, such as those that produce long HCDR3 CD4 bs bnAbs, may be expanded when an Env immunogen that binds bnAb UCAs is included in a boosting regimen. In the RV305 clinical trial, RV144 vaccine-recipients who had previously received the initial ALVAC-HIV + AIDSVAX B/E gp120 immunization regimen (0,1,3,6 months) and remained HIV-1- uninfected were boosted with ALVAC-HIV, AIDSVAX B/E gp120, or ALVAC-HIV + AIDSVAX B/E gp120 6–8 years later (S1 Fig). We found that boosting of RV144 vaccinees led to an increased frequency of memory B cells producing envelope-specific antibodies with long HCDR3s. Several of the mature antibodies and inferred unmutated common ancestors (UCA) neutralized both neutralization sensitive HIV-1 isolates (tier 1) and a difficult-to-neutralize (tier 2) HIV-1 CRF01 AE isolate, CNE8.
After two boosts (6-month interval) with the same immunogens 6–8 years after the completion of the RV144 primary immunizations (S1 Fig), plasma neutralizing antibody (nAb) responses were assayed in the A3R5 pseudovirus neutralization assay [23] against a panel of 11 CRF01 AE isolates (S2A Fig). Previous work has shown that neutralization of neutralization resistant (tier 2) HIV-1 isolates by antibodies is more readily detected in the A3R5 cell based assay than in the TZM-bl cell based assay [23]. Here the A3R5 cell based assay was used to search for vaccinees who had robust antibody responses to Env. We selected four vaccinees for study who had high magnitude and breadth of neutralization. Two were from RV305 Group 1 who received ALVAC-HIV plus AIDSVAX B/E gp120 boosts (3043, 3070), and two were from RV305 Group 2 who received only AIDSVAX B/E gp120 boosts (3064, 3053) (S2B Fig). In all four vaccinees, the RV305 boosts increased both autologous (AE.A244gp120) and heterologous (B.6240gp120) plasma IgG-gp120 binding responses, to levels higher than those observed after the initial RV144 regimen (S3A Fig) The RV305 boosts also increased the magnitude of B.MN and AE.92TH023 neutralization in the TZM-bl neutralization assay by plasma from all four vaccinees, but there was no plasma tier 2 neutralizing activity seen (S3B Fig).
We isolated AE.A244 gp120 Env-specific post-RV305 boost memory B cells from the four vaccinees- 3043, 3070, 3064 and 3053 (S4 Fig) and from the same vaccinees post-RV144 samples for three of the four vaccinees for whom PBMCs were available. Comparison of the gp120-reactive mAbs from post-RV144 (n = 184 mAbs) with the gp120-reactive mAbs from post-RV305 (n = 242 mAbs) showed that the mean VH nucleotide mutation frequency increased over 2-fold in each vaccinee from a mean of 3.1% to 6.9% (Fig 1A and 1C).
Mobilizing and expanding the pool of long HCDR3 antibodies will be critical for the eventual induction of V2-glycan, V3-glycan, or HCDR3-loop binding bnAbs since many of these bnAbs have HCDR3s longer than 22 amino acids (aa) [24–28]. A meta-analysis of antibodies isolated from post-RV144 studies found that the frequency of Env-reactive B cells with HCDR3s ≥ 22 aa was 2.1%. An analysis of the post-RV305 antibodies indicated that the frequency of Env-reactive B cells with HCDR3s ≥ 22 aa was 20.7% (S5 Fig). To confirm that the increased frequency of Env-reactive long HCDR3 mAbs was related to late boosting, we analyzed the B cell repertoires of three of the four vaccinees (3043, 3053 and 3064) for whom blood samples were available both 2 weeks after the initial RV144 immunization and 2 weeks after the RV305 immunizations. The average frequency of Env-reactive long HCDR3 antibodies within the same vaccinees increased from 7.6% to 20.7%. (Fig 1B and 1D). The HCDR3 length is dictated primarily by V(D)J recombination and can be diversified through secondary means: VH replacement, D-D fusion, insertions, N-nucleotide addition and P-nucelotide addition. Long HCDR3 antibodies have been shown to be biased towards DH2, DH3 gene and JH6 gene segment usage [29]. Coinciding with this observation 72% of the Env-reactive long HCDR3 antibodies isolated post-RV305 and utilized DH2 or DH3 and 58% used JH6 (S1 Table). To determine if this phenomenon was unique to B cell repertoires from late boosting of RV144 vaccinees, we compared these data with the frequency of Env-reactive long HCDR3 found in other HIV-1 Env based immunization regimens. In the GSK PRO HIV-002 human clinical trial, vaccine-recipients received gp120 immunizations in AS01B adjuvant, and the frequency of Env-reactive mAbs with long HCDR3s was 6.9% (n = 58) [30]. In the DNA prime Ad5 boost HIV-1 vaccine regimen used in the HVTN 505 efficacy trial, the frequency of gp140-reactive mAbs with long HCDR3s was 4.1% [31] (S2 Table p< 0.05 compared to RV305 boost data; Fisher’s Exact Test). These data suggested that other immunization regimens without boosting did not expand memory B cell pools with long HCDR3s to the extent achieved with the RV305 boosts.
In vaccinee 3053, seven gp120-reactive B cell clonal lineages were present after the initial RV144 vaccine regimen that persisted and had expanded after boosting 6–8 years later in RV305, one of which, DH678, had a long HCDR3. In vaccinee 3043 nine gp120-reactive B cell clonal lineages were identified after RV144 that were also represented in the samples taken after the RV305 boosts. The antibodies in two of these lineages, DH686 and DH576, had long HCDR3s (S6 Fig). These data demonstrate that memory B cells producing antibodies with long HCDR3s were induced by the initial RV144 regimen and could be expanded with boosting 6–8 years later.
All antibodies isolated were assayed by ELISA as transient transfection supernatants and we selected twenty-seven Env-binding antibodies derived from blood memory B cells post-RV305 boosts based on HCDR3-length (≥ 22 aa) as a representative set of antibodies for characterization (S3 Table). Nine of the 27 mAbs neutralized the neutralization sensitive (tier-1) virus AE.92TH023 in the TZM-bl neutralization assay [23,32,33](S4 Table). The epitopes of these nine neutralizing mAbs with long HCDR3s were then mapped by ELISA for activity in blocking soluble (s) CD4 binding to Env and for binding to mutant Envs. All 9 long-HCDR3 antibodies that neutralized HIV-1 blocked sCD4 binding by ≥70% (Fig 2A) and also blocked binding of CD4 bs bnAbs VRC01 and CH31 (S7 Fig).
Env mutations I371, P363, R476 and D368 generally reduce binding by CD4bs Abs [34]. When assayed with Δ371I/P363N and D368R CD4bs Env mutants, binding of three neutralizing mAbs (DH576, DH576.2, and DH577) was measurably lower compared to wild-type Env (Fig 2B). Seven of nine long HCDR3 sCD4 blocking mAbs (Fig 2A) bound to B.YU2gp120. The binding epitopes of these seven mAbs were mapped by yeast display using B.YU2gp120 core (ΔV1, V2, V3 loops) and B.YU2gp120 cores with mutations that reduce binding by known CD4bs Abs [35]. In contrast to epitope mapping on A244gp120 Env binding of six of seven mAbs were D368R sensitive (Fig 2C). The four Abs not sensitive to the D368R mutation in A244gp120 likely have a higher affinity for the A244gp120 protein then YU2gp120 and their epitope is less dependent on Env D368. Abs DH576 and DH576.2 shared with the CD4bs bnAb B12 sensitivity to 3 CD4bs-critical mutations (D368R, R419G, T455E) and 2 of 3 additional mutation sensitivities (K282V and I467K) [36,37] suggesting these vaccine-induced CD4bs mAbs have a specificity more similar to that of B12 than to that of the non-bnAb CD4bs mAb, B6 which is not sensitive to D368R, R419G and T455E mutations (Fig 2C).
In the TZM-bl cell assay, all neutralizing CD4bs mAbs neutralized not only AE.92TH023 but also the heterologous tier 2 CRF01 isolate AE.CNE8 isolate. DH583 was the broadest neutralizing antibody, also neutralizing the tier 1 viruses B.SF162, B.MN, and the tier 1B (intermediate neutralization sensitivity) primary isolate C.6644 (Fig 2D). Long HCDR3 neutralizing mAbs were assayed against four additional tier 2 CRF01 AE isolates but showed no additional neutralization breadth (S5 Table).
In RV144, infection risk correlated inversely with V1V2 antibody responses [8]. Two V1V2 binding antibodies, CH58 and CH59, neutralized the autologous tier 1 isolate AE.92TH023 in the TZM-bl neutralization assay and also mediated ADCC against tier 2 virus infected cells [9]. To determine whether the long HCDR3 CD4bs mAbs isolated after the RV305 boosts also mediated ADCC, the Abs were expressed in an IgG1 backbone optimized for FcγRIIIa binding[38] and assayed for ADCC against virus-infected cells. DH583 mediated ADCC against B.WITO and C.1086C virus infected cells, with an endpoint concentration of approximately 0.1μg/ml and overall ADCC activity, as evaluated by positive area under the dilution curve, similar to that observed for the CD4bs bnAb CH31. The other eight long HCDR3 CD4 bs mAbs had little to no ADCC activity against any of the isolates tested (S8 Fig).
The most heavily mutated member of the long HCDR3 CD4bs DH576 B cell clonal lineage was DH576.2 (VH nucleotide mutations of 10.33%), but the additional mutations did not broaden or strengthen HIV-1 tier 2 CRF01 AE.CNE8 neutralization (Fig 2) with respect to neutralization by less mutated lineage members such as DH576 (VH mutations of 7.33%) (S3 Table). To determine the effects of affinity maturation, we assayed the UCA, IAs and three naturally occurring DH576 clonal lineage mAbs for neutralization of the autologous tier 1 virus AE.92TH023 and the heterologous tier 2 virus CRF01 AE.CNE8. The DH576 UCA neutralized both the tier 1 HIV-1 AE.92TH023 and the tier 2 HIV-1 CRF01 AE.CNE8. As affinity maturation progressed, there was a difference in the ratio of neutralization potencies for tier 1 and tier 2 viruses. Affinity maturation increased DH576 ineage neutralization potency (IC50) against the tier 1 AE.92TH023 by over 3 logs, but increased its potency (IC50) against the tier 2 AE.CNE8 by less than 1 log (Fig 3). These data can be explained in part as follows. The UCA of DH576 had a higher affinity for AE.CNE8gp120 than did the UCA for AE.A244gp120 (nearly identical in sequence to AE.92TH023). Binding assays to the two gp120s showed affinity maturation of < 1 log to AE.CNE8gp120 while there was > 2 log increase in affinity maturartion for AE.A244gp120 (S6 Table).
To determine whether neutralization of HIV by a UCA was a common property of HCDR3-loop CD4 bs binding mAbs, we assayed the UCAs of the other vaccine-induced CD4bs mAbs and found that 3 of 8 nAb UCAs neutralized both AE.92TH023 and AE.CNE8 (S7 Table).These data indicated that the vaccination regimens in both RV144 and RV305 trials could elicit long HCDR3 CD4bs mAbs, whose germline genes could mediate tier 2 neutralization of HIV-1 AE.CNE8.
Progression from sporadic tier 2 neutralization to increased tier 2 virus neutralization breadth depends upon the epitope specificity [39] and the precise footprint of the Ab on Env [26]. We analyzed by negative stain electron microscopy (EM) a CH505 SOSIP.664 trimer bound with DH576. A 3D reconstruction showed DH576 bound to an open trimer—that is, to Env in a conformation related to the one stabilized by CD4 binding (Fig 4A, S9 Fig). A top view of the complex suggested that the DH576 footprint might resemble those of bnAbs B12 and CH103 (Fig 4B). The bnAbs CH103, CH235, CH31, VRC01, and PGV04, as well as CD4 itself, project away from the center of the trimer, avoiding interference with adjacent gp120 subunits in the closed trimer conformation, whereas DH576 may require the open form in order to avoid overlap. The DH576 Fab has an orientation with respect to Env quite similar to that of the B12 Fab, but turned by ~90° about its long axis (Fig 4B).
The CD4bs bnAb B12 interaction with gp120 depends upon an aromatic residue at the apex of the HCDR3 loop, aromatic residues around the base of the HCDR3 region, a tyrosine at the apex of the HCDR2 loop and positively charged amino acids in the LCDR1[40]. An alignment of the DH576 inferred UCA and naturally occurring clonal lineage members with the B12 heavy sequence showed that, like B12, the DH576 clonal lineage contained an aromatic residue at the apex of the HCDR3 loop, aromatic residues around the base of the HCDR3 and a tyrosine in the HCDR2 loop (S10 Fig). The HCDR3s of B12 and DH576 protrude at different angles and when DH576 is superimposed on the B12-gp120 complex, the HCDR3 of DH576 sterically clashes with gp120. Thus it is not suprising that DH576 rotates by approximately ~90° when it binds to gp120 (S10 Fig).
Negative stain EM of 92Br SOSIP.664 with DH583, the broadest mAb identified, showed that DH583 also binds an open form of the trimer, even though this trimer is stable in the closed form (S11 Fig). These observations suggest that antibodies elicited in the RV305 trial bind epitopes generally shielded in closed trimers, consistent with the use of gp120 (rather than a closed Env trimer) as a principal component of the original, RV144 vaccine.
In this paper we demonstrate that late (6–8 year) boosting of RV144 vaccinees with ALVAC-HIV and AIDSVAX gp120 B/E increased the VH chain gene mutation frequency and expanded clonal lineages of CD4bs antibodies with long HCDR3 regions. Increased somatic hypermutation and affinity maturation by repetitive immunization with a gp120-protein has previously been reported in humans and non-human primates [30,41]. In this study the boosting of RV144 vaccinees occurred several years later suggesting that in spite of the rapid waning in plasma IgG seen in the RV144 vaccine trial, long lived memory B cells were induced that could be recalled with subsequent boosting. The observation that three CD4bs clonal lineage UCAs could neutralize tier 2 CRF01 AE AE.CNE8 raised the hypothesis that the AE.A244 gp120 Env in the boost selectively stimulated expansion of a pool of pre-existing tier 2 neutralizing clonal lineages. An antibody HCDR3 arises from recombination of immunoglobulin heavy variable (VH), diversity (DH), and joining (JH) genes; its overall length is determined by gene usage [20,29,42], D-D fusion [25,42,43], N nucleotide additions [22,42], or VH gene replacement [44,45]. While B cells that give rise to long HCDR3 antibodies frequently undergo productive gene rearrangement [42], they can experience negative selection during B cell development because of autoreactivity or polyreactivity [21,22]. Thus, in uninfected individuals, only approximately 4% of the naïve repertoire consists of long HCDR3 antibodies, and this population contracts by ~ 50% due to negative selection in the bone marrow at the first immune tolerance checkpoint [22,29].
Virus neutralization by a fully reverted, inferred UCA has been reported for V1V2 and CD4bs bnAbs [25,46–49] that came from HIV-1 chronically infected individuals. Pancera et al [49] and Bonsignori et al [25] found that V1V2 bnAb UCAs of PG16 and CH01 could neutralize several primary HIV strains. Both UCAs neutralized clade C ZM233, clade A AQ23 and clade B WITO[25,49]. More recently Gorman et al [47] and Andrabi et al [48] have shown that the combining sites of multiple V1V2 bnAbs share binding motifs, and their UCAs frequently neutralize the same HIV-1 primary isolates, suggesting that these primary isolate Envs might be candidates for use as immunogens.
The fundamental question raised is whether the CD4bs B cell clonal lineages primed by RV144 and expanded with the repetitive boost of the same vaccine can, with continued boosting, affinity mature into bnAbs. The epitopes of the vaccine-induced CD4bs mAbs described here appear to overlap those of other CD4bs antibodies and that of bnAb B12 in particular. Electron microscopy of negatively stained complexes showed that the vaccine-induced mAbs DH576 and DH583 bound an open form of Env, consistent with a gp120 being used in the vaccine-regimen, and the images were consistent with the CD4bs epitope mapping. Seven of the nine long HCDR3 CD4bs mAbs characterized here had the same VH3 gene usage as the CD4 bs bnAbs CH98 [36] and HJ16[50]; one of the nine used VH1-69 (S3 Table), like VRC13 [26]. One mAb also used a VL κ4–1 like HJ16 and six of the nine long HCDR3 CD4 bs mAbs used either a VL κ3–20 or VL κ1–33, which are VL chain genes used by the CD4bs bnAbs B12, VRC01, VRC-PGV04, VRC30-34, 3BNC117, 3BNC60, NIH45-46, 12A12, 12A21 and 8ANC131 (reviewed in [16]). Nonethless, after 6–8 years and 4 boosts, the induced mAbs neutralized only 1 of 40 tier 2 viruses that were assayed with DH583 and DH576. Moreover, the neutralizing IC50 of the DH576 clonal lineage for CRF01 AE.CNE8 changed only marginally during affinity maturation, strongly suggesting that theAE.A244gp120, although it could bind to the UCA, did not select clonal lineage members that could undergo affinity maturation and exhibit greater breadth. Rather it was only neutralization of the tier 1 virus AE.92TH023 for which vaccine boosting led to a 3 log increase in IC50. Thus, it is likely that AE.A244 gp120 selected antibody responses that neutralized viruses with an “open” Env conformation, consistent with known conformational properties of the free gp120 fragment. As previously shown in non-human primates antibodies that exclusively bind an open Env sterically clash with Env variable regions leaving little chance of maturing to a bnAb [51,52]. We do not yet know whether a de novo series of prime-boost immunizations with stable, closed trimer as proposed by others [51,53,54] would engage the UCAs of long HCDR3 antibodies such as DH576 and induce affinity maturation to neutralization breadth. In general, Envs of tier 1 viruses open readily, while those of tier 2 viruses do not. The Env of CRF01 AE.CNE8 apparently opens readily enough to bind the antibodies we have characterized, but most other tier 2 Envs do not. The boosts that expanded the pool of long HCDR3 mAbs occurred several years after the completion of the RV144 trial. We do not know what effect the interval between boosting has on the vaccine-induced antibody repertoire. In the RV306 HIV-1 clinical trial (NCT01931358), vaccine-recipients received the same ALVAC-HIV and AIDSVAX B/E prime-boost regimen and were boosted again with a shorter rest period. Characterization of the Env-reactive mAb repertoire in these vaccine-recipients may provide some insight into whether the length of the rest period necessary for expansion of long HCDR3 mAbs.
In summary, study of the B cell repertoires of memory B cells induced by the RV305 trial vaccine-regimen has defined a set of CD4bs-reactive B cell clonal lineages that were initiated by the RV144 vaccine-regimen and expanded after late boosting with the ALVAC-HIV and AIDSVAX B/E immunogens. These antibodies derived from UCAs with some degree of tier 2 virus neutralization capability.
The RV305 clinical trial (NCT01435135) received approvals from Walter Reed Army Institute of Research, Thai Ministry of Public Health, Royal Thai Army Medical Department, Faculty of Tropical Medicine, Mahidol University, Chulalongkorn University Faculty of Medicine, and Siriraj Hospital. Written informed consent was obtained from all clinical trial participants.The Duke University Health System Institutional Review Board approved all human specimen handling.
The RV305 clinical trial (NCT01435135) was a randomized double blinded placebo-controlled boosting of 162 RV144 clinical trial participants (NCT00223080) that occurred in Thailand. The RV305 clinical trial was sponsored by the U.S. Army Office of the Surgeon General and conducted in collaboration with the U.S. Army Medical Research and Materiel Command and the Thailand Ministry of Public Health. The primary objective was to characterize the cellular and humoral immune response after boosting and to evaluate the safety and tolerability of late and repetitive boosting with the ALVAC-HIV (vCP1521) and AIDSVAX B/E immunogens. Six-eight years after the conclusion of RV144, RV305 volunteers were randomized into three groups and boosted two times with a six month interval with either AIDSVAX B/E + ALVAV-HIV (vCP1521), AIDSVAX B/E or ALVAC-HIV (vCP1521) or a placebo. After commencement no changes were made to the vaccine-regimen. All HIV-1 uninfected RV144 participants that had completed the full RV144 vaccine-regimen, were at low risk for HIV-1 infection based on self-reported behavioral habits, able to pass a Test of Understanding, gave written consent and were in general good health were eligible. Female volunteers had to be on adequate birth control 45 days prior to the first inject and consent to remaining on birth control. For safety reasons women that were pregnant, nursing or planning on becoming pregnant were excluded. Volunteers with a conflict of interest, psychological or medical conditions, or those unable to complete a Test of Understanding were excluded. Vaccine safety was measured by self-reporting on a diary card local and systemic reactions for three days post-vaccination. All adverse events and serious adverse events were recorded throughout the trial and up to three months post final boost.
Peripheral blood mononuclear cells (PBMCs) were stained with Aqua vital dye ((AqVd) Invitrogen), IgM-FITC, IgD-PE, CD3 -PECy5, CD14-BV605, CD16-BV570, CD235a-PECy5, CD27-PECy7, CD38-APC-AF700, CD19-APCCy7, along with AF647 and BV421 conjugated antigens. Viable antigen-specific B cells (AqVd-CD14-CD16-CD3-CD235a-CD19+IgD-CD38all, AF647 and BV421 double positive) were single-cell sorted with a BD FACSAria II- SORP (BD Biosciences, Mountain View, CA) into 96 well PCR plates and stored at -80°C.
Immunoglobulin variable heavy and light chain variable regions (VH and VL) were RT-PCR amplified using AmpliTaq360 Master Mix (Applied Biosystems) with conditions previously described [55]. PCR products were purified (Qiagen, Valencia, CA) and sequenced with a BigDye Sequencing kit (Applied Biosystems) on an ABI 3700 sequencer. VH and VL chain gene rearrangements, clonal relatedness, UCA and intermediate ancestor (IA) inferences were made using Cloanalyst [56].
PCR-amplifed sequences were transiently expressed as previously described [55]. Briefly, linear expression cassettes were constructed by placing the PCR-amplified VH and VL chain genes under the control of a CMV promoter along with an IgG constant region and poly A signal sequence. These linear expression cassettes were then co-transfected into 293T cells and after three days the cell culture supernatants were harvested and concentrated.
For large scale expression, the VHDHJH and VLJL genes were synthesized (VH chain in the IgG1 4A backbone) and transformed into DH5α cells (GeneScript, Piscataway, NJ). Plasmids were expressed in Luria Broth, purified (Qiagen, Valencia, CA) and ~ 5x106 293i cells were transfected with 1 mg of Ig (VH) and light (VL) chain genes using poly-ethylenimine (PEI) or with 0.4mgs of heavy- and light chain-gene using ExpiFectamine™ (Life Technologies, Carlsbad, CA) following the manufacturers protocol. After five days mAbs were concentrated, purified from the cell culture supernatant by an overnight incubation with Protein A beads and buffer exchanged into PBS.
High affinity 384-well microplates (Costar 3700) were coated overnight at 4°C with 30ng/well of protein in 0.1% Sodium Biocarbonate. For binding, a direct ELISA was performed in which monoclonal antibodies (mAbs) beginning at 100ug/mL were diluted 3-fold in blocking buffer and added to the plates for 1 hour. Antibody binding was detected using IgG-HRP (Rockland, Limerick, PA) diluted 1:10,000 in azide-free blocking buffer. For the blocking ELISA, mAbs of interest were diluted and added to the plate for one hour. Plates were washed and a biotinylated mAb was added for one hour. Blocking was evaluated by adding streptavidin-HRP. The direct binding and blocking ELISAs were developed using SureBlue Reserve TMB One Component microwell peroxidase substrate (catalog no. 53-00-03; KPL) and the reactions were stopped with 0.1% HCL. Plates were read on a plate reader (Molecular Devices) at 450 nm. Palivizumab (Synagis) (MedImmune, LLC; Gaithersburg, MD) was used as a negative control. The plasma was screened with the binding Ab multiplex assay (BAMA) as previously described. The antibody B12 was a gift from QBI and the Vaccine Research Program, Division of AIDS, NIAID contract # HSN272201100023C.
Neutralization assays were performed in both TZM-bl and A3R5 cell lines as previously described [23,32]. Data were reported as ID50 titers for plasma and IC50 titers for mAbs.
Purified mAbs were epitope mapped on B.YU2gp120 core proteins (ΔV1, V2, V3 loops) displayed on S. cerevisiae as previously described [35,36]. Briefly, mAbs that bound B.YU2gp120 core protein were assayed for binding to 31 different B.YU2gp120 core proteins with point-mutations and the wild type protein. Antigen-specific recognition was confirmed by the observation that mAbs did not show binding to non-displaying S. cerevisiae. Data was recorded as the percent binding to a mutant relative to the wild type core proteins. The B12 binding data in Figure 3 are from [36].
Surface plasmon resonance was performed on a BIAcore 4000 instrument. The purified recombinant mAb was immobilized to a CM5 sensor chip and envelope binding was measured in real time with continuous flow of PBS (150mM NaCL, 0.005% surfactant P20 [pH 7.4] at 10–30 μl/min. Data was analyzed with BIAevaluation 4.1 software (BIAcore).
ADCC mediated by the mAbs was assessed according to previously published procedures [57,58]. Briefly, HIV-1 reporter virus used was a replication-competent infectious molecular clone (IMC) designed to encode the HIV-1 env genes in cis within an isogenic backbone that also expresses the Renilla luciferase reporter gene and preserves all viral open reading frames [59]. CEM.NKRCCR5 cells (NIH AIDS Reagent Program, Division of AIDS, NIAID, NIH: CEM.NKR CCR5+ Cells from Dr. Alexandra Trkola [60] were infected with HIV-1 IMCs encoding the subtype AE CM235 (accession number AF259954), B WITO (accession number JN944948), and C Ce1086.c (accession number FJ444395) env genes within an NL4-3 backbone [59]. Whole PBMC from an HIV-seronegative donor with the heterozygous 158F/V genotype for Fc-gamma receptor IIIa were used as effector cells at an effector cell to target cell (E:T) ratio of 30:1. MAb A32 (James Robinson; Tulane University, New Orleans, LA), Palivizumab (MedImmune, LLC; Gaithersburg, MD; used as negative control) and vaccine induced mAbs were tested at a final concentration range of 10–0.039μg/ml using 4-fold serial dilutions. All the conditions were evaluated after 6 hour incubation at 37°C and 5%CO2. The ADCC activity was reported as % specific killing calculated as [(RLU in control well − RLU in test well)/ RLU of control well] ×100. The results were considered positive if ADCC activity was ≥15% specific killing. ADCC activities are reported either as the endpoint concentration (EC), defined as the mAb concentration that intersects the positive cutoff of 15% specific killing, or as positive area under the curve (pAUC), calculated by the trapezoidal rule using the values ≥15% specific killing.
To generate the autologous HIV-1 CH505 SOSIP.664 and clade B 92Br SOSIP.664 expression constructs we followed established SOSIP design parameters [61]. Briefly, the SOSIP.664 trimer was engineered with a disulfide linkage between gp120 and gp41 by introducing A501C and T605C mutations (HxB2 numbering system) that covalently links the two subunits of the heterodimer [61]. The I559P mutation was included in the heptad repeat region 1 (HR1) of gp41 for trimer stabilization, and a deletion of part of the hydrophobic membrane proximal external region (MPER), in this case residues 664–681 of the Env ectodomain [61]. The furin cleavage site between gp120 and gp41 (508REKR511) was altered to 506RRRRRR511 to enhance cleavage [61]. The resulting, codon-optimized CH505 SOSIP.664 env gene was obtained from GenScript (Piscataway, NJ) and cloned into pVRC-8400 using Nhe1 and NotI restriction sites and the tissue plasminogen activator signal sequence.
Fabs were expressed by transient transfection of HEK 293F suspension cells, using linear PEI following the manufacturer’s suggested protocol. After 5 d, supernatants were clarified by centrifugation and diluted twofold with 1x PBS buffer, and the protein isolated from the diluted spernatant using CaptureSelect LC-Kappa (Hu) affinity matrix (Thermo Fisher Scientific, Waltham, MA), according to manufacturer’s protocols. Fractions containing the protein of interest were pooled, concentrated, and further purified by gel filtration chromatography using a Superdex 200 analytical column (GE Healthcare Life Sciences, Pittsburgh, PA) in a buffer of 2.5mM Tris, pH 7.5, 350mM NaCl, and 0.02% sodium azide.
Each SOSIP.664 construct was transfected into 293F cells together with a plasmid encoding the cellular protease, furin, at a 4:1 Env:furin ratio. The cells were allowed to express the soluble trimer for 5–7 days. Culture supernatants were collected, cells removed by centrifugation at 3800 x g for 20 min, and the supernatant filtered with a 0.2 μm pore size filter. The soluble SOSIP was purified by flowing the filtered supernatant over a lectin (Galanthus nivalis) affinity chromatography column overnight at 4°C. The lectin column was washed with 1x PBS, followed with 1x PBS supplemented with 0.5 M NaCl, and proteins were eluted with 1 M methyl-α-D-mannopyranoside dissolved in 1x PBS. The eluate was concentrated and loaded for further purification onto a Superdex 200 10/300 GL column (GE Healthcare Life Sciences, Pittsburgh, PA) prequilibrated in a buffer of 5 mM Hepes, pH 7.5, 150 mM NaCl and 0.02% sodium azide for analysis by EM.
Purified SOSIP.664 trimer was incubated with a five molar excess of Fab at 4°C for 1 hour. A 3 μL aliquot containing ~0.01 mg/ml of the complex was applied for 30 s onto a carbon coated 400 Cu mesh grid that had been glow discharged at 20 mA for 30 s, followed by negative staining with 2% uranyl formate for 20 s. Samples were imaged using a FEI Tecnai T12 microscope operating at 120kV, at a magnification of 52,000x, resulting in a pixel size of 2.13 Å at the specimen plane. Images were acquired with a Gatan 2K CCD camera using a nominal defocus of 1500 nm at 10° tilt increments, up to 50°. The tilts provided additional particle orientations to improve the image reconstructions.
Particles were picked semi-automatically using EMAN2 [62] and put into a particle stack. Initial, reference-free, two-dimensional (2D) class averages were calculated and particles corresponding to complexes (with one, two, or three Fabs bound) were selected into a substack for determination of an initial model for the DH576: CH505 SOSIP.664 complex. The initial model was calculated in EMAN2, imposing 3-fold symmetry, and subsequent refinement in EMAN2 also imposed 3-fold symmetry. In total, 22,929 particles were included in the final reconstruction. The resolution of the final model was determined using a Fourier Shell Correlation (FSC) cut-off of 0.5.
The cryo-ET structure of b12-bound gp120 trimer (PDB ID: 3DNL) [63] and an Fab model were manually docked into the EM density and refined with the UCSF Chimera ‘Fit in map’ function [64]. The gp120 subunit of crystal structures with different Fabs were superposed on other gp120 cores from the PDB by least-squares fitting in Coot [65]
The DH576 Fab was crystallized at 10–15 mg/mL. Crystals were grown in 96-well format using hanging drop vapor diffusion and appeared after 24–48 h at 20°C. Fab crystals were obtained in the following conditions: 20% PEG 4000, 100mM Hepes, pH 7.0, 1M NaCl. Crystals were harvested and cryoprotected by the addition of 20–25% glycerol to the reservoir solution and then flash-cooled in liquid nitrogen.
Diffraction data were obtained at 100 K from beam line 24-ID-C at the Advanced Photon Source using a single wavelength. Datasets from individual crystals were processed with HKL2000[66]. Molecular replacement calculations for the free Fab were carried out with PHASER[67], using the variable domains of PGT135 [Protein Data Bank (PDB) ID 4JM2] and the constant domains of VRC01 from the VRC01/gp120 complex [Protein Data Bank (PDB) ID 4LSS] as the starting models for molecular replacement.
Refinement was carried out with PHENIX[68], and all model modifications were carried out with Coot[65]. During refinement, maps were generated from combinations of positional, group B-factor, and TLS (translation/libration/screw) refinement algorithms. Secondary-structure restraints were included at all stages for all Fabs. Structure validations were performed periodically during refinement using the MolProbity server[69]. The final refinement statistics are summarized in (S8 Table).
All statistical analysis was performed in SAS by the Duke Human Vaccine Institute statistical team. The statistical test and p value are recorded where used.
The EM reconstruction has been deposited in the Electron Microscopy Data Bank as EMD-8573. The crystal structure of DH576 has been deposited in the Protein Data Bank as PDB ID5UIX. The VH and VL chain genes described have been submitted to Genbank with accessioning numbers KY499910-KY499949.
|
10.1371/journal.pgen.1000589 | P-Type ATPase TAT-2 Negatively Regulates Monomethyl Branched-Chain Fatty Acid Mediated Function in Post-Embryonic Growth and Development in C. elegans | Monomethyl branched-chain fatty acids (mmBCFAs) are essential for Caenorhabditis elegans growth and development. To identify factors acting downstream of mmBCFAs for their function in growth regulation, we conducted a genetic screen for suppressors of the L1 arrest that occurs in animals depleted of the 17-carbon mmBCFA C17ISO. Three of the suppressor mutations defined an unexpected player, the P-type ATPase TAT-2, which belongs to the flippase family of proteins that are implicated in mediating phospholipid bilayer asymmetry. We provide evidence that TAT-2, but not other TAT genes, has a specific role in antagonizing the regulatory activity of mmBCFAs in intestinal cells. Interestingly, we found that mutations in tat-2 also suppress the lethality caused by inhibition of the first step in sphingolipid biosynthesis. We further showed that the fatty acid side-chains of glycosylceramides contain 20%–30% mmBCFAs and that this fraction is greatly diminished in the absence of mmBCFA biosynthesis. These results suggest a model in which a C17ISO-containing sphingolipid may mediate the regulatory functions of mmBCFAs and is negatively regulated by TAT-2 in intestinal cells. This work indicates a novel connection between a P-type ATPase and the critical regulatory function of a specific fatty acid.
| Fatty acids serve diverse functions in organisms, including roles at the cell membrane to coordinate cell signaling processes. Monomethyl branched-chain fatty acids (mmBCFAs) are a special type of fatty acid that is commonly present in animals. Because mmBCFAs are a small component of the total fatty acid pool, their functions have not been a major research focus and are largely unclear. We tackled the problem using the nematode C. elegans. Our laboratory previously found that without mmBCFAs, worms cannot develop normally and die. To understand how these obscure fatty acids perform such important roles, we searched for other factors involved in the process by conducting a mutagenesis screen to uncover mutant worms that can recover the ability to grow without the presence of mmBCFAs. We found several such mutations in a single gene that codes for a protein called TAT-2. TAT-2 is one of several poorly understood P-type ATPases that likely help maintain the proper lipid structure in cell membranes. Our work indicates that TAT-2 antagonizes the growth regulatory function of mmBCFAs in intestinal cells. Studies on how mmBCFAs and this protein functionally interact explore a novel, interesting, and important problem that is only beginning to be understood.
| Lipids play many critical roles in cellular function ranging from providing structural support within cell membranes to mediating signaling events. The importance of the particular fatty acid constituents of complex lipids is only beginning to be understood. While extensive analyses have been conducted to elucidate the specific roles of straight-chain saturated and unsaturated fatty acids, there is little known about the roles of monomethyl branched-chain fatty acids (mmBCFAs) in animals. We have previously identified an essential role for mmBCFAs in regulating post-embryonic development in the model organism C. elegans, where the deletion of elo-5, a gene encoding a very long chain fatty acid (VLCFA) elongase, causes larval arrest and death. When the fatty acid composition of worms lacking functional ELO-5 was compared to that of wild-type worms, two detectable mmBCFAs (C15ISO and C17ISO) were missing. In addition, when the ELO-5 deficient strain was supplemented with exogenous C15ISO and C17ISO, the nematodes recovered from larval arrest and were able to proliferate similar to wild-type worms [1]. The L1 arrest that occurs in the absence of C17ISO is strikingly similar to the L1 diapause that occurs in the absence of food [2]. In both cases, the worms arrest at an early L1 stage, just after hatching and prior to the first M cell division, and the arrest is reversible upon restoration of the depleted component. The DAF-2/DAF-16 insulin-signaling pathway has been shown to be involved in the starvation-induced arrest [3]. Further characterization revealed that C17ISO may play a critical role in activating post-embryonic development in C. elegans through a novel pathway that is independent of the DAF-2/DAF-16 food-sensing pathway [2]. CKI-1 is a cyclin-dependent kinase inhibitor that has been shown to be required for exit from the cell cycle and is likely a critical downstream target of DAF-16 with its expression level normally decreasing during L1 development [3],[4]. The L1 arrested larvae seen in the absence of C17ISO exhibit stable expression of CKI-1 [2], suggesting that C17ISO deprivation may contribute to the upregulation of CKI-1 through a previously uncharacterized, DAF-16-independent signaling pathway. Consistent with the critical regulatory role played by mmBCFAs in worms, we have found that mmBCFA homeostasis is maintained through a feedback regulatory mechanism involving SBP-1 [1],[2]. SBP-1, the C. elegans orthologue of mammalian SREBP-1c [5], may sense a deficiency in mmBCFA levels and respond by upregulating the transcription of monomethyl branched-chain fatty acid biosynthesis enzymes including ELO-5 and ACS-1, a long chain fatty acid acyl-CoA ligase.
Cell membranes are generally asymmetric in nature, containing higher levels of aminophospholipids on the cytosolic leaflet and exhibiting an enrichment of sphingolipids and choline-containing lipids on the extracellular/lumenal leaflet [6]. Alterations in cell surface properties due to the loss of membrane asymmetry are associated with various normal and pathological outcomes including apoptosis and platelet activation [7],[8]. However, with the exception of phosphatidylserine externalization during apoptosis, the precise mechanisms by which the asymmetric distribution of lipids across cellular membranes regulates membrane and lipid functions are not well understood. A subfamily of P-type ATPases, called aminophospholipid translocases or flippases, are proteins that are thought to contribute to this asymmetry in an ATP-dependent manner [9],[10]. The postulated function of flippases is to aid in the inward movement of phosphatidylserine (PS) and phosphatidylethanolamine (PE) from the extracellular or lumenal leaflet to the cytosolic leaflet of cellular membranes. To date, there have been very limited functional studies on this family of proteins in animals. In yeast, an aminophospholipid translocase, DRS2, has been shown to mediate the flipping of a fluorescent analog of PS in isolated late Golgi membranes [11]. DRS2 mutants have been shown to have defects in protein trafficking at the trans Golgi network [12],[13], suggesting that membrane asymmetry may be important for intracellular transport events. In humans, mutations in FIC1, a putative aminophospholipid translocase, have been found in familial hereditary cholestasis [14]. However, the mechanism by which this mutation contributes to the pathology has yet to be elucidated. The C. elegans genome contains six predicted aminophospholipid translocases known as tat-1 through tat-6 [15],[16]. Recent work based on the expression pattern of reporter transgenes suggests that each of the TAT proteins may have distinct functions and four of them are not essential under regular growth conditions [17]. Mutations in tat-2, tat-3 and tat-4 genes were shown to enhance the sensitivity of C. elegans to cholesterol deficiency [17]. TAT-1 has been shown to be involved in the movement of phosphatidyl serine during apoptosis and regulate lysosome biogenesis and endocytosis [16],[18],[19]. However, the cellular and physiological functions of TAT proteins largely remain to be explored in animal cells.
Sphingolipids are structurally diverse lipids that can have equally diverse functions. In addition to altering membrane content, sphingolipid metabolites have also been implicated in intracellular and extracellular signaling (reviewed in [20]). The rate-limiting step in sphingolipid biosynthesis is the condensation of serine with fatty acid-CoA to generate 3-ketosphinganine, a precursor to the sphingoid base. This step is catalyzed by serine palmitoyltransferase (SPT) and as the name implies, the fatty acyl-CoA substrate is generally palmitoyl-CoA [21],[22]. In C. elegans, sphingoid bases of glucosylceramides and sphingomyelin have been shown to exclusively contain a branched-chain base constituent, perhaps produced from an iso-branched fatty acid [23]. This may suggest a potential link between the essential function of mmBCFAs and sphingolipids.
In this paper, we describe the identification and characterization of mutations in the tat-2 P-type ATPase gene as suppressors of the L1 arrest caused by depleting the mmBCFA, C17ISO. We provide evidence that TAT-2 acts in intestinal cells to specifically antagonize mmBCFA activity in regulating post-embryonic growth, and such a function may be mediated by an mmBCFA-containing sphingolipid. This work provides a novel link between the critical regulatory function of a specific fatty acid and the maintenance of lipid bilayer asymmetry.
The elo-5(gk208) deletion mutant animals die in various larval stages and consequently do not reproduce. When the elo-5(gk208) deletion worms are supplemented with C15ISO or C17ISO, they grow to wild-type adults and can continue to grow and propagate in the presence of this supplementation [1]. However, upon supplementation with a short-chain mmBCFA, C13ISO, elo-5(gk208) worms grow to wild-type adults in the first generation and produce F1 progeny (Figure 1A). The progeny uniformly arrest in the first larval stage for several days before dying. To identify genes that may potentially act downstream of C17ISO to mediate this regulatory function, we screened for suppressor mutations that would allow C13ISO-supplemented elo-5(gk208) animals to grow past the L1 stage in the F1 generation (Materials and Methods). Among several suppressor mutations, three alleles were able to grow indefinitely in the presence of C13ISO supplementation (Figure 1B). Without any supplementation, these worms grew similarly to wild type in the first generation, but their progeny arrested and died as early larvae (Table 1), indicating that, although the suppressor mutations were able to overcome the L1 arrest in the absence of the longer chain mmBCFA, C17ISO, they were not able to bypass all mmBCFA functions.
Molecular cloning indicated that these three mutations are alleles of a single gene, tat-2 (see below). We were able to conclude that these three mutations are likely loss of function mutations in the tat-2 gene and that loss of tat-2 gene function is responsible for the suppressor phenotype based on results from the following three sets of experiments. Firstly, we determined that tat-2 RNAi phenocopies the mutations. When tat-2(RNAi), but no mmBCFA supplementation, was applied, elo-5(gk208) deletion mutants grew to healthy adults in the first generation and produced progeny that were able to develop into egg-laying adults, with the next generation arresting in early larval stages (Table 1). Secondly, we acquired a 222-bp deletion mutation in tat-2, tm2332, from the National Bioresource Project in Japan, and showed that it can also suppress the L1 arrest phenotype of elo-5(RNAi). When treated with elo-5(RNAi) in the absence of mmBCFA supplementation, tat-2(tm2332) mutant worms grew to wild-type adults in the first generation and continued to grow and propagate normally when transferred to fresh elo-5 RNAi plates for multiple generations. The stronger suppressing effect seen in this assay likely reflects that fact that elo-5(RNAi) has a weaker mutant phenotype as that of elo-5(gk208) [2]. A similar effect was seen with another tat-2 deletion mutation, tm1773 (data not shown). Finally, we were able to rescue the tat-2(ku449) mutant phenotype (eliminating the suppressive role of ku449) by expressing the TAT-2 protein from a transgene under the control of the putative TAT-2 promoter as described below (Table 2). Therefore, combined with the finding that TAT-2 is not involved in the biosynthesis of mmBCFAs (see below), our results indicate that TAT-2 plays a negative role in C17ISO-mediated L1 growth regulation.
To identify the gene defined by the suppressor mutations, we mapped the mutations to the middle of chromosome IV using genetic and single nucleotide polymorphism markers (Materials and Methods; Figure 2A and 2B). Sequence analysis identified all three alleles as distinct molecular lesions in the tat-2 gene: ku452, ku449, and ku450 contained amino acid changes from Thr 412 to Asp, Ser 510 to Leu, and Gly 617 to Glu, respectively (Figure 2C). Upon examination of the tat-2 cDNA, we found that it coded for an additional 92 amino acids not included in the protein predicted by Wormbase (data not shown).
The protein encoded by tat-2 is a 1313-amino acid P-type ATPase that belongs to a family of proteins also known as flippases. Proteins with structural similarity to TAT-2 are found in yeast, mammals and C. elegans (Figure 2C). The protein encoded by tat-2 shows conservation of domains that are characteristic of P-type ATPases as well as features that are specific to the Drs2p-subfamily of ATPases. Drs2p and other P-type ATPases are 10-pass transmembrane proteins with a cytoplasmic domain containing conserved residues that form an aspartyl-phosphate intermediate during ATP hydrolysis [24]. There are also two cytoplasmic sequences specifically conserved in the Drs2p family of ATPases [25]. The residue change in tat-2(ku452) (threonine 412 to asparagine) is two residues away from the catalytic aspartate and thus, may disrupt the catalytic function of the TAT-2 protein. This mutation may be of particular interest because an analogous mutation has been identified in patients diagnosed with familial cholestasis, in which this conserved threonine is changed to a methionine [26]. The tat-2(ku449) lesion (S510T) is in a cytoplasmic domain that is conserved in Drs2p-like ATPases, and the tat-2(ku450) lesion (G617E) affects a region that is not obviously conserved (Figure 2).
To determine that the suppressor effect is not due to a recovery of C17ISO biosynthesis, we examined the fatty acid composition of these mutants by gas chromatography (GC). We found that the suppressor mutants grown in the presence of C13ISO did not recover the ability to synthesize mmBCFAs (Figure 1C). The GC analysis also detected statistically significant differences in the levels of two other fatty acids, stearic acid (C18:0; P = .016) and methyleneoctadecanoic acid (C19Delta; P = .0015) between wild-type, elo-5(gk208), elo-5(gk208); tat-2(tm2332) and tat-2(tm2332) animals. Because C19Delta was significantly increased in the tat-2(tm2332); elo-5(gk208) mutants compared to elo-5(gk208) worms, we supplemented this fatty acid to the food given to elo-5(gk208) worms in the presence of C13ISO, but no rescue of the growth arrest of elo-5(gk208) mutants was observed (data not shown). Therefore, while these tat-2 mutations allow mmBCFA-deficient animals to by-pass the L1 arrest, they do not do so by recovering the ability to synthesize mmBCFAs, nor by increasing the levels of other fatty acids.
Upon detailed examination of the tat-2(tm2332) single deletion mutant, we saw no obvious abnormalities, including normal fatty acid composition (data not shown; Supplemental Data; Figure 1). When we constructed the complete double deletion mutant, tat-2(tm2332);elo-5(gk208), we found that the animals behaved similarly to the point mutants in that they can grow indefinitely in the presence of C13ISO. However, a small proportion of worms (<20%) grown without supplementation progressed an additional generation to F2 prior to the arrest, instead of arresting in the F1 generation as seen with the point mutants (Table 1). These data support the idea that the absence of tat-2 can suppress some, but not all defects caused by mmBCFA deficiency. It also suggests that in the absence of tat-2, C13ISO (at a low concentration) can substitute for C17ISO in some post-embryonic developmental aspects. One possibility is that, in tat-2(-) animals, the absorption of C13ISO may be increased to achieve a higher cellular concentration of C13ISO. We tested this by feeding high amounts of supplemental C13ISO to tat-2(-) and N2 worms. By comparing the fatty acid composition of these worms, we determined that there was not a significant difference in absorption of C13ISO (data not shown).
To determine where tat-2 is expressed in C. elegans, we generated a transcriptional GFP reporter that included approximately 3 kb of sequence upstream of the tat-2 initiation codon. Consistent with the report by Lyssenko et al. [17], we also observed that tat-2 is expressed primarily in the excretory cell, spermatheca and intestine (Figure S1 and Text S1). ELO-5 is also expressed in the intestine and several amphid neurons in the head [1]. Since fatty acids generated in the intestine are likely transported to various tissues of the body, it is possible that rescue of the elo-5 deletion mutant phenotype depends on TAT-2-mediated effects on mmBCFAs in non-intestinal tissues. In order to address this, we first verified that the tat-2 cDNA expressed under the control of the 3-kb promoter described above produces a functional protein by generating transgenic lines harboring a tat-2Prom::tat-2 cDNA in the tat-2(ku449);elo-5(gk208) genetic background. The transgenic worms were assessed for the elo-5(lf) phenotype (second-generation larval arrest) in the presence of C13ISO and compared to transgenic tat-2(ku449);elo-5(gk208) worms expressing GFP under the control of the tat-2 promoter. As shown in Table 2, 99.1% of the transgenic F1 worms expressing tat-2 cDNA arrested at the L1 stage, whereas 0% of the negative control transgenic F1s arrested as L1s. This indicated that the tat-2 cDNA expressed under the control of the 3-kb tat-2 promoter is sufficient to rescue TAT-2 function. Next, we generated several constructs for the tissue-specific expression of the tat-2 cDNA, and each of these constructs was injected into tat-2(ku449); elo-5(gk208). The transgenic lines were then examined for second-generation larval arrest in the presence of C13ISO and the results are summarized in Table 2. Expression of tat-2 cDNA from the intestinal-specific ges-1 promoter [27] was able to fully rescue the elo-5 suppression phenotype of tat-2(ku449);elo-5(gk208), however, tat-2 driven by the spermathecal promoter, sth-1 [28], the excretory cell promoter, exc-5 [29], or the hypodermal-specific promoter col-10 [30], was not able to restore the elo-5 phenotype (Table 2). These results demonstrate that loss of functional TAT-2 in the intestine is necessary for suppression of elo-5(gk208). Therefore, TAT-2 acts in the intestine for its function in C17ISO-defiency-triggered growth arrest.
To test if the rescue of mmBCFA deficiency was a shared characteristic of depletion of other aminophospholipid translocases in C. elegans, we acquired deletion mutations for tat-1(tm1034), tat-3(tm1275), tat-4(tm1801) and tat-6(ok1984) and checked for rescue of the elo-5(lf) phenotype. Due to the poor viability of the tat-5 deletion mutant and the penetrant embronyic lethal phenotype displayed by rrf-3(-) animals on tat-5 RNAi, we treated elo-5(gk208);rrf-3(ok629) animals with tat-5 RNAi and assayed rescue. None of these tat mutants rescued the elo-5(lf) phenotype (Table 3). In addition, we noted that tat-4(tm1801) and tat-6(ok1984) seemed to have a more severe phenotype in the presence of elo-5(lf), resulting in death of the tat-4 and tat-6 mutant animals in the first generation (Table 3). While TAT-2 is the only TAT protein that plays a prominent negative role in mmBCFA-mediated growth regulatory functions, it is possible that TAT-4 and TAT-6 play some positive roles in mmBCFA-involved functions.
We then wanted to test if tat-2(lf) can also affect the growth defects caused by general fatty acid depletion. We grew tat-2(tm2332) and wild-type control animals on bacteria expressing RNAi against either fasn-1, the C. elegans fatty acid synthase orthologue or fat-2, a Δ12 fatty acid desaturase involved in the normal formation of polyunsaturated fatty acids [31],[32]. tat-2(tm2332) exhibited no rescue of the lethality caused by fasn-1(RNAi) or fat-2(RNAi) (data not shown).
As mentioned earlier, sphingolipids in C. elegans have been shown to contain monomethyl branched-chain sphingoid bases [23]. We reasoned that tat-2(lf) may suppress the mmBCFA growth defect through suppression of the activity of more complex lipids that are formed from mmBCFAs, such as sphingolipids. We explored the idea that TAT-2 may act through its postulated flippase activity to modulate the activity of one or more sphingolipids localized in the membrane.
We examined the effects of depleting the C. elegans homolog (sptl-1) of one of the subunits (SPTLC1) of the mammalian serine palmitoyl-CoA transferase complex. We found that sptl-1(RNAi) results in 95.8% lethality in 10 days (Figure 3G). These animals also displayed an uncoordinated phenotype resulting from the dead larva within the mother and vacuolization and degradation of the intestine (Figure 3A and 3C). These results suggest that sphingolipids are essential for C. elegans survival and development. Surprisingly, when tat-2(tm2332) animals were grown in the sptl-1(RNAi) background, 96.3% were alive on day 10 (Figure 3G). The animals were egg-laying competent and most embryos developed normally (Figure 3E). A few embryos (>10 per worm) completed embryonic development inside the eggshell within the mother, but did not hatch out of the eggshell (data not shown). tat-2(-);sptl-1(RNAi) laid approximately 45% of the number of eggs compared to control tat-2(-) worms on the mock RNAi plates, and 96% of the resulting larvae were alive on day 10 (Figure 3G). The F1 larvae remain alive in various larval stages as long as 2 months but did not continue to propagate. In addition, no vacuoles were observed in tat-2(tm2332); sptl-1(RNAi) worms (Figure 3F). This suppression is not likely due to a reduced efficiency of the RNAi treatment in the tat-2 mutant, since no obvious reduction in RNAi efficiency was observed when the tat-2 mutant was treated by RNAi of several other metabolic genes such as fasn-1, spb-1 and acs-1 (data not shown). These results demonstrate that tat-2 deficiency not only rescues the growth defects caused by depletion of C17ISO, but can also partially rescue defects caused by depletion of sphingolipids.
Since cholesterol and sphingolipids reside together in lipid rafts, we tested tat-2 mutants for the ability to suppress the cholesterol-depleted phenotype. We found no rescue of the growth arrest phenotype caused by cholesterol-depletion in tat-2(tm2332) worms as compared with wild-type worms (data not shown; [33]). We also sought to determine if tat-2(tm2332) would be able to rescue general lipid misregulation phenotypes. When we knocked down expression of the C. elegans homolog of SREBP gene, C. elegans sbp-1, which is thought to affect global lipid homeostasis [5], tat-2(tm2332) mutants showed no differences as compared to control worms under the same conditions (data not shown). Finally, we tested the ability of tat-2(-) to alter the phenotype of a deletion mutant in a C. elegans orthologue of phosphatidylserine synthase, pssy-2(tm1955), an enzyme that converts phosphatidylcholine to phosphatidylserine. We found that the depletion of TAT-2 did not change the sterility phenotype of this mutant (data not shown).
The rescue of both sphingolipid and C17ISO depletion by tat-2 mutants led us to ask whether there were sphingolipids in elo-5-depleted worms and if so, what the fatty acid constituents of those sphingolipids were. Due to the difficulty of analyzing every type of sphingolipid in C. elegans in various mutant backgrounds, we chose to isolate glycosphingolipids in synchronized populations of mutant versus wild-type worms and separate them using thin layer chromatography (TLC) in the manner described in Griffitts et al., [34]. The bands representing ceramide-based sphingolipids with one to four saccharide molecules (as inferred by reference standards) were scraped from the TLC plates and the fatty acid side chains were extracted and analyzed by gas chromatography. The results are displayed in Figure 4. Analysis of wild-type control worms indicated that C15ISO and C17ISO represent ∼20–30% of the fatty acid chain composition in the glycosphingolipids obtained in this assay. In the elo-5-depleted worms, the fatty acid composition in the TLC scrape revealed the combined C15ISO and C17ISO proportions were down to ∼5%. The elo-5(lf) and wild-type worms supplemented with C13ISO showed similar results. In addition, we noted that the tat-2(lf);elo-5(lf) animals supplemented with C13ISO had similar results to the elo-5(lf) animals (Figure 4). This likely rules out the possibility that tat-2(-) is somehow rescuing the growth defect by altering the incorporation of mmBCFAs into the fatty acid side-chains of sphingolipids. When the relative amount of glycosphingolipids in the wild-type worms was compared with that in the elo-5(RNAi) worms, they were found to be similar (data not shown). Interestingly, elo-5(RNAi) worms incorporate other fatty acids into glycosphingolipids in order to compensate for the missing mmBCFAs, with a trend towards a significant increase in stearic acid (C18:0). We postulate that the substituted glycosphingolipids are unable to carry out the essential functions of mmBCFA-containing sphingolipids required for post-embryonic growth.
Little is known about the functions of most of the P-type ATPases in animals. In this report, we identified a novel physiological role for a P-type ATPase in a specific lipid-mediated developmental function. In our previous studies, we showed that depletion of long monomethyl branched-chain fatty acids by mutating the elo-5 gene causes C. elegans to robustly arrest their post-embryonic growth. We show here that loss-of-function mutations in the putative aminophospholipid translocase TAT-2 are able to overcome this arrest, indicating a link between TAT-2 function at cell membranes and C17ISO-mediated growth regulation. A straightforward interpretation of the genetic data is that TAT-2 antagonizes C17ISO function in promoting post-embryonic growth (Figure 5). Given what is known about this family of P-type ATPases, we speculate that TAT-2 may regulate the function of a lipid of more complex structure that mediates the C17ISO functions. TAT-2 could function to direct the localization of this lipid through its potential lipid translocase activity or through its potential role in secretory and/or endocytic trafficking.
Our work presented in this paper also suggests that such a lipid affected by TAT-2 function may be a sphingolipid. We have found that tat-2 mutations also partially suppress the developmental defects caused by RNAi of a subunit of the serine-palmitoyl transferase complex that is required for sphingolipid biosynthesis. We further determined that ceramide-based glycosphingolipids contain a significant proportion of C15ISO and C17ISO in their fatty acid side chains and that these are depleted in an elo-5(lf) background. Although it is currently corollary evidence, these results strongly implicate a functional linkage between mmBCFAs and sphingolipids.
In a recent report, mutations in tat-2, tat-3 and tat-4 genes were shown to enhance the sensitivity of C. elgans to cholesterol deficiency, implicating the roles of these P-type ATPases in cholesterol metabolism [17]. These roles appear to be distinct from the tat-2 function in C17ISO-mediated growth regulation for two reasons. First, regarding the effects on the animal growth, the enhancement of the defects caused by cholesterol deficiency is the opposite to the rescue of the defects caused by C17ISO depletion or by disrupting sphingolipid biosynthesis. Second, the tat-2 mutant effect on C17ISO function is highly specific to this particular P-type ATPase; none of the other tat genes has any observable effects (Table 3).
Disruption of mmBCFA biosynthesis causes multiple developmental defects that include, but are not limited to L1 growth arrest [1]. There is a clear functional difference between the shorter mmBCFA C13ISO and the long chain mmBCFA C17ISO [2]. Low levels of exogenous C13ISO (1 mM) can overcome almost all developmental defects caused by depleting endogenous mmBCFAs in elo-5 mutants except the L1 growth arrest that can be overcome by feeding a low level of dietary C17ISO or a high level of C13ISO (10 mM). Our data in this study indicate that tat-2 mutations can suppress the L1 arrest phenotype but not bypass the requirement of mmBCFA for some other developmental functions. In our typical L1 arrest assay, the elo-5(lf) mutants were grown on plates supplemented with a low level of C13ISO. These results indicate that C17ISO, not C13ISO, plays the major role in mediating development past the first larval stage in normal worms, and that C13ISO, at a very high concentration, can function as C17ISO for the growth regulatory function. The latter raises the possibility that the tat-2 mutations may suppress the growth arrest by dramatically changing the dynamics or efficiency of C13ISO. For example, tat-2 mutant worms may be able to absorb a higher level of exogenously provided mmBCFAs, leading to a higher internal C13ISO concentration. Alternatively, in the absence of TAT-2, the subcellular distribution of C13ISO may be altered through a novel mechanism in such a way that allows the molecule to function as effectively as C17ISO.
Several lines of evidence argue that the above scenarios are unlikely. First, under conditions without any mmBCFA supplement, elo-5(-);tat-2(-) mutants can typically bypass the L1 arrest and grow another generation before the animals arrest at various larval stages due to other developmental defects (Table 1). In addition, when elo-5(RNAi) is applied to wild-type worms, L1 arrest is observed for the F1 animals. However, when elo-5(RNAi) is applied to tat-2(-) mutants without any mmBCFA supplementation, the animals can actually propagate multiple generations without developmental arrest (Table 1). These results suggest that the tat-2 mutations are able to overcome the L1 arrest from depleting C17ISO without the presence of C13ISO. Second, we found that internal C13ISO levels were not significantly different between tat-2 and wild-type worms, arguing against the possibility that a tat-2 mutation enhances the absorption of C13ISO into animal cells. Third, the C15ISO and C17ISO levels are similar in the elo-5;tat-2 double mutants when compared to the elo-5 single mutant, inconsistent with the idea that a tat-2 mutation can increase the stability of mmBCFAs. Finally, if the C17ISO function in post-embryonic growth is mediated by a glycosphingolipid as we proposed above, the fact that tat-2 mutations can also suppress the defects in sptl-1(RNAi) worms does not appear to be consistent with the effect of the tat-2 mutations altering the subcellular distribution and effectiveness of C13ISO.
Our previous work indicated that the mmBCFA C17ISO can move from the intestinal cells of adult worms into their eggs [2]. Therefore, mmBCFA produced in a given tissue can potentially move to other tissues, making it difficult to determine in which tissue the mmBCFA is promoting post-embryonic growth. In this study, by expressing the TAT-2 protein behind tissue-specific promoters, we were able to show that expression of TAT-2 in intestinal cells is sufficient and likely necessary for TAT-2 to execute its role in mmBCFA-mediated regulatory function (Table 2). The latter statement is supported by the data that the expression of TAT-2 in three other tissues, including the major hypodermal tissue, could not rescue the mutant phenotype. Consistent with this, the mmBCFA biosynthesis enzymes ELO-5, ELO-6 and ACS-1 are all expressed at high levels in intestinal cells. These results suggest that the post-embryonic growth regulatory function of mmBCFAs and TAT-2 primarily occurs in intestinal cells, where exogenic mmBCFAs would have been first absorbed during feeding. However, because depleting C17ISO causes arrest of all post-embryonic tissues including muscle cells and hypodermal cells [2], a growth inhibitory signal such as an increase in cki-1 expression that may be triggered by C17ISO deficiency may be able to spread from intestinal cells to neighboring cells. The nature of this high-order lipid molecule and how this molecule executes this signaling process are currently not known. Additional genetic screening and biochemical analyses are being carried out to search for the answers.
All strains were maintained at 20 °C on OP50 bacteria on nematode growth media (NGM) according to standard protocol [35] unless RNAi treatment was performed. The wild-type strain was Bristol strain, N2. Mutant strains used were elo-5(gk208)IV, tat-2(ku449)IV, tat-2(ku450)IV, tat-2(ku452)IV, tat-2(tm2332)IV, tat-2(1773)IV, tat-1(tm1034)III, tat-3(tm1275)III, tat-4(tm1801)II, tat-6(ok1984)V, rrf-3(ok629)II, dpy-13(e184)IV and unc-5(e53)IV. The elo-5(gk208)IV, tat-6(ok1984)V, rrf-3(ok629)II, dpy-13(e184)IV and unc-5(e53)IV strains were obtained from the C. elegans Gene Knockout Consortium. The tat-2(tm2332)IV, tat-2(1773)IV, tat-1(tm1034)III, tat-3(tm1275)III, tat-4(tm1801)II, pssy-2(tm1955)III deletion strains were obtained from the National Bioresource Project in Japan. The CB4856 Hawaiian strain was used in SNP mapping procedures.
The elo-5(gk208)IV strain and suppressor and double mutant strains with the elo-5(gk208)IV allele were maintained on OP50 bacteria supplemented with the mmBCFA-producing S. maltophilia bacteria [1].
elo-5(gk208) L4-staged P0 worms grown in the presence of 2 mM C17ISO were mutagenized with EMS. Mutagenized F1 worms were grown to gravid adulthood on 2 mM C17ISO-supplemented plates and F2 eggs released by alkaline hypochlorite treatment were plated on 1 mM C13ISO-supplemented plates. Suppressor candidates were determined by the ability of their progeny to grow past the L1-stage. From ∼8000 haploid genomes, five suppressor mutants were isolated. Three tat-2 alleles, ku449, ku450, ku452, were isolated in separate rounds of screening.
For all survival and rescue experiments, strains were maintained on plates containing S. maltophilia to provide an exogenous source of mmBCFAs. In order to score for elo-5 growth arrest phenotypes, worms were collected from S. maltophilia containing plates and subjected to alkaline hypchlorite treatment to eliminate contaminating S. maltophilia and eggs were collected. Clean eggs were then spotted onto AMP plates spotted with HT115 transformed with empty vector (pPD 129.36) without mmBCFA supplements or AMP plates spotted with HT115 transformed with empty vector to which indicated concentrations of C13ISO were added from a 10 mM stock in DMSO. Following several days of growth, worms were scored for their developmental stage. HT115, when transformed with pPD 129.36, was used in these experiments because of its resistance to ampicilin, which aids in preventing the growth of contaminating bacteria that can interfere with accurate rescue scoring. Phenotypes on standard OP50 were identical.
In the sptl-1(RNAi) experiments, gravid wild-type or tat-2(tm2332) adults were bleached and eggs were spotted on bacteria expressing sptl-1(RNAi). After 2 days of development, 20 single worms were cloned to individual sptl-1(RNAi) plates and survival of the P0 and F1 generation was quantified as a percent of the total number of animals scored. This was repeated 3 times, in parallel for wild-type and tat-2(tm2332) mutant worms. An unpaired Student's t-test was used to compare the values.
Lipid extraction, fatty acid methyl ester preparation and gas chromatography were performed in the manner described [1],[2],[36]. Worms grown from eggs isolated by bleaching were spotted on plates with: 1) HT115 (transformed with pPD129.36), 2) HT115 supplemented with C13ISO where final concentration of the fatty acid in the bacteria culture was equal to 1 mM, or 3) HT115 transformed with pPD129.36 vector expressing double-stranded RNA against the indicated genes. Adults and their L1 progeny were collected with water, washed and frozen until use for lipid extraction. For each type of sample, five replicates were assayed.
For analysis of the fatty acid chains in glycosphingolipids, glycosphingolipids were first isolated from worm extracts and separated by TLC as described below. Bands representing glycosphingolipids were scraped from the silica TLC plates. Extraction, methyl ester preparation and gas chromatography of fatty acids from TLC-fractionated lipids were performed as outlined above. In these experiments, three replicates were done for each sample type, except where noted. For both analyses, average, standard deviation and standard error were calculated using the Excel program. Statistical significance was determined using the single-factor ANOVA method.
A SNP mapping method [37] was used to map tat-2(ku452) near the middle of chromosome IV. Due to the proximity of tat-2(ku452) to the elo-5(gk208) mutation, which was necessary to have in the background to observe the suppressor phenotype, the mapping strategy was switched to traditional three-point mapping, narrowing the region containing ku452 to between the markers dpy-13 and unc-5. Recombinants were generated by crossing ku452 to a dpy-13(e184); elo-5(gk208); unc-5(e53) strain. F2 Unc non-Dpy and Dpy non-Unc animals were isolated, homozygosed and scored for the presence of ku452 suppressor phenotype. Of these recombinants, 11/23 Unc non-Dpy and 45/63 Dpy non-Unc recombinants retained the ku452 mutation. Pools of four fosmids (15 ng/µL) along with sur-5::GFP (90 ng/µL) [38] were co-injected into ku452 worms and examined for rescue. Due to the incomplete fosmid coverage of the tat-2 gene, no fosmid pools rescued. Therefore, pJB18 (see below) was injected and rescue was observed. The coding region of tat-2 was PCR-amplified from ku449, ku450 and ku452 mutants and sequenced to identify the molecular legions.
HT115 bacteria expressing double-stranded RNA from the indicated genes was seeded onto NGM agar plates with IPTG and ampicilin for feeding [39]. Bacterial strains were obtained from the C. elegans genome-wide RNAi feeding library (Geneservice). Eggs isolated from the indicated C elegans strains by hypochlorite treatment were allowed to develop at 20°C with the RNAi feeding strains as their food source. Controls were fed with HT115 transformed with an empty pPD129.36 plasmid.
The tat-2 transcriptional reporter construct, pJB10, was generated by PCR amplification of 3 kb upstream of the tat-2 initiation codon from N2 genomic DNA using the following primers: For 5′-AAGGATCCTTTCCATGACTCACGCTG-3′, Rev 5′-AACCCGGGCCCTCCGCCACCTCCTTT-3′. The PCR product was ethanol precipitated, digested with BamHI and SmaI and ligated into the GFP vector pPD95.69 (generously provided by the Fire lab). The resulting construct, pJB10, was co-injected with pBluescript into the following strains: N2, elo-5(gk208), elo-5(gk208);ku452, and tat-2(tm2332) to obtain transgenic lines. Multiple lines were generated in each strain to verify the expression. A construct expressing tat-2 cDNA under the native 3 kb tat-2 promoter was generated by first cloning the promoter into pPD49.26 using the same primers and procedure described above for pJB10. The resulting construct, pJB14, was subsequently used to subclone the tat-2 cDNA. A first strand cDNA synthesis reaction was carried out using a gene-specific tat-2 primer and RNA isolated from N2 worms as template. The tat-2 cDNA was then PCR amplified from the first strand reaction using the following primers: For 5′-CACACCCGGGATGTTCAGTTGGTTGCCATG-3′, Rev 5′-AAACCCGGGTTAAAGGCGAGTGATTACGTCTG-3′. The PCR product was ligated into pJB14 as a SmaI fragment to yield pJB18. To generate tissue-specific constructs tat-2 cDNA was amplified from the first strand reaction described above using primers: For 5′-CACAGGATCCATGTTCAGTTGGTTGCCATG-3′, Rev 5′-AAACCCGGGTTAAAGGCGAGTGATTACGTCTG-3′. The PCR product was ligated into pPD49.26 as a BamHI/SmaI fragment to generate pJB15. Tissue-specific promoters were PCR amplified as follows: 1.8 kb upstream of the ges-1 start codon, 2.2 kb upstream of the exc-5 start codon and 3.4 kb upstream of the sth-1 start codon. Each PCR product was purified, digested and subcloned separately into pJB15 upstream of the tat-2 cDNA start codon. For the heat shock construct, pES11, the tat-2 cDNA generated above was cloned into the heat shock promoter vector, pPD49.83 as a BamHI/SmaI fragment. pES11 was co-injected with pBluescript and with a sur5::GFP marker (Yochem et al. 1998) construct into N2 worms.
Analysis of GFP expression and phenotypic abnormalities were conducted with Nomarski optics using a Zeiss Axioplan2 microscope and a C4742-95 CCD camera. Plate phenotypes were observed through a Leica MZ16F dissecting microscope and pictures were taken with a Canon Powershot A620 digital camera with a Canon Scopetronix Maxview Plus adapter.
Lower phase lipids were isolated and resolved via TLC in the manner described [34] from the indicated strains. In short, a 500–800 µL pellet of packed worms was used as starting material. The lower phase extracts were re-suspended in 500–800 µL volume of 1∶1 chloroform∶methanol. HPTLC plate (Supelco, Bellefonte, PA) lanes were loaded with 50 µL volumes of individual samples. Seven lanes were done for each sample type. The plates were developed in a TLC chamber containing 45∶18∶3 chloroform∶methanol∶water. Glycosphingolipids were resolved with an orcinol/sulfuric acid stain, followed by heating for 120 °C for 10 minutes. As mentioned in the results, a neutral glycosphingolipid standard was used for comparison. After cooling, the brownish-red bands corresponding to glycolipids were scraped from the plate and the bands from the seven lanes were pooled to obtain adequate material for fatty acid extraction and GC analysis.
|
10.1371/journal.pntd.0002505 | The Improbable Transmission of Trypanosoma cruzi to Human: The Missing Link in the Dynamics and Control of Chagas Disease | Chagas disease has a major impact on human health in Latin America and is becoming of global concern due to international migrations. Trypanosoma cruzi, the etiological agent of the disease, is one of the rare human parasites transmitted by the feces of its vector, as it is unable to reach the salivary gland of the insect. This stercorarian transmission is notoriously poorly understood, despite its crucial role in the ecology and evolution of the pathogen and the disease. The objective of this study was to quantify the probability of T. cruzi vectorial transmission to humans, and to use such an estimate to predict human prevalence from entomological data. We developed several models of T. cruzi transmission to estimate the probability of transmission from vector to host. Using datasets from the literature, we estimated the probability of transmission per contact with an infected triatomine to be 5.8×10−4 (95%CI: [2.6 ; 11.0]×10−4). This estimate was consistent across triatomine species, robust to variations in other parameters, and corresponded to 900–4,000 contacts per case. Our models subsequently allowed predicting human prevalence from vector abundance and infection rate in 7/10 independent datasets covering various triatomine species and epidemiological situations. This low probability of T. cruzi transmission reflected well the complex and unlikely mechanism of transmission via insect feces, and allowed predicting human prevalence from basic entomological data. Although a proof of principle study would now be valuable to validate our models' predictive ability in an even broader range of entomological and ecological settings, our quantitative estimate could allow switching the evaluation of disease risk and vector control program from purely entomological indexes to parasitological measures, as commonly done for other major vector borne diseases. This might lead to different quantitative perspectives as these indexes are well known not to be proportional one to another.
| Chagas disease is a parasitic disease affecting about 10 million people, often living in poor conditions, and the disease contributes to impede their development. As several other infectious diseases (malaria, dengue or sleeping sickness), it is transmitted by blood-feeding insect vectors. While most other human pathogens are directly injected with the vector's saliva, Chagas disease parasite is transmitted through the insect's feces that are deposited on the skin during bloodmeals, which seems to be a very inefficient process. The probability of such transmission to human has thus been very hard to estimate, although it is crucial to predict where people are at risk and design effective control strategies. Using mathematical models integrating data on vectors and humans collected across Latin America, we estimated that for several vector species transmission occurs in 1 over 900–4000 contacts with infected insects. We further showed that our estimate allows prediction of human infection rates in various ecological conditions. These models will provide health policy makers with improved indexes to better prioritize/evaluate of the outcomes of vector control programs.
| Vector-borne diseases represent one of the biggest challenges to current and future human wellbeing. They have severe impacts on many tropical and subtropical countries, where they are responsible for ∼10% of human deaths and contribute to impoverishment by imposing an annual burden of >50 millions of DALYs [1]. They also are an emerging threat for more developed countries as climate change and increasing international exchanges modify the geographic distributions of vectors and pathogens [2].
Vectorial transmission is traditionally thought to critically depend on the incubation period and the survival rate of the pathogen in the vector, and on the frequency of vector feeding on humans. This is well reflected in classical measures of transmission such as the vectorial capacity, entomological inoculation rate or the basic reproductive number, which are central to empirical and theoretical literature on the ecology, evolution and control of vector-borne diseases [3], [4]. Also appearing in these standard measures is the parasite transmission efficacy from infected vectors to hosts, whose effects on vector-borne diseases has received less attention, and has frequently been assumed to be systematic, in particular for malaria transmission [5]. However, efficacy of vector transmission may become a key parameter when it takes on unusually low values, as even small variations could then have major effects on disease dynamics and the resulting prevalence in hosts [6].
The vast majority of causal agents of human vector-borne diseases, such as Plasmodium, Leishmania, dengue and other flaviviruses, are ‘salivarian’ pathogens. After entering the vector during a blood meal, the pathogens multiply inside the gut or haemolymph before spreading to the salivary glands to be directly injected to a human or a reservoir host during a subsequent blood meal. The probability of transmission from vector to a given host species is a complex process that depends on the size of the inoculate and on the within-host dynamics following inoculation. For some salivarian pathogens, the number of pathogens injected at the biting site can be measured [7], as well as the subsequent dynamics of the host-pathogen interactions [8]. Quantitative assessments of the overall resulting probability of transmission based on experimental infections gave values of 0.5 to ∼1 per bite for Plasmodium spp. and dengue virus, 0.3–0.6 for African trypanosome, 0.2–0.4 for Leishmania spp., and as low as 0.01–0.04 for the Japanese encephalitis virus [6], [9].
There also are pathogens for which this probability of transmission can be much lower as they are unable to reach the salivary glands of the vector. The so-called ‘stercorarian’ transmission, sometimes considered as the ancestor of salivarian transmission [10], occurs through the contact of vector's feces and the biting wound (or a mucosa). Successful transmission requires an extraordinary combination of somewhat unlikely events. An infected vector has to defecate sufficiently close to the biting site whilst or shortly after feeding, the infected feces must be brought to the bite wound by the host by scratching, and the pathogen then has to cross the skin of the host to initiate infection [11].
Trypanosoma cruzi is one of the rare parasites that has managed to establish an endemic human infection through this transmission route, and despite its presumably low probability of transmission from vector to human, it has become a major public health problem. It is indeed the etiological agent of American trypanosomiasis, also called Chagas disease, a widely distributed neglected tropical disease in Latin America, with an estimated 8–9 million infected persons and another 25–90 million at risk of infection [12]. Although maternal and oral transmissions have been documented [13], vectorial transmission remains the main cause of human infection. This protozoan kinetoplastid parasite is transmitted by a large diversity of hematophagous bugs of the Reduviidae family to multiple species of sylvatic and domestic mammalian hosts, and at least 20 species of triatomines are involved in transmission to humans [14]. Nonetheless, parasite transmission through these highly diverse vector and host communities remains poorly understood, mostly because the probability of stercorarian transmission can hardly be estimated from experimental infection due to the complexity and rareness of the process. Estimates thus rely on indirect approaches based on a combination of entomological and epidemiological studies at fine temporal and spatial scales. Given the difficulty to collect such integrative datasets, there are currently only three estimates of the probability of stercorarian transmission of T. cruzi to its hosts. This probability was found to be per contact with an infected vector for transmission to human [11], to guinea pigs, a typical domestic host in many Latin American regions [15], and to opossums, the likely ancestral mammalian host of T. cruzi [16]. Although these point estimates seem rather consistently low, their usefulness remains limited as there is no information on their confidence intervals, and no sensitivity analysis to uncertainties in the entomological and epidemiological raw data has been performed. A significant benefit of gaining a robust estimate of the probability of stercorarian transmission would be to establish a clear link between basic entomological data and the prevalence of T. cruzi infection in humans. Potentially, this could allow evaluating disease risk and vector control program in terms of parasitological rather than entomological indexes, as commonly done for other human vector borne diseases such as malaria [4] or dengue [17].
In this contribution we thus aimed at (i) providing a robust quantitative estimate of the probability of stercorarian transmission of T. cruzi to humans based on information available in the literature, and at (ii) determining if this estimate allows predicting the prevalence of the infection from basic entomological data.
Our estimate of the probability of stercorarian transmission of T. cruzi to humans was derived from an indirect approach relying on a standard mathematical relationship which links the number of healthy hosts to become infected, i.e. the incidence of T. cruzi infection, to the average number of potentially infective contacts per individual host, and the probability of pathogen transmission per potentially infective contacts [4]. Given such a relationship, field measurements of the first two quantities allow estimating indirectly the probability of transmission per contact from infected vector to host.
We expanded this standard modelling of transmission to derive the expected distribution of the number of susceptible humans acquiring the parasite in order to obtain a maximum likelihood estimate of the probability of stercorarian transmission. We first used this framework to re-analyze the data from [11], which provides entomological (vector abundance, infection rate, feeding frequency and the proportion of blood meals on humans) and epidemiological (incidence of infection in humans) data at the household scale. Next, we adjusted our model to estimate the probability of transmission to humans (i) when entomological and/or epidemiological data are available at the village rather than at the household scale, and (ii) when epidemiological data consist of human prevalence rather than incidence. These adjustments allowed estimating the probability of stercorarian transmission from four additional datasets. We then performed sensitivity analysis of each estimate to the uncertainties in the measurement of the entomological variables.
Finally, we aimed at testing whether our modelling framework and estimates would allow predicting the prevalence of T. cruzi infection from limited entomological data. We thus predicted the prevalence of infection in different human populations representing nine different epidemiological settings, based on vector abundance and its T. cruzi infection rate combined with our estimate of the stercorarian transmission probability, and compared these predictions to the observed human prevalence.
Vectorial transmission of T. cruzi to humans typically takes place inside the domestic habitat, and an accurate description of transmission thus needs to focus at this scale. Considering a household () with susceptible individuals, the probability of observing new cases during a finite period of time , follows a binomial law characterised by the probability for a susceptible human from that house to become infected during that period:(1)The probability can be linked to the per contact transmission probability of parasites from an infected vector to a susceptible human, , and the per human number of potentially infectious contacts with vectors during a single time unit, . Formally, and as in [11], we have for the household ‘’:(2a)which, as T is typically small for T. cruzi, can be approximated by the catalytic model [18], [19];(2b)where stands for the ‘force of infection’ [19] at the household scale.
The number of potentially infectious contacts per human per time unit can, in turn, be derived from the number of vectors present, ; the proportion of infectious vectors, ; the biting rate of vectors per time unit, ; the feeding rate on humans, ; and finally the total number of individuals in the household, :(3)Combining equations 2 and 3 leads to a non-linear relationship between the abundance of vectors and the incidence in humans.
When all parameters appearing in equations 1–3 are known from field measurements at the household scale, the probability T of transmission can be estimated together with a confidence interval using a standard maximum likelihood approach [20]. Using Eqs. 1–3 the log likelihood function to be maximized can be defined as:(4)where stands for the binomial coefficient. An interval estimate of T is then obtained drawing a maximum likelihood profile as a function of T [20].
However, in most cases, entomological and/or epidemiological data are available at the village rather than at the household scale (see Table 1). Still, a point estimate of T can be derived from Eqs. 1–3, but this requires further assumptions and calculations. Primarily, we have to assume that the parameters appearing in Eq. 3 have the same value across houses, leading to a common number of potentially infectious contacts per human (). Under such an assumption, the probability to become infected is the same in each household (), and the distribution of incidence for the entire village can be modelled by a unique Binomial , where stands for the total number of susceptible individuals in the village. The expectation of this law provides an estimate of as the ratio between numbers of newly infected and susceptible individuals summed over all houses: . From Eq. 2b, one can derive a simple point estimate of T:(5)In addition, epidemiological studies most frequently report human prevalence rather than incidence. Here again, one can use Eqs. 1–3 to derive a point estimate of T under simple additional assumptions. Prevalence typically results from a dynamical equilibrium between the rate at which individuals become infected and the rate at which they die. Assuming that both the force of infection, , and the human death rate, , are constant over time, the prevalence observed in the whole population, , can be used to estimate the force of infection; (see Appendix S1 for details). Using equation 2b, one can then propose another simple point estimate of T;(6)Prevalence data are sometimes provided for children under a given age A, . The force of infection can still be derived under similar assumptions, although it has to be evaluated numerically from the following equation (see Appendix S1 for calculations):(7)From Eq. 2b one can again derive a point estimate of T from the estimated value of since(8)We again point out that both equations 6 or 7 are consistent with other known non-linear relationships between the abundance of vectors and the prevalence of infection in humans. Box 1 provides a guideline to incorporate the entomological and epidemiological knowledge into the modelling proposed in this contribution. Together with Appendices S1 and S2, it also summarizes the assumptions of the models and the potential limitation of this integrative approach.
The assessment of the probability of transmission T using the indirect approaches described above relies on the estimates of the various quantities appearing in equations 1–3. Since they all are subject to estimation uncertainty (Table 1), we performed a sensitivity analysis to such uncertainties in and by determining the distribution of T estimates that resulted from variations in the raw data.
The distributions of plausible estimates were obtained by randomly sampling 1000 values of and in independent zero-truncated normal distributions with mean and standard deviation estimated from the data (Table 1). When no information on the variability in measurements was reported by the authors, we used the standard deviation calculated from dataset 2, as this dataset provides the most comprehensive information about these three parameters. When species-specific information was missing, species-aggregated estimates were used. Since the average values of these parameters were typically larger than 5% (Table 1), sampling from an exact zero truncated binomial distribution would not change the results of the sensitivity analysis.
The sensitivity analysis to the density of vectors, , was performed in a slightly different manner. Since bug collection tends to underestimate vector densities, a correction factor is usually applied to estimate actual vector densities according to the efficacy of the particular collection method. Two different methods were used in these studies; M1 - timed-manual collection with insecticide spraying or aerosol to kill or dislodge triatomines (datasets 1–4), and M2 - passive surveillance by inhabitants or community participation (dataset 5). The correction factor of M1 and M2 have been previously estimated to be around 2.5 [11] and 10 [21], respectively. The sensitivity analysis of the estimate of T to variations in Nvj was performed by varying the correction factor according to a uniform distribution within a range defined as its value ±1.5. This range was chosen as the maximum possible range to guarantee that samples collected with M1 reflected at least the density of bugs actually found (i.e. for the lowest correction factor to be 2.5-1.5 = 1), while keeping the mean of the correction factor equals to the value of 2.5 that was inferred from field data [11]. For consistency the same range was applied to other collection methods (see below for the definition of a last method - M3 - used for datasets that allowed evaluation of the model predictive ability). Accordingly, the range tested for the different methods (M1 : 1–4, M3 : 4.5–7.5, M2: 8.5–11.5) did not overlap, which is consistent with the general understanding that M1 is more efficient than M2, while M3 has an intermediate efficacy.
We then explored the usefulness of our models and our best estimate of the probability of transmission T (i.e. from dataset 1, see results) to predict the prevalence of infection in humans. From the literature, we selected nine additional studies (Table 2) reporting data on triatomine density and infection rates, as well as human prevalence of infection, so that predictions of human prevalence could be derived and compared with the observed values.
As for the sensitivity analyses above, uncertainties in the raw data were taken into account for the predictions of prevalence of infection in humans. We first constructed a distribution of the number of potentially infectious contacts per human () accounting for the variability in and , according to the same zero-truncated normal and uniform distributions as described for the estimates of T. The standard deviations of the zero-truncated normal distributions were again estimated from the data whenever possible, or we used the standard deviation calculated from the most comprehensive dataset (dataset 2). The range of the uniform distribution was determined according to the insect collection method as described above. However, in 3 case studies (datasets 6, 11 and 12), bugs were collected using a third method; M3 - a timed-manual collection by trained personnel without insecticide spraying. We considered this method to have an intermediate efficacy, and we thus set the correction factor to 6 and varied its value within a range set to its value ±1.5.
We then determined the expected distribution of the number of infected individuals by sampling into the binomial distribution given by equation 1, with probabilities calculated from T and the distribution of , and with the number of susceptible humans given by the number of individuals that were tested for Chagas disease in the prevalence studies. Since the expected number of infected individuals could be low, we applied a continuous form of the binomial distribution [22] and sampled it using the ‘acceptance-rejection’ methods [23]. This allowed determining a 95% confidence interval for the expected prevalence, and calculating a p-value as the probability for the observed prevalence to belong to the distribution of predicted prevalence.
We first determined the probability of T. cruzi transmission using dataset 1 which includes data collected at the scale of the household [11]. A profile likelihood was drawn as a function of the probability of transmission (Figure 1). Such profile provided a maximum likelihood estimate of per contact with infected bugs, which was close to the original point estimate of obtained with the same data. The profile likelihood also provided a confidence interval around this estimate, which was per contact, which means that on average 900–4000 contacts with infected vector are needed for a host to become infected (Figure 1). This confidence interval was confirmed by the sensitivity analysis to variations in entomological raw data ( and ). The 95% range of the sensitivity estimates was per contact and the corresponding distribution remained well within the confidence interval derived from the likelihood approach. Only 10% of values from the sensitivity analysis felt outside of the 95% likelihood confidence interval.
From the other four datasets, where entomological and/or epidemiological data were provided at the village scale, we derived additional point estimates of the probability of transmission of T. cruzi to human. These estimates varied from and for T. infestans in Argentina and for T. longipennis and T. barberi in Mexico, to for T. dimidiata in Mexico (Figure 2). They all lied within or very close to the confidence interval derived above from dataset 1 indicating that there is no major difference between the maximum likelihood estimates of T and these four estimates. The estimate for dataset 4 can be viewed as a species-averaged probability of transmission as the data do not allow to distinguish infection from either T. longipennis or T. barberi. The sensitivity analyses to variation in the entomological raw data ( and ) further confirmed the high consistency of those results. The distribution of estimates obtained for the other two datasets on T. infestans were slightly less dispersed than in the first data set (Figure 2A and B) with 95% of the values ranging in for dataset 2, and in for dataset 3. The distribution obtained for T. longipennis and T. barberi (dataset 4) in Mexico were slightly broader with 95% of the estimates found within (Figure 2C). Finally the distributions obtained for T. dimidiata in Mexico (dataset 5), were the most variable since 95% of the values laid within , and for the villages of Teya, Sudzal and the city of Merida, respectively (Figure 2D, E and F). Nevertheless, all estimates were still very consistent with the likelihood-based confidence interval. According to our analysis, including a likelihood estimation with confidence interval, point estimations and sensitivity analyses, the probability of stercorarian transmission of T. cruzi to human is estimated to be in the range - per contact.
We then attempted to predict human prevalence of infection based on the probability of transmission determined above and basic entomological data, using nine independent case-studies (Table 3). Those included T. infestans in Peru and in southern Cochabamba in Bolivia, T. barberi, T. mexicana, and T. dimidiata in Mexico and Costa Rica, T. brasiliensis and T. pseudomaculata in Brazil, T. pallidipennis and T. longipennis in Mexico, and R. prolixus and P. geniculatus in Venezuela.
In seven of these cases, our model satisfactorily predicted human prevalence (Table 3), as indicated by a lack of significant differences between observed and predicted prevalence. In three cases (T. infestans in northern Cochabamba, Bolivia, and in Paraguay, and T. maculata in Venezuela) there were statistical differences between the observed and predicted prevalence. However, in the later two cases, the observed and expected prevalences of infection were of very similar magnitude and almost included in the 95% confidence interval. Accordingly, the difference between the high level of transmission by T. infestans in Paraguay and the weak level of transmission by T. maculata in Venezuela were thus properly predicted, so that our model would not lead to any lack of appreciation of a serious health issue. Only in one instance, T. infestans in Northern Cochabamba, Bolivia, a large discrepancy was found and could not satisfactorily be explained from available data.
We thus obtained accurate predictions of human prevalence of T. cruzi infection over a broad range of epidemiological conditions ranging from low to high prevalence of infection (), a wide geographic range (with 7 countries across Latin America), and 12 species of triatomines.
The lack of a quantitative estimate of the probability of T. cruzi transmission to human through the feces of the vector has hindered the development of approaches that integrate ecological and epidemiological information on Chagas disease. These approaches have had an impressive influence in mitigating several vector-borne diseases including malaria [47], dengue [48] or leishmaniasis [49], and would help better understand the complex features of the transmission of T. cruzi and compare it with other vector-borne diseases. Based on data from the literature we built here epidemiological models to derive 6 estimates of this probability of transmission, all being of the order of 10−4–10−3 per contact. This primarily illustrates the paradox of Chagas disease; despite the ‘milli-transmission’ of the parasite from vectors to humans, the disease affects millions of people across the Americas. The quantitative knowledge of its transmission probability also opens new perspectives for the study of the disease, with key implications for both parasite evolution and public health policy.
The probability of transmission of T. cruzi from triatomine vectors to humans was found to be very small, per contact with infected vector (95% CI ), relatively consistent across the different study systems, with point estimates ranging from to , and in agreement with the only other point estimate available in the literature [11]. This narrow range of probabilities was observed in spite of marked differences in vector density, vector species (taxonomic, ecologic and behavioural differences), prevalence of infection in humans and vectors, resulting in very different epidemiological situations. Such a broad consistency was confirmed by our sensitivity analyses, which further supported that estimates are robust to changes in the entomological and epidemiological raw data used for their calculation.
These estimates of parasite transmission to human are similar to what has been calculated for guinea pigs [15], but differ substantially with the probability of stercorarian transmission to opossums that was estimated to be 10–100 times larger [17]. This suggests a reduced adaptation of T. cruzi to domestic hosts compared to its likely ancestral and sylvatic host, which is consistent with the much shorter period of coevolution between T. cruzi and humans. Indeed, estimates suggest around ∼10000 years of coevolution of T. cruzi with humans, compared to ∼80 millions years with the opossum [50]. Such a low probability of transmission does not mean that humans are of secondary importance or even ‘dead ends’ in term of parasite transmission, as suggested by the ∼40% prevalence of infection in humans observed in 9000-years old mummies [51] as well as in today's highly endemic areas [45]. In fact, the potential amplification and dilution effects [52] that human and other domestic hosts could have on the populations of T. cruzi still remain to be properly quantified (but see [53]).
A low probability of vectorial transmission of T. cruzi was expected, given the succession of unlikely events required to occur and the many parameters involved. However, the narrow range of probabilities was more surprising given that all of these parameters could potentially affect parasite transmission quite dramatically. This suggests that these parameters combine in an independent way to produce an almost universal efficacy of transmission of T. cruzi from vectors to humans. While more accurate data may allow refining our estimate of the probability of transmission of T. cruzi to human, potentially detecting species specificity, the residual variations in the probability of transmission are expected to have little impact on the prevalence of infection in humans. Indeed, we were able to predict rather accurately the prevalence in humans from infected vector density, the frequency of human blood meal, and a unique probability of vectorial transmission. Triatomine vectorial capacity is thus primarily dependent upon vector density and feeding frequencies on specific hosts, a conclusion which is consistent with the key influence of those parameters on the spread and persistence of the disease [6]. As vector demography has been documented for a variety of triatomine species and entomological context [21], [53], [54], [55], the emerging eco-epidemiology of Chagas disease would benefit from a substantial improvement of our knowledge on vector feeding ecology. The emergence of methods based on the use of metagenomics [56] and stable isotopes, which potentially allow identifying vector trophic networks [57], should shortly allow tackling the transmission of T. cruzi in the context of host communities, as it has already been done successfully for the transmission of plague [58].
The very low probabilities of transmission of T. cruzi from vector to vertebrate hosts raises an obvious evolutionary question: why has T. cruzi not evolved from a stercorarian to a salivarian mode of transmission while closely related species such as T. rangeli [59] and T. brucei [60] or Leishmania [61] have been able to do so? A first line of explanation could lie in a lower ‘evolvability’ of T. cruzi [62]. However, there is no evidence that T. cruzi has a lower mutation rate compared to other kinetoplastids [63], and T. cruzi does experience reproduction and recombination [64] as demonstrated for other related taxa [65], [66], [67]. A lower ‘evolvability’ could also result from specific features in T. cruzi genotypes to phenotypes mapping function, which may be evaluated by mutagenesis and artificial selection experiments [62]. A second line of explanations that could explain T. cruzi ineffective mode of transmission are the costs associated with the migration of the parasite across the midgut, the escape of the immune response in the haemocoel, and the invasion of the salivary glands, which all may exert selective pressure to restrict the parasite to the gut. Those costs have not been identified yet, although molecular studies are progressively unravelling the interaction between T. cruzi and its vector [14], [68], and insights could be gained by comparative analysis of vector-parasite interactions of the various kinetoplastids [14]. Comparisons with T. rangeli, a closely related and sympatric parasite that shares hosts and vectors with T. cruzi [69], should be especially informative as it is able to colonize the haemocoel and reach the salivary glands of its vector [59]. Finally, the selective pressure on T. cruzi may be too weak given its potential for direct transmission which is known to be of evolutionary and epidemiological importance in opossum [16], [70] as well as in human either because of oral or maternal transmission [13].
Our quantitative assessment of the probability of transmission of T. cruzi offers new opportunities to tackle these key eco-evolutionary questions, as it allows quantifying standard epidemiological measures such as R0 or related quantities [3], [4], [71] which have been consistently missing in the epidemiology of Chagas disease [6], [ but see 72], while they are central tenets of the study of the epidemiological dynamics of malaria [4], dengue [17] and others human and livestock vector-borne diseases [73].
Our study demonstrates that the quantitative estimates of the probability of T. cruzi transmission from vector to humans allow expressing infection risk in terms of human incidence or prevalence, rather than in terms of purely entomological indexes such as vector presence/absence [74] or abundance [75]. Importantly, the entomological data used to make those predictions are basic estimates of vector abundance and infection rates that can inferred from typical entomological collections achieved by trained personnel or even low-cost studies based on community participation [76]. Although a proof of principle study is necessary to validate the proposed models' predictive ability in an even broader range of ecological settings, this approach offers a much more affordable way than large-scale serological surveys to estimate human prevalence over geographic areas and obtain better descriptions of the global and local burden of the disease. Such prevalence/incidence data would be more straightforward and explicit to interpret at all levels of public health systems for the design of epidemiological surveillance and vector control operations. In addition, the risk expressed in incidence or prevalence would likely differ from that expressed in vector presence/absence or abundance because, according to the catalytic model [18], the relationship between these variables is non-linear and follows a cumulative exponential distribution. At low vector densities the risk of human transmission increases rapidly with vector abundance, and is thus likely to be underestimated by the sole measure of vector abundance, while at large vector densities, human incidence and prevalence reach a plateau, so that variations in vector abundance have little or no influence on the already high transmission to human.
Our model can also profoundly help the assessment of the efficacy of vector control interventions, which is traditionally measured in terms of reduction in vector abundance or vector presence (infestation index) [53], [77], [78]. Typically, current guidelines for vector control in several endemic countries aim at reducing triatomine presence below the somewhat arbitrary level of 5% of the houses of a community [79], based on the assumption that this may be sufficient to dramatically reduce or even interrupt parasite transmission to humans. The modelling developed here opens the possibility to convert a reduction of vector abundance into a variation in the actual level of parasite transmission to humans, allowing to rapidly define more rational target/threshold levels of infestation for vector control. Again, one expects that for high vector densities, very large reduction in vector populations will be needed to reduce human prevalence, while at low vector abundance, small reductions in vectors could result in significant decrease in human prevalence. Nonetheless, even small residual populations of bugs due to incomplete treatment [80], development of insecticide resistance [81] or infestation by non-domiciliated vectors [21], [55], [82], [83], [84], [85] will be sufficient to maintain an active transmission of T. cruzi to humans, which clearly appeals for the use of highly sensitive tools for entomological surveillance following the ‘action’ stage of control program [86].
In conclusion, we provided estimates of the probability of T. cruzi transmission from vector to human, which were shown to be highly consistent across vector species and epidemiological conditions. Such a new quantitative knowledge could allow expanding purely entomological indexes, which are typically calculated for triatomines, into parasitological measures (such as R0, the so-called force of infection and resulting incidence and prevalence), as routinely done for other human vector-borne diseases such as malaria or dengue. This offers the possibility to develop a better understanding of the ecology and evolution of one of the rare stercorarian human parasites. This also is of primary interest in public health, as parasitological measures provide a more straightforward evaluation of the disease risk and a better description of the outcomes of vector control program in terms of human infection rather than vector abundance. Studies specifically designed to validate our models' predictive ability in an even broader range of entomological and ecological settings would now be worth performing to strengthen the proposed approach, and to allow for its use in a large scale operational/policy setting.
|
10.1371/journal.ppat.1006589 | The transcription factor CHOP, an effector of the integrated stress response, is required for host sensitivity to the fungal intracellular pathogen Histoplasma capsulatum | The ability of intracellular pathogens to manipulate host-cell viability is critical to successful infection. Some pathogens promote host-cell survival to protect their replicative niche, whereas others trigger host-cell death to facilitate release and dissemination of the pathogen after intracellular replication has occurred. We previously showed that the intracellular fungal pathogen Histoplasma capsulatum (Hc) uses the secreted protein Cbp1 to actively induce apoptosis in macrophages; interestingly, cbp1 mutant strains are unable to kill macrophages and display severely reduced virulence in the mouse model of Hc infection. To elucidate the mechanism of Cbp1-induced host-cell death, we performed a comprehensive alanine scanning mutagenesis and identified all amino acid residues that are required for Cbp1 to trigger macrophage lysis. Here we demonstrate that Hc strains expressing lytic CBP1 alleles activate the integrated stress response (ISR) in infected macrophages, as indicated by an increase in eIF2α phosphorylation as well as induction of the transcription factor CHOP and the pseudokinase Tribbles 3 (TRIB3). In contrast, strains bearing a non-lytic allele of CBP1 fail to activate the ISR, whereas a partially lytic CBP1 allele triggers intermediate levels of activation. We further show that macrophages deficient for CHOP or TRIB3 are partially resistant to lysis during Hc infection, indicating that the ISR is critical for susceptibility to Hc-mediated cell death. Moreover, we show that CHOP-dependent macrophage lysis is critical for efficient spread of Hc infection to other macrophages. Notably, CHOP knockout mice display reduced macrophage apoptosis and diminished fungal burden and are markedly resistant to Hc infection. Together, these data indicate that Cbp1 is required for Hc to induce the ISR and mediate a CHOP-dependent virulence pathway in the host.
| Histoplasma capsulatum is the causative agent of histoplasmosis, a fungal infection that can be fatal in a wide range of mammalian hosts, including otherwise healthy, immunocompetent individuals. Histoplasma cells replicate to very high levels within host macrophages, eventually causing macrophage death and the release of live fungal cells. Here, we show that Histoplasma yeast use the secreted protein Cbp1 to activate a specific signaling pathway in the host cell to cause macrophage death during infection. Importantly, this signaling cascade is essential for pathogenesis, and mice deficient for a central component of this pathway are resistant to Histoplasma infection. Our study is the first demonstration that Histoplasma employs a secreted effector to alter host signaling pathways to promote virulence, and thus provides key insight into the pathogenesis strategies of this important human fungal pathogen.
| Death of the host cell has profound consequences for an intracellular pathogen. It may eliminate the pathogen’s replicative niche, or it may promote pathogen spread. The specific form of cell death can impact immune activation of the host: pyroptosis and necroptosis are highly inflammatory types of cell death whereas apoptosis is generally characterized as immunologically quiet, although it does promote cross-presentation of antigens by dendritic cells. Thus, it is unsurprising that many intracellular pathogens have evolved various mechanisms and strategies to manipulate host-cell survival and death [1,2]. For example, Mycobacterium tuberculosis has been shown to inhibit apoptosis in macrophages to promote bacterial survival and replication, but also promote necrosis, allowing for bacterial spread [3].
The intracellular fungal pathogen Histoplasma capsulatum (Hc) is adept at manipulating the viability of macrophages during infection. Hc causes respiratory and systemic disease in a wide range of mammalian hosts, including immunocompetent individuals. Hc is a thermally dimorphic fungus, growing in the soil at ambient temperatures as a saprophytic mold that produces spores. When the soil is disturbed, spores and hyphae aerosolize and can be inhaled by mammalian hosts. After inhalation, exposure to mammalian body temperature is sufficient to induce a morphogenetic switch of the fungus, resulting in unicellular yeast-form growth. Hc yeast survive and replicate within host macrophages, eventually triggering host-cell death and allowing the release of live yeast cells.
Hc-dependent macrophage lysis is dependent on Cbp1, a secreted virulence factor of Hc that is produced specifically by the yeast phase of Hc [4]. In fact, Cbp1 is the most abundant secreted protein present in Hc culture supernatants [4], and it is critical for robust macrophage death and in vivo pathogenesis [5,6]. We previously demonstrated that Cbp1 is dispensable for high fungal burden but required for host-cell death, suggesting that it actively promotes macrophage death during infection. Indeed, during Hc infection, caspase-3/7 activation and host-cell death is fully dependent on Cbp1 and partially dependent on the pro-apoptotic host proteins Bax and Bak, suggesting that Cbp1 promotes apoptosis in Hc-infected macrophages [6].
Interestingly, Cbp1 has no identifiable protein domains and few identified homologs [7], and the mechanism by which it induces host-cell death is unknown. To generate tools to explore the host pathways that are impacted by Cbp1, we performed an alanine-scanning mutagenesis of Cbp1 and identified point mutants that were unable to lyse macrophages during Hc infection. These alleles were used to explore the mechanism by which Hc induces apoptosis in a Cbp1-dependent manner. We determined that Cbp1 activates the integrated stress response (ISR), which is one of several pathways that can trigger apoptosis in mammalian cells. The ISR is an intracellular signaling cascade that is activated after exposure to a variety of stresses, including endoplasmic reticulum (ER) stress and amino acid starvation. The central signaling event is the phosphorylation of the α-subunit of the eukaryotic initiation factor eIF2, which leads to a global reduction in translation levels. Despite this general translational inhibition, the translation of several key transcripts is promoted, resulting in a stress response that initially facilitates return to homeostasis but leads to apoptosis if the stress is unresolved. Here we demonstrate that Hc strains expressing lytic CBP1 alleles trigger the integrated stress response (ISR) in macrophages during infection, leading to an induction of the pro-apoptotic genes CHOP and Tribbles 3 (TRIB3). Non-lytic CBP1 alleles are unable to induce CHOP and TRIB3. We further show that these components of the ISR are necessary for robust macrophage death during in vitro infection. Finally, we show that the fungal burden of Hc is significantly reduced in CHOP knockout mice, which are highly resistant to Hc infection. Together, these data demonstrate that Hc uses the secreted protein Cbp1 to activate the ISR in macrophages, resulting in full virulence of the pathogen in in vitro and in vivo infection models.
To generate alleles of Cbp1 that perturb the ability of Hc to lyse macrophages, we conducted alanine scanning mutagenesis on Cbp1 in the highly virulent G217B strain background. Each of the 63 non-alanine amino acids in the 78 amino acid Cbp1 sequence were individually changed to alanine, and the resultant mutant CBP1 alleles were expressed in the cbp1 mutant strain. The resulting strains were then screened to determine which mutant proteins were present at high levels in yeast culture supernatants. Of the 63 mutant proteins, 20 were not detected in yeast culture supernatants (S1 Table; S1 Fig). We next examined whether the 43 secreted mutant proteins were able to rescue the defect in macrophage lysis seen with our cbp1 mutant strain. To do so, we used the corresponding 43 strains to infect J774.1 cells, a murine macrophage-like cell line whose lysis during Hc infection is dependent on Cbp1 [6]. Mammalian cell death was qualitatively assessed by fixing and visualizing the monolayer with methylene blue (S1 Fig). Of the secreted mutants, 26 alanine mutants showed a reduction in macrophage lysis (S1 Table). The lysis defects of two of these secreted mutants, D10A and D58A (S2A Fig), were validated quantitatively using infected murine bone marrow-derived macrophages (BMDMs). The D10A mutant was able to partially complement the macrophage lysis defect of the cbp1 strain, whereas D58A was completely defective for macrophage lysis (Fig 1). Importantly, the reduction in macrophage lysis after infection with the D10A mutant was not due to any differences in fungal burden; this mutant showed the same intracellular growth rate as wildtype Hc (S2B Fig). In contrast, the non-lytic D58A mutant showed reduced growth kinetics in macrophages (S2B Fig).
In addition to generating an allelic series of CBP1 mutants by mutagenesis, we also examined the function of the CBP1 homolog from the closely related fungus Paracoccidioides brasiliensis, another thermally dimorphic intracellular pathogen. P. brasiliensis, which is also able to colonize macrophages [8], is one of the few fungi with a clear ortholog of Hc Cbp1, although the function of the Pb protein is completely uncharacterized. Pb Cbp1 differs from Hc Cbp1 at 32 of its 77 amino acids (S3 Fig). Based on our perusal of general transcriptomics and proteomics studies in Pb, the CBP1 transcript is highly expressed in P. brasiliensis yeast cultures compared to mycelia [9], and the protein is highly abundant in yeast culture supernatants [10]. We expressed a fusion protein consisting of the G217B Cbp1 signal peptide followed by the predicted mature Cbp1 from the Pb03 P. brasiliensis strain in our Hc cbp1 mutant strain. After confirming that the protein (Pb_Cbp1) was secreted into Hc culture supernatants (S2A Fig), we then determined that it was able to fully rescue the lysis defect of the Hc cbp1 mutant (Fig 1), with no change in intracellular growth kinetics compared to wildtype Hc (S2B Fig). We decided to use the wild-type Hc Cbp1, the nonlytic D58A mutant, the partially lytic D10A mutant, and the fully lytic Pb_Cbp1 strains to elucidate how Cbp1 causes macrophage death during Hc infection.
We previously identified a set of host genes that are upregulated in Hc-infected macrophages in a Cbp1-dependent manner, and many of these same genes are induced during ER stress [6]. This finding was intriguing for two reasons: many intracellular pathogens manipulate ER stress pathways in the host [11], and unresolved ER stress can lead to apoptosis through many mechanisms [12]. To determine if Hc triggers ER stress in macrophages, we assessed the activation of the unfolded protein response (UPR) during infection. The mammalian UPR, which counters the effects of ER stress, consists of three distinct signaling pathways triggered by three ER-resident proteins that are able to detect ER stress: IRE1, ATF6, and PERK [12–14] (S4 Fig). Activated IRE1 is an RNase which removes a non-canonical intron in the transcript encoding the transcription factor Xbp1 [15]. To examine IRE1 activation, we compared levels of unspliced and spliced Xbp1 transcripts (Xbp1u and Xbp1s, respectively) in macrophages infected with Hc or treated with the potent ER stress-inducer tunicamycin (Tm), which activates all three branches of the UPR. Infection with Hc did not result in an increase in Xbp1s compared to uninfected macrophages, whereas Xbp1s was detected in macrophages as early as 9 hours after Tm treatment (Fig 2A). To further assess IRE1 activation, we also examined the induction of the Xbp1 target gene ERdj4 [16]. Similar to Xbp1s, ERdj4 was upregulated 9 hours after the addition of Tm but was not upregulated in Hc-infected macrophages (Fig 2B). We then examined ATF6, which is cleaved during ER stress, resulting in the transcription factor ATF6n. To assess the activation of ATF6, we monitored the upregulation of one of its specific targets, SEL1L [17]. Similar to ERdj4, SEL1L was induced by Tm but not induced in Hc-infected macrophages (Fig 2C). We also examined the transcript levels of the chaperone protein BiP/GRP78, a classic marker of ER stress whose upregulation is dependent on both IRE1 and ATF6 [18]. BiP was strongly induced after Tm treatment, but was not upregulated in macrophages after Hc infection (Fig 2D). Next, we looked at the third and final mammalian UPR sensor, PERK, which oligomerizes and autophosphorylates upon activation [19,20]. We were unable to detect phospho-PERK in Hc-infected macrophages (Fig 2E), suggesting that Hc infection does not activate PERK. To confirm this hypothesis, we utilized the PERK-specific inhibitor GSK2606414 [21]. Consistent with previous observations [22], inhibition of PERK signaling increased caspase-3/7 activity in cells after tunicamycin treatment. However, GSK2606414 had no effect on caspase-3/7 activity in Hc-infected macrophages (Fig 2F), consistent with our observation that PERK is not activated during Hc infection. Taken together, these data demonstrate that Hc does not activate the UPR in infected macrophages.
After determining that Hc-infected macrophages are not undergoing ER stress, we then re-examined our transcriptional profiling results [6]. Many of the genes that are upregulated during infection in a Cbp1-dependent manner are involved in the integrated stress response (ISR), a pathway activated in response to a variety of cellular stresses, including ER stress. Thus, we shifted our focus to the ISR. The central event in the ISR signaling cascade is the phosphorylation of the α-subunit of the translation initiation factor eIF2 [23,24]. Phosphorylation of eIF2α leads to an overall reduction in translation, but allows for the increased translation of select, stress-responsive transcripts due to unique features in their mRNAs. To first examine ISR activation, we assessed phospho-eIF2α levels in macrophages by Western blot. We observed an increase in phospho-eIF2α in macrophages infected with Hc strains expressing lytic alleles of CBP1. Though it was a modest increase in phosphorylated eIF2α levels in macrophages infected with lytic Hc strains, it was comparable to the increase seen in cells treated with Tm (Fig 3A), and small changes in phospho-eIF2α levels are known to have large effects in the cell [25]. Importantly, we observed this increase in phospho-eIF2α at 12 hours post-infection (hpi), which is well before the onset of macrophage lysis (Fig 1). To confirm ISR activation, we next looked at the levels of ATF4, a transcription factor that is preferentially translated as phospho-eIF2α levels increase [26,27]. As expected, we saw robust levels of ATF4 in macrophages infected with Hc expressing lytic alleles of CBP1. ATF4 induction preceded the onset of macrophage lysis and correlated with phospho-eIF2α levels (Fig 3B).
To continue our exploration of the ISR, we also examined the expression of two downstream targets of ATF4, the transcription factor CHOP/DDIT3 and pseudokinase TRIB3 [28], which were previously implicated by our transcriptional analysis of infected cells [6]. Upregulation of CHOP and TRIB3, a target of CHOP, has been shown to be pro-apoptotic in a variety of cell types in response to multiple stresses [29–33]. Both CHOP and TRIB3 showed robust induction in macrophages after infection with lytic Hc strains (Fig 3C), and this correlated with protein levels (S5 Fig). Notably, there was intermediate induction in macrophages infected with Hc expressing the partially lytic D10A CBP1 alanine mutant (Fig 3C). Furthermore, robust TRIB3 expression after infection with Hc secreting wildtype Cbp1 is dependent on CHOP, as CHOP-/- macrophages showed reduced TRIB3 induction after infection with wildtype Hc or the complemented cbp1 mutant strain (S6 Fig). Together, these data demonstrate that Hc triggers the ISR in macrophages before the onset of host-cell death, and the activation of the ISR is dependent on the expression of lytic CBP1 alleles. Importantly, at 48 hpi, we did not see an increase in phospho-eIF2α, ATF4, CHOP, or TRIB3 levels in macrophages infected with non-lytic Hc strains, even though the fungal burden of the non-lytic strains at this time point exceeds that of the lytic strains at 12 hpi (S2 Fig). Thus, a high intracellular fungal burden is not sufficient to activate the ISR in infected macrophages.
To determine if ISR activation was a general response of different macrophage types to divergent Histoplasma strains, we varied either the Hc strain or the host macrophage and assessed ISR acviation by examining CHOP and TRIB3 induction. First, we infected BMDMs with G186AR, a clinical isolate from a different Histoplasma clade than G217B [34]. We observed robust CHOP and TRIB3 induction in infected BMDMs before the onset of macrophage death (S7A Fig). Next, we used the G217B strain to infect U937 cells, a human cell line that can be differentiated to a macrophage-like state. Similar to what we observed with BMDMs, these human cells displayed Cbp1-dependent lysis as well as Cbp1-dependent CHOP and TRIB3 induction during Hc infection (S7B Fig). Together, these data indicate that ISR activation in macrophages during Hc infection occurs in multiple biological contexts.
The pro-apoptotic pseudokinase Trib3 is hypothesized to promote apoptosis by binding Akt and preventing its phosphorylation, thereby reducing its pro-survival and anti-apoptotic signaling [35]. After determining that TRIB3 is induced during Hc infection, we then evaluated phospho-Akt levels in infected macrophages. Macrophages infected with lytic Hc strains showed a reduction in phospho-Akt levels at 12 hpi, whereas macrophages infected with non-lytic Hc strains showed no such reduction (Fig 4A). Importantly, those macrophages infected with non-lytic Hc strains showed no reduction in phospho-Akt, even 48 hpi (Fig 4B), when the fungal burdens of the non-lytic strains exceed those of the lytic strains at 12 hpi (S2 Fig).
In addition to reducing phospho-Akt levels, ISR signaling has been shown to promote apoptosis by activating caspase-8 [36]. Caspase-8 is an initiator caspase that is able to directly activate downstream caspases, thereby bypassing the need for mitochondrial outer membrane permeabilization (MOMP), canonically considered the central signaling event in the intrinsic apoptosis pathway [37]. Macrophages infected with lytic Hc strains showed an increase in caspase-8 activity that was comparable to that of macrophages treated with tunicamycin. In contrast, macrophages infected with non-lytic Hc strains showed no increase in caspase-8 activity when compared to uninfected macrophages (Fig 5). Thus, expression of functional Cbp1 led to caspase-8 activation in macrophages during Hc infection.
Having concluded that the ISR was activated before the onset of lysis, we next wanted to assess the relevance of this signaling pathway during in vitro infection. To this end, we generated BMDMs from CHOP-/- [33] and TRIB3-/- [38] knockout mice. Both CHOP-/- and TRIB3-/- macrophages showed a reduction in caspase-3/7 activity after infection with wildtype Hc (Fig 6A). The reduction in the activity of these executioner caspases correlated with a reduction in host-cell lysis after infection with wildtype Hc (Fig 6B). Importantly, these reductions in macrophage death were not due to any differences in fungal burden (S8 Fig). Thus, macrophages individually deficient for two different components of the ISR showed a reduction in lysis, demonstrating that activation of the ISR by Hc is required to trigger robust host-cell death during infection.
Intracellular pathogens can trigger host-cell death to facilitate efficient exit from a host cell, thereby promoting subsequent spread of infection to other host cells. Since optimal host-cell death during Hc infection is dependent on CHOP, we next tested whether CHOP affects Hc spread to other macrophages. We hypothesized that the reduction in lysis in CHOP-/- macrophages would result in the release of fewer fungal cells, thereby reducing the spread of Hc to other macrophages. To test this hypothesis, we utilized the underside of transwells to seed wildtype or CHOP-/- BMDMs that were mock infected or infected with wildtype Hc. These samples were then placed over uninfected, wildtype BMDMs so that these host cells were accessible to Hc released from the original wildtype or CHOP-/- macrophages (Fig 7A). After the transwell macrophages had begun to lyse, but before the bottom macrophages lysed, we removed the transwells and assessed fungal burden in the lower macrophages. The bottom macrophages that had received Hc from infected CHOP-/- BMDMs had a significantly lower fungal burden and reduced host-cell death compared to the bottom macrophages that had received Hc from wildtype BMDMs (Fig 7B and 7C). These data support the hypothesis that the reduced death of CHOP-/- macrophages results in reduced spread of Hc, thereby limiting pathogenesis.
Having determined that the ISR is necessary for full Hc virulence during in vitro infection, we next wanted to examine the relevance of the ISR during in vivo infection. We first asked if there was reduced macrophage apoptosis in CHOP-/- mice following Hc infection. We infected CHOP-/- and C57BL/6 mice intranasally with 1 x 106 mCherry-producing Hc yeast [39] and performed flow cytometry analysis on lung homogenates at 3 dpi. There was no difference in the percentage of infected (i.e., mCherry+) cells in wild-type and CHOP-/- animals at this time point (S9A Fig), although we were unable to determine the precise fungal burden per cell. Despite comparable numbers of infected cells, there was a significant reduction in the percentage of apoptotic alveolar macrophages in the lungs of infected CHOP-/- animals compared to infected wild-type animals (Fig 8A).
We then tested whether this reduction in host-cell death correlated with a reduced fungal burden in vivo, especially in light of our observation that CHOP is required for efficient spread of Hc in vitro (Fig 7). We infected female CHOP-/- and C57BL/6 mice intranasally with a sub-lethal dose (3 x 105 colony-forming units (CFU)/mouse) of wildtype Hc and then monitored fungal burden in the lungs and spleens of the infected animals. Even by 3 dpi, we observed a significant reduction in fungal burden in the lungs of CHOP-/- mice compared to wildtype mice (Fig 8B). Similar to what we observed in the lungs, there was a significant reduction in Hc recovered from the spleens of CHOP-/- mice compared to wildtype animals beginning at 5 dpi (Fig 8B). Importantly, this difference in fungal burden arises early during infection, before the activation of the adaptive immune response [39,40], consistent with our hypothesis that CHOP plays an important role in innate immune cells, such as macrophages, during Hc infection. Furthermore, as further evidence of ISR induction in vivo, we monitored TRIB3 expression in the lungs of CHOP-/- mice at 1 dpi, when there is no statistically significant difference in fungal burden between wildtype and mutant mice. Consistent with our in vitro observations, we observed that robust TRIB3 induction in vivo is dependent on CHOP (S9B Fig; Fig 8B).
Finally, we then examined if the difference in fungal burden correlated with the outcome of infection. We infected female CHOP-/- and C57BL/6 mice intranasally with a lethal dose (1 x 106 CFU/mouse) of wildtype Hc and then monitored the animals daily for symptoms, including weight loss, hunching, panting, and lack of grooming. While all infected animals developed symptoms, including weight loss (S9C Fig), the wildtype mice (n = 11) showed more severe symptoms, with over half of the animals succumbing to infection. Notably, the CHOP-/- mice were significantly more resistant to infection (Fig 8C), indicating that CHOP is required for virulence of Hc in vivo. Together, these data support the model that Hc activates the ISR in a CHOP-dependent manner, thereby promoting host-cell death and facilitating pathogen spread during infection.
Hc is a primary fungal pathogen that is able to replicate to very high levels within host macrophages before host-cell lysis occurs. However, as we previously demonstrated, high intracellular fungal burden is not sufficient to induce host-cell death. This conclusion is based on the phenotype of mutant strains that lack the small, secreted protein Cbp1: the cbp1 mutant replicates to high levels within macrophages but fails to lyse them [6]. Here, we demonstrate that Hc strains expressing functional alleles of CBP1 induce a host signaling pathway called the integrated stress response (ISR) in infected macrophages. Two pro-apoptotic components of ISR signaling, the transcription factor CHOP and the pseudokinase TRIB3, are necessary for robust macrophage death during in vitro infection. Furthermore, we show that CHOP is required for efficient spread of Hc during macrophage infection. Finally, in the mouse model of Histoplasma infection, CHOP is required for optimal macrophage apoptosis, fungal burden, and host sensitivity to Hc. The phrase “Integrated Stress Response” is a term that has been used for over a decade to describe a signaling cascade that is activated after mammalian cells undergo a variety of stresses, including ER stress, amino acid starvation, and iron deprivation [23,24,41]. The key event in this signaling pathway is the phosphorylation of the alpha subunit of the translation initiation factor eIF2. There are four known eIF2α kinases: general control non-depressible 2 (GCN2), which is activated by amino acid starvation; heme-regulated eIF2α kinase (HRI), which is activated during heme deprivation; protein kinase R (PKR), which is activated by cytosolic double-stranded RNA; and PKR-like ER kinase (PERK), which is activated by ER stress [42,43]. Phosphorylation of eIF2α by these kinases causes a reduction in global translation within the cell. Specific transcripts, however, are preferentially translated as phospho-eIF2α levels increase, resulting in a response geared to help the cell adapt to the stress. If the cell is unable to overcome the stress that initiated ISR signaling, the cascade ultimately leads to apoptosis of the cell through induction of many pro-apoptotic factors, including the transcription factor CHOP and its target, the pseudokinase Tribbles 3 [12].
The ability of the ISR to either combat stress or drive the cell towards apoptosis makes it a key target for intracellular pathogens. One well-studied mechanism of ISR manipulation by pathogens is alteration of ER stress signaling pathways. Different types of intracellular pathogens, ranging from viruses [44], to bacteria [11], to eukaryotic parasites [45,46], trigger the ISR in their host cells by inducing ER stress. Unsurprisingly, activation of the ISR has a range of outcomes, depending on the pathogen and the host cell. For example, the intracellular bacterium M. tuberculosis induces ER stress in macrophages, ultimately leading to apoptosis [47]. However, activation of this host signaling pathway is detrimental for the bacteria, as increasing phospho-eIF2α levels reduces bacterial burden, and similarly, reducing CHOP levels increases bacterial burden [48].
In the case of Hc, the ISR is triggered without any evidence of ER stress induction, and indeed it is clear that other microbial pathogens promote ISR activation via multiple pathways. For example, conditioned media from Pseudomonas aeruginosa cultures has been shown to induce ISR activation that is dependent on HRI. Additionally, P. aeruginosa also induces ER stress, although the functional outcome of these two ISR activating events during infection is unclear [49]. Reoviruses have been shown to trigger the ISR in host cells by activating PKR, an eIF2α kinase that directly senses cytosolic dsRNA [50,51]. This signaling benefits the pathogen, as attenuation of the ISR response by blocking eIF2α phosphorylation or ATF4 accumulation results in reduced viral replication [52]. In the case of Hc, the mechanism of Cbp1-dependent ISR induction is unknown. Since Cbp1 is a protein with no known functional domains, it is intriguing to consider exactly how this virulence factor triggers the ISR during Hc infection, and work to elucidate this mechanism is underway.
To advance a mechanistic understanding of Cbp1, we identified amino acids in Cbp1 that are critical for its ability to trigger host-cell lysis (S1 Table). These residues may affect Cbp1 protein structure or a potential ligand interaction site. Thus, the alanine mutant library generated here provides a significant tool for future structure-function studies. Additionally, we demonstrated that the presumptive Cbp1 ortholog from the closely related fungal pathogen P. brasiliensis (Pb) is able to fully complement the lysis defect of Cbp1-deficient Hc. To our knowledge, this is the first functional characterization of Pb Cbp1. It is intriguing to note that the sequences of Hc Cbp1 and Pb Cbp1 show many differences, with nearly half of the residues differing between G217B (Hc) and Pb03 (Pb) Cbp1 (S3 Fig). While our alanine scan results suggest that many of the N-terminal amino acid residues in Hc Cbp1 are necessary for macrophage death (S1 Table), Pb Cbp1 lacks these charged residues but is still able to induce robust host-cell lysis (S3 Fig). Additionally, and perhaps unsurprisingly, residues that are highly conserved between Hc Cbp1 and Pb Cbp1, such as D58, are essential for robust macrophage death. Interestingly, one of the few other fungal organisms that contains a Cbp1 homolog in its genome is Emmonsia crescens (S3 Fig). Emmonsia spp. are emerging dimorphic fungal pathogens [53], and the role of Cbp1 in their pathogenesis has yet to be explored. The acquisition of new sequence information in related fungi will enhance our ability to gain insight on structure-function analysis of Cbp1.
One of the highlights of the current study is the exploration of the role of CHOP in Hc infection, since the role of ISR signaling during in vivo infection with microbial pathogens is relatively unexplored. To our knowledge, only one study has previously determined the susceptibility of CHOP-deficient mice to bacterial infection: CHOP-/- mice show reduced mortality in a polymicrobial sepsis model, and this increase in survival correlates with a decrease in bacterial burden [54]. Our data are the first to show that CHOP plays a role in host susceptibility to fungi, since we found that CHOP-/- mice are resistant to Hc infection. Intriguingly, Hc was unable to achieve a robust increase in fungal burden in the absence of CHOP signaling in the host. This decreased fungal burden manifested in the lungs of CHOP-/- mice by 3 dpi and in the spleen by 5 dpi. Importantly, the relatively early failure of Hc to thrive in the CHOP-/- mice is consistent with an altered interaction between the fungus and innate immune cells such as alveolar and tissue macrophages. Based on our observations in cell culture and the reduced apoptosis in alveolar macrophages in infected CHOP-/- mice, we speculate that Hc yeast are unable to robustly lyse macrophages in the CHOP-/- mouse, thereby limiting the spread of the pathogen and the extent of fungal growth.
In sum, we propose the following model: during Hc infection, intracellular yeast produce the secreted protein Cbp1. As the fungus replicates, levels of Cbp1 continue to rise, ultimately resulting in ISR activation and an increase in the levels of CHOP. CHOP in turn induces expression of TRIB3, which triggers a decrease in phospho-Akt levels. This decrease in phospho-Akt leads to an increase in the activity of several pro-apoptotic BH3-only proteins [55], which promote Bax/Bak oligomerization in the outer mitochondrial membrane. As we showed previously, Bax/Bak are required for optimal susceptibility of macrophages to Hc-mediated killing [6]. Bax/Bak oligomerization, in addition to the increase in caspase-8 activity stimulated by Hc infection, leads to the activation of the executioner caspases-3/7, ultimately resulting in apoptosis of the infected cell and the release of live yeast that are able to colonize new host cells in subsequent rounds of infection (Fig 9). One intriguing facet of this model is that Cbp1 acts as a molecular timer of host-cell lysis: the ISR is triggered only when a threshold level of Cbp1 is attained, thus allowing sufficient fungal replication to occur before the demise of the host cell.
Finally, the discovery that Cbp1 activates the ISR implies that Cbp1 represents a new category of Hc virulence factors. Of the handful of known Hc virulence factors [56], the others identified thus far promote pathogenesis by enabling Hc yeast to survive in the potentially hostile environment of the host. Examples include a siderophore used for iron acquisition within the classically iron-limited environment of the host cell [57], as well as catalases [58] and a secreted superoxide dismutase [59] necessary for combatting reactive oxygen species. Based on our observations, we propose that one function of Cbp1 is to promote pathogenesis by activating a host signaling cascade that ultimately results in macrophage death, allowing for escape of live Hc yeast from host cells and efficient spread of the pathogen throughout the host organism.
Bone marrow-derived macrophages (BMDMs) from 6 to 8 week old female mice were isolated from femurs and tibias and maintained in BMM (bone marrow macrophage media) as described previously [57]. J774.1 cells (ATCC) were maintained in Dulbecco’s modified Eagle’s medium (DMEM) high glucose (UCSF Cell Culture Facility) with 10% Fetal Bovine Serum (FBS; Hyclone or Gibco), penicillin and streptomycin (UCSF Cell Culture Facility). U937 cells (ATCC) were maintained in RPMI medium 1640 (Gibco) with 10% heat-inactivated FBS (Hyclone or Gibco), 2 mM glutamine (UCSF Cell Culture Facility), 110 μg/mL sodium pyruvate (UCSF Cell Culture Facility), penicillin (UCSF Cell Culture Facility), and streptomycin (UCSF Cell Culture Facility). U937 cells were differentiated with 100 nM phorbol 12-myristate 13-acetate (PMA; UCSF Cell Culture Facility). H. capsulatum cultures were grown in liquid Histoplasma macrophage medium (HMM) using an orbital shaker or on HMM agarose plates [60]. All cells were maintained at 37°C with 5% CO2. GSK2606414 (PERK inhibitor I; EMD Millipore) and tunicamycin (Santa Cruz Biotechnology) were dissolved in DMSO.
H. capsulatum strain G217B ura5Δ (WU15) was a kind gift from William Goldman (University of North Carolina, Chapel Hill). For all studies involving the cbp1 mutant, “wildtype” refers to G217 ura5Δ transformed with a URA5-containing episomal vector (pLH211), cbp1 refers to G217Bura5Δcbp1::T-DNA as previously described [6] transformed with the same URA5-containing episomal vector, and “complemented” strain refers to G217Bura5Δcbp1::T-DNA transformed with the URA5-containing episomal plasmid bearing the wild-type CBP1 gene (pDTI22) as previously described [6].
Primers for alanine scanning mutagenesis of CBP1 were designed using PrimerX (http://www.bioinformatics.org/primerx/). Primer sequences are included in supplemental material (S2 Table). Mutations were introduced into pDTI22 [6] with site-directed mutagenesis using standard cloning techniques. The fusion gene consisting of the sequence encoding the G217B Cbp1 signal peptide and the sequence encoding the mature Cbp1 from P. brasiliensis strain Pb03 was synthesized by Genewiz, Inc. The fusion gene was then cloned into pDTI22, replacing G217B CBP1 while leaving the flanking sequences unchanged.
For all Cbp1 constructs, approximately 25ng of PacI-linearized DNA was electroporated into the cbp1 mutant and the G217Bura5Δ parental strain [6] as previously described [61]. The plasmid pLH211 was used as the vector control. Transformants were selected on HMM agarose plates.
The genomic sequences of known CBP1 homologs are consistent with a conserved two-intron gene structure, but the associated protein predictions are often inconsistent with this conserved splicing pattern (e.g., Emmonsia crescens KKZ65414.1) or protein predictions are not available (e.g., Histoplasma capsulatum Tmu GCA_000313325.1). Therefore, protein sequences were inferred by TBLASTN [62] alignment of the Histoplasma capsulatum G186AR protein sequence (AAC39354.1) to the genome assemblies of Histoplasma capsulatum G217B (GCA_000170615.1), WU24 (NAm1, GCA_000149585.1), H88 (GCA_000151005.2), H143 (GCA_000151035.1), and Tmu (GCA_000313325.1); Paracoccidioides lutzii Pb01 (GCA_000150705.2); Paracoccidioides brasiliensis Pb03 (GCA_000150475.2) and Pb18 (GCA_000150735.2); and Emmonsia crescens UAMH 3008 (GCA_001008285.1). The inferred protein sequences were aligned with PROBCONS [63], and alignments were formatted with JALVIEW [64].
Four-day Hc cultures were collected and yeast were pelleted by centrifugation. The supernatants were sterile-filtered through 0.22μm filters, and the filtrates were concentrated using Amicon Ultra Centrifugal Filter Units with a 3 kDa cutoff (EMD Millipore). Protein concentration was quantified using the Bio-Rad protein assay (Bio-Rad Laboratories). Equal amounts of protein were separated by SDS-PAGE, and proteins were visualized by staining the gel with Coomassie Brilliant Blue G-250 (Fisher BioReagents).
Macrophage infections with Hc strains in the G217B background were performed as described previously [6,57]. Briefly, the day before infection, macrophages were seeded in tissue culture-treated dishes. On the day of infection, yeast cells from logarithmic-phase Hc cultures (OD600 = 5–7) were collected, resuspended in the appropriate macrophage media, sonicated for 3 seconds on setting 2 using a Fisher Scientific Sonic Dismembrator Model 100, and counted using a hemacytometer. Infection with G186AR required a slightly different protocol due to the tendency of this strain to form large aggregates of cells. G186AR yeast cells from a logarithmic-phase culture were sonicated for 30 1-second pulses, vortexed for 30 seconds, sonicated for 30 1-second pulses, and then vortexed for another 30 seconds to disrupt clumps of yeast cells. The cells were then pelleted, resuspended in the appropriate media, sonicated for 30 1-second pulses, and vortexed for 30 seconds. The cells were then spun at low speed (400 rpm) to pellet large yeast aggregates. The remaining cells in the supernatant, which were single cells or very small aggregates of two or three yeast, were then counted using a hemacytometer. Depending on the multiplicity of infection (MOI), the appropriate number of yeast cells was then added to the macrophages. After a 2-hour phagocytosis period, the macrophages were washed once with PBS and then fresh media was added. For infections lasting longer than 2 days, fresh media was added to the cells approximately 48 hpi.
U937 cells were differentiated with 100 nM PMA three days prior to infection. One day prior to infection, adherent cells were collected by scraping, counted with a hemacytometer, and seeded at a density of 4x105 cells per well in a 24-well plate. The following day, the cells were infected in quadruplicate as described above, although the phagocytosis period was extended to 12 hours to accommodate the reduced phagocytosis rate of differentiated U937 cells for Hc cells.
To screen strains expressing mutant alleles of CBP1, J774.1 cells were seeded (3.75 x 104 cells per well of a 24-well plate) and infected as described above in triplicate wells per time point. Three independent transformants per alanine mutant were tested. Each day for 4 days after infection, macrophage monolayers were washed once with PBS, fixed and stained with methylene blue staining solution (0.2% methylene blue [Sigma Aldrich], 20% ethanol) at room temperature for 15 min, washed 3 times with PBS, and then imaged.
To quantify macrophage lysis, BMDMs were seeded (7.5 x 104 cells per well of a 48-well plate) and infected as described above. At the indicated time points, the amount of LDH in the supernatant was measured as described previously [65]. BMDM lysis is calculated as the percentage of total LDH from uninfected macrophages lysed in 1% Triton-X at the time of infection. Due to continued replication of BMDMs over the course of the experiment, the total LDH at later time points is greater than the total LDH from the initial time point, resulting in an apparent lysis that is greater than 100%.
BMDMs were seeded (7.5 x 104 cells per well of a 48-well plate) and infected in triplicate as described above. At the indicated time points, culture supernatants were removed and 250 μl of ddH2O was added. After incubating at room temperature for 10 min, the macrophages were mechanically lysed by vigorous pipetting. The lysate was collected, sonicated to disperse any clumps, counted, and plated on HMM agarose in appropriate dilutions. After incubation at 37°C with 5% CO2 for 12–14 days, CFUs were enumerated. To prevent any extracellular replication from confounding the results, intracellular replication was not monitored after the onset of macrophage lysis.
The day before infection, 1x107 wildtype or CHOP-/- BMDMs were seeded in 15 cm tissue culture dishes. The following day, the cells were infected with wildtype Hc at an MOI of 2 or mock infected as described above. The same day, wildtype BMDMs were also seeded in 6-well plates at a density of 7.5x105 cells per well. Approximately 24 hours after infection, the wildtype and CHOP-/- BMDMs from the 15 cm plates were collected by scraping, and viable cells, as determined by trypan blue exclusion, were counted using a hemacytometer. These BMDMs were then seeded on the underside of 0.4 μm polyester transwells (Costar) at a density of 4x105 cells per transwell in 1 mL BMM. After 2 hours, excess liquid was aspirated, and the transwells were inverted over the wildtype BMDMs in 6-well plates. Approximately 18 hours later, when the infected transwell BMDMs had begun to lyse but the lower macrophage monolayer was still intact, the transwells were removed, and the lower BMDMs were washed once with PBS and fresh BMM was added. CFUs and cytoxicity were measured as described above.
For RNA isolation from cultured cells, BMDMs were seeded (1x 106 cells per well of a 6-well plate) and infected in triplicate as described. Triplicate wells of infected macrophages were lysed in 1 mL total of QIAzol (Qiagen). U937 cells were seeded at 4x105 cells per well of a 24-well plate and infected in quadruplicate as described. Quadruplicate wells were lysed in 1 mL total of QIAzol (Qiagen). Lung samples were thawed on ice, and 500 μL of homogenate was pelleted at 4°C and resuspended in 1 mL of QIAzol (Qiagen). After addition of chloroform, total RNA was isolated from the aqueous phase using Econo-spin columns (Epoch Life Science) and then subjected to on-column PureLink DNase (Invitrogen) digestion. To generate cDNA, 2–4 μg total RNA was reverse transcribed using Maxima Reverse Transcriptase (Thermo Scientific), an oligo-dT primer, and pdN9 primers following manufacturer’s instructions. Splice isoforms of Xbp1 were detected by performing non-quantitative PCR on 1:10 dilutions of cDNA using Phusion High-Fidelity DNA polymerase (New England BioLabs) with the following cycling conditions: 98°C for 30s, followed by 35 cycles of 98°C (10 s), 65°C (30 s), and 72°C (30 s), followed by 72°C for 10 min. The resulting amplicons were separated and visualized on a 2.5% agarose gel containing ethidium bromide. Quantitative PCR was performed on 1:10 to 1:50 dilutions of cDNA template using FastStart SYBR Green MasterMix with Rox (Roche). Reactions were run on an Mx3000P machine (Stratagene) and analyzed using MxPro software (Stratagene). Cycling parameters were as follows: 95°C for 10 min, followed by 40 cycles of 95°C (30 s), 55°C (60 s), and 72°C (30 s), followed by dissociation curve analysis. Abundances of ERdj4, SEL1L, BiP, CHOP, and TRIB3 were normalized to HPRT levels. Primer sequences are listed in S2 Table.
For protein isolation, BMDMs were seeded (1x 106 cells per well of a 6-well plate) and infected in triplicate as describe. Triplicate wells of macrophages were lysed in a 300 μl total of radioimmunoprecipitation assay (RIPA) buffer (50 mM TrisHCl pH8, 150 mM NaCl, 1% NP-40, 0.5% sodium deoxycholate, 0.1% SDS) with Halt Protease and Phosphatase Inhibitor Cocktail (Thermo Scientific). Insoluble debris was removed by centrifugation. Protein concentrations were determined using Pierce BCA Protein Assay Kit (Thermo Scientific). Equivalent amounts of protein were separated by SDS-PAGE and transferred to nitrocellulose. Membranes were incubated with antibodies per manufacturer’s suggestions. Blots were imaged on an Odyssey CLx and analyzed using ImageStudio2.1 (Licor). The following primary antibodies were used: phospho-PERK (ThermoFisher Scientific G.305.4), phospho-eIF2α (Cell Signaling Technology 9721), α-tubulin (Santa Cruz Biotechnology YOL1/24; sc-53030), ATF4 (Cell Signaling Technology D4B8; 11815), CHOP (Cell Signaling Technology L633F7; 2895), TRIB3 (Calbiochem ST1032), phospho-Akt (Cell Signaling Technology C313E5E; 2965), Akt (Cell Signaling Technology 40D4; 2920). Images were processed with Adobe Photoshop, which was utilized on occasion to change the order of lanes in the image to group like samples (e.g. lytic strains or non-lytic strains) together.
BMDMs were seeded at 2.5 x 104 cells per well in solid white 96-well plates and infected as described above. Caspase-3/7 and caspase-8 activities were measured using Caspase-Glo 3/7 Assay and Caspase-Glo 8 Assay, respectively, according to manufacturer’s instructions (Promega). GSK2606414 (EMD Millipore) was dissolved in DMSO and added to macrophages after the 2-hour phagocytosis period at a final concentration of 3 μM where indicated. Luminescence was measured on a Wallac Victor2 microplate reader (PerkinElmer).
Eight-to-twelve week-old female C57Bl/6 (Jackson Laboratory stock 000664) or CHOP-/- (B6.129S(Cg)-Ddit3tm2.1Dron/J; Jackson Laboratory stock 005530) mice were anesthetized with isoflurane and infected intranasally with wildtype Hc yeast. The inoculum was prepared by collecting mid-logarithmic phase (OD600 = 5–7) yeast cultures, washing once with PBS, sonicating for 3 seconds on setting 2 using a Fisher Scientific Sonic Dismembrator Model 100, counting with a hemacytometer, and diluting in PBS so that the final inoculum was approximately 25 μl. To monitor survival, animals were infected with 1x106 yeast per mouse. To monitor in vivo colonization, animals were infected with 3x105 yeast per mouse. Infected mice were monitored daily for symptoms of disease, including weight loss, hunching, panting, and lack of grooming. For survival curve analysis, mice were euthanized after they exhibited 3 days of sustained weight loss greater than 25% of their maximum weight in addition to one other symptom. For in vivo colonization, five mice per genotype were sacrificed at the indicated time points, and lungs and spleens were harvested. The organs were homogenized in PBS and plated on brain heart infusion (BHI) agar plates supplemented with 10% sheep’s blood (Colorado Serum Company). After plating, the remaining organ homogenate was snap frozen in liquid nitrogen and stored at -80°C. CFUs were enumerated after 10–12 days of growth at 30°C.
For flow cytometry analysis, mice were sacrificed with avertin, and lungs were perfused with PBS and then removed. Lungs were then dissociated in Hanks Buffered Salt Solution (HBSS) containing 0.5 mg DNase I (Roche) per mouse, and 0.75 U Collagenase P (Roche) per mouse using a GentleMACS Tissue Dissociator (Milteny Biotec). Red blood cells were hypotonically lysed with ACK lysis buffer and the remaining cells were filtered through a 70 μM cell strainer (BD Biosciences.) 2x106 lung cells were resuspended in FACS buffer (PBS containing 1% heat-inactivated FBS, 1 mM EDTA, 10 μg/mL CD16/32, and 0.1% sodium azide) and stained with CellEvent Caspase-3/7 Green Detection Reagent (Invitrogen) according to manufacturer’s instructions. The cells were then washed with FACS buffer and stained with Fixable Viability Dye eFluor 450 (eBiosciences) for 20 minutes. Cells were then incubated with CD16/32 (BioLegend) for 20 minutes, then stained for 30 minutes with the appropriate antibodies, fixed in BD Stabilizing Fix (BD Biosciences), and stored at 4°C until analysis on an LSR II (BD Biosciences). Antibodies used to identify alveolar macrophages were as follows: BV650-CD45 (BioLegend), PE-SiglecF (BD Biosciences), BV605-CD11b (BioLegend), Alexa700-CD11c (BioLegend), and BV711-F4/80 (BioLegend). Flow cytometry data were analyzed using FlowJo version 10. Apoptotic cells were defined as Fixable Viability Dye eFluor 450+ and Caspase-3/7 Green Detection Reagent+ double-positive cells.
Two-tailed t-tests were performed using Excel (Microsoft). Two-tailed Mann-Whitney tests and statistical analysis for mouse survival and colonization experiments were performed using Prism (GraphPad Software) as described in the figure legend.
All mouse experiments were performed in compliance with the National Institutes of Health Guide for the Care and Use of Laboratory Animals and were approved by the Institutional Animal Care and Use Committee at the University of California San Francisco (protocol AN155431-01). Mice were euthanized by CO2 narcosis and cervical dislocation consistent with American Veterinary Medical Association guidelines.
|
10.1371/journal.pcbi.1002965 | The Fidelity of Dynamic Signaling by Noisy Biomolecular Networks | Cells live in changing, dynamic environments. To understand cellular decision-making, we must therefore understand how fluctuating inputs are processed by noisy biomolecular networks. Here we present a general methodology for analyzing the fidelity with which different statistics of a fluctuating input are represented, or encoded, in the output of a signaling system over time. We identify two orthogonal sources of error that corrupt perfect representation of the signal: dynamical error, which occurs when the network responds on average to other features of the input trajectory as well as to the signal of interest, and mechanistic error, which occurs because biochemical reactions comprising the signaling mechanism are stochastic. Trade-offs between these two errors can determine the system's fidelity. By developing mathematical approaches to derive dynamics conditional on input trajectories we can show, for example, that increased biochemical noise (mechanistic error) can improve fidelity and that both negative and positive feedback degrade fidelity, for standard models of genetic autoregulation. For a group of cells, the fidelity of the collective output exceeds that of an individual cell and negative feedback then typically becomes beneficial. We can also predict the dynamic signal for which a given system has highest fidelity and, conversely, how to modify the network design to maximize fidelity for a given dynamic signal. Our approach is general, has applications to both systems and synthetic biology, and will help underpin studies of cellular behavior in natural, dynamic environments.
| Cells do not live in constant conditions, but in environments that change over time. To adapt to their surroundings, cells must therefore sense fluctuating concentrations and ‘interpret’ the state of their environment to see whether, for example, a change in the pattern of gene expression is needed. This task is achieved via the noisy computations of biomolecular networks. But what levels of signaling fidelity can be achieved and how are dynamic signals encoded in the network's outputs? Here we present a general technique for analyzing such questions. We identify two sources of signaling error: dynamic error, which occurs when the network responds to features of the input other than the signal of interest; and mechanistic error, which arises because of the inevitable stochasticity of biochemical reactions. We show analytically that increased biochemical noise can sometimes improve fidelity and that, for genetic autoregulation, feedback can be deleterious. Our approach also allows us to predict the dynamic signal for which a given signaling network has highest fidelity and to design networks to maximize fidelity for a given signal. We thus propose a new way to analyze the flow of information in signaling networks, particularly for the dynamic environments expected in nature.
| Cells are continuously challenged by extra- and intracellular fluctuations, or ‘noise’, [1]–[3]. We are only starting to unravel how fluctuating inputs and dynamic interactions with other stochastic, intracellular systems affect the behavior of biomolecular networks [4]–[9]. Such knowledge is, however, essential for studying the fidelity of signal transduction [10], [11] and therefore for understanding and controlling cellular decision-making [12]. Indeed, successful synthetic biology requires quantitative predictions of the effects of fluctuations at the single-cell level, both in static and dynamic environments [13]. Furthermore, sophisticated responses to signals that change over time are needed for therapeutics that involve targeted delivery of molecules by microbes [14], [15] or the reprogramming of immune cells [16]. Here we begin to address these challenges by developing a general framework for analysing the fidelity with which dynamic signals are represented by, or ‘encoded’ in, the output of noisy biomolecular networks.
For cellular signaling to be effective, it should maintain sufficient fidelity. We wish to quantify the extent to which the current output of an intracellular biochemical network, , can represent a particular feature of a fluctuating input (Fig. 1). This signal of interest, , is generally a function of the history of the input, denoted . By its history, we mean the value of the input at time and at all previous times. The signal could be, for example, the level of the input at time or a time average of the input over a time window in the most recent past. The output of the signaling network, , is able to perfectly represent the signal if can be inferred exactly from at all times, . The system then has zero fidelity error. However, for a stochastic biochemical mechanism, a given value of will map to multiple possible values of the output, .
We will assume that the conditional mean, , is an invertible function of : it takes different values for any two values of . It is then a perfect representation of . The output will, however, usually be different from and have a fidelity error, defined as the difference between and . The notation is read as conditioned on, or given, the value of the variable at time . We use , as for example in , to denote averaging over all random variables except those given in the conditioning. Therefore is itself a random variable: it is a function of the random variable (we give a summary of the properties of conditional expectations in the SI).
Many response functions, , in biochemistry and physiology (for example, Hill functions) satisfy the requirement of invertibility or can be made to do so by defining appropriately—for example, when a response exactly saturates for all input values above a threshold, those values can be grouped to form a single input state. Furthermore, we know from the properties of conditional expectations that is closer to in terms of mean squared fidelity error than to any other representation (function) of (SI).
The difference between the conditional expectations and, for example, is important. The former, , is the average value of the output at time given a particular history of the input . It will often coincide with the deterministic (macroscopic) solution when the same input trajectory is applied to the network. The output shows random variation around this average, , for identical realisations of the trajectory of . By contrast, is the average value of given that the trajectory of up to time ends at the value . By the properties of conditional expectations, this is also the average value of over all trajectories ending in the value : that is, . These mathematical definitions are illustrated diagrammatically in Fig. 2.
We distinguish between two types of error that reduce fidelity between and .
We can decompose the output into the sum of the faithfully transformed or transmitted signal, , the dynamical error, and the mechanistic error:(5)for all times . Eq. 5 is an orthogonal decomposition of the random variable —each pair of random variables on the right-hand side has zero correlation (Methods). The variance of therefore satisfies(6)where the magnitude of the fidelity error is given by , which is because of the orthogonality. This magnitude of the fidelity error is also equal to the expected conditional variance of the output, . We note that we can generalize this decomposition, and thus extend our approach, for example, to study different components of the mechanistic error (Methods).
To compare signaling by different biochemical mechanisms, we normalize by the square root of its variance, writing , and define the fidelity as a signal-to-noise ratio:(7)for some signal of interest, . Eq. 7 is dimensionless and a montonically decreasing function of . Indeed, we have shown that the maximal mutual information between and across all possible signal distributions is bounded below by a decreasing function of (and so an increasing function of our fidelity), for a suitable choice of distribution of the signal and when is an invertible function of [7].
Comparing biochemical systems using the fidelity measure is equivalent to comparison based on the magnitude of the fidelity error, , where and the error is measured in units of the standard deviation of the output. Eq. 7 is maximized when is minimized. One minus the magnitude of the fidelity error is the fraction of the variance in the output that is generated by the signal of interest. In information theoretic approaches, normalizing the output by its standard deviation is also important, because the normalization allows determination of the number of ‘unique’ levels of output that can be distinguished from one other despite the stochasticity of the output, as least for Gaussian fluctuations [18].
When and have a bivariate Gaussian distribution, the instantaneous mutual information, , is monotonically related to the fidelity and exactly equal to [7], where denotes the correlation coefficient. Also in this Gaussian case, is equal to the minimum mean squared error (normalised by ) between and the linear, optimal estimate, . (This is the optimal ‘filter’ when only the current output is available, although typically a filter such as the Wiener filter would employ the entire history of up to time .) Gaussian models of this sort for biochemical signalling motifs were considered in [19], with instantaneous mutual information expressed in terms of a signal-to-noise ratio equivalent (for their models) to the fidelity of Eq. 7. Such Gaussian models (if taken literally, rather than used to provide a lower bound on the information capacity [19]) would imply that the input-output relation, , is linear and that does not depend on (by the properties of the multivariate normal distribution). Our approach requires neither assumption.
Whenever is a linear function of , that is for constants and , we consider to be the gain for the signal of interest [19]. The fidelity then depends on the ratio of the squared gain to the fidelity error and is given by .
Methods of analysis of stochastic systems with dynamic inputs are still being developed. We argue that deriving expectations of network components conditional upon the histories of stochastic inputs is a powerful approach. We have developed three methods to determine components of Eqs. 5 and 6 (SI):
We note that our methods require that the inputs can be modeled as exogenous processes that are unaffected by interactions with the biochemistry of the signaling network (a distinction emphasised in [20]). By an exogenous process we mean one whose future trajectory is independent, given its own history, of the history of the biochemical system. This model for an input is reasonable, for example, when the input is the level of a regulatory molecule, such as a transcription factor, that has relatively few binding sites in the cell.
Transcriptional regulation is a primary means by which cells alter gene expression in response to signals [21]. We now provide an exact, in-depth analysis of a two-stage model of gene expression [22] where the fluctuating input, , is the rate (or propensity) of transcription and the signal of interest, , equals the current value of the input, . For example, may be proportional to the extracellular level of a nutrient or the cytosolic level of a hormone regulating a nuclear hormone receptor.
The cellular response should account for not only the current biological state of but also future fluctuations. If we consider an input that is a Markov process, future fluctuations depend solely on the current value , and the cell would need only to ‘track’ the current state as effectively as possible and then use the representation in protein levels to control downstream effectors. These ideas are related to those underlying predictive information [23], [24].
Our analysis requires only the stationary mean and variance of the input and that has exponentially declining ‘memory’ (SI). Consequently, the autocorrelation function of is a single exponential with autocorrelation time (the lifetime of fluctuations in ). Examples include a birth-death process or a two-state Markov chain. We can generalize using, for example, weighted sums of exponentials to flexibly model the autocorrelation function of .
Solving the ‘conditional’ master equation with a time-varying rate of transcription, we find that the conditionally expected protein level is a double weighted ‘sum’ of past levels of the signal (SI):(9)(where for simplicity the equation is stated for the case of zero initial mRNA and protein). We denote the rate of translation per molecule of mRNA by , the rate of mRNA degradation per molecule by , and the rate of degradation of protein per molecule by . The most recent history of the input exerts the greatest impact on the current expected output, with the memory of protein levels for the history of the input determined by the lifetimes of mRNA and protein molecules. Eq. 9 gives the signal of interest, (a function of the history of the fluctuating transcription rate), that gene expression transmits with the highest fidelity to protein levels (see Eq. 8). Notice that the current value of the input, , cannot be recovered exactly from , which is therefore not a perfect representation of .
We find, by contrast, that is an invertible, linear function of :(10)when the dynamics reach stationarity, and that the stationary unconditional mean is (SI). Notice that does not converge for large to the average ‘steady-state’ solution for a static , but depends on . The discrepancy between Eqs. 9 and 10 results in dynamical error with non-zero magnitude (Fig. 3B).
Using our solutions for the conditional moments, we can calculate the variance components of Eq. 6 (SI). For the faithfully transformed signal, when , we have(11)where is the ratio of the lifetime of mRNA to the lifetime of fluctuations in , and is the ratio of the lifetime of protein to the lifetime of fluctuations in . The magnitude of the dynamical error is in this case proportional to Eq. 11(12)and the magnitude of the mechanistic error satisfies(13)When the autocorrelation time of becomes large ( and tending to zero), the dynamical error therefore vanishes (Eq. 12). In this limit, the output effectively experiences a constant input during the time ‘remembered’ by the system.
To gain intuition about the the effect of relative lifetimes on the fidelity of signaling, we first suppose the mechanistic error is small relative to . Eq. 7 then becomes simply if protein lifetime is large relative to mRNA lifetime, (as expected for many genes in budding yeast [25]). The fidelity thus improves as the protein lifetime decreases relative to the lifetime of fluctuations in , and the output is able to follow more short-lived fluctuations in the signal. This observation is only true, however, for negligible mechanistic error.
It is the aggregate behavior of dynamical and mechanistic errors as a fraction of the total variance of the output that determines signaling fidelity, Eq. 7. Effective network designs must sometimes balance trade-offs between the two types of error.
Our framework naturally adapts to the scenario of controlling a network output to approach a desired ‘target’ response when, for example, the cell's environment changes. Combined with model search procedures for synthetic design [32], it is a promising approach to the design of synthetic biomolecular networks. If the target response is given by , which is a function of the input history, then to guide the design process, we can decompose the error analogously to Eq. 5 and find an equivalent to Eq. 6, a dissection of the network performance into orthogonal components (SI).
Cells use the information conveyed by signaling networks to regulate their behavior and make decisions. Not all features of the input trajectory will, however, be relevant for a particular decision, and we define the fidelity between the output of the network and a signal of interest, , which is a function of the input trajectory. Information encoded in upstream fluctuations must eventually either be lost or encoded in current levels of cellular constituents. We have therefore focused on the fidelity with which is represented by the current output, .
Using an orthogonal decomposition of the network's output into the faithfully transformed signal and error terms, we are able to identify two sources of error – dynamical and mechanistic. We assume the transformed signal, , to be an invertible function of . The aggregate behavior of the two types of error determines the signaling fidelity, and ignoring either may cause erroneous conclusions. We interpret as the current cellular estimate or ‘readout’ of the faithfully transformed signal. The magnitude of the fidelity error relative to the variance in , Eq. 7, is a dimensionless measure of the quality of that estimate since . Furthermore, we have shown that is related to the mutual information between the input and output [7].
To apply our approach experimentally, we can use microfluidic technology to expose cells to the same controlled but time-varying input in the medium [33], and a fluorescent reporter to monitor the network output, . This reporter could measure, for example, a level of gene expression or the extent of translocation of a transcription factor. The transformed signal, , and its variance (for a given probability distribution of the input process) can then be estimated with sufficient amounts of data by monitoring in each cell and in the microfluidic medium. We can determine the mechanistic error by measuring the average squared difference between the output of one cell and that of another — because the outputs of two cells are conjugate given the history of the input [7] –and hence determine the dynamical error by applying Eq. 6.
Our analysis is complementary to one based on information theory and the entire distribution of input and output [7]. Without making strong assumptions about the network and the input, calculation of mutual information is challenging for dynamic inputs. Previous work has considered either the mutual information between entire input and output trajectories with a Gaussian joint distribution of input and output [19], [34], or the ‘instantaneous’ mutual information between input and output at time [19] (applicable in principle to non-Gaussian settings). Our approach, however, depends only on conditional moments and avoids the need to fully specify the distribution of the input process, which is often poorly characterized.
The environments in which cells live are inherently dynamic and noisy. Here we have developed mathematical techniques to quantify how cells interpret and respond to fluctuating signals given their stochastic biochemistry. Our approach is general and will help underpin studies of cellular behavior in natural, dynamic environments.
Define , the transformed signal with zero mean. Then the signal and error components of Eq. 5 are pairwise uncorrelated:(14)
Eq. 5 is a special case of the following general decomposition for any random variable (with finite expectation), here denoted . Consider a filtration, or increasing sequence of conditioning ‘information sets’, , where and . Let for , and let . Then the decomposition(15)satisfies for all since the sequence is a martingale difference sequence with respect to the filtration (SI). Therefore, .
|
10.1371/journal.pmed.1002659 | Community delivery of antiretroviral drugs: A non-inferiority cluster-randomized pragmatic trial in Dar es Salaam, Tanzania | With the increase in people living with HIV in sub-Saharan Africa and expanding eligibility criteria for antiretroviral therapy (ART), there is intense interest in the use of novel delivery models that allow understaffed health systems to successfully deal with an increasing demand for antiretroviral drugs (ARVs). This pragmatic randomized controlled trial in Dar es Salaam, Tanzania, evaluated a novel model of ARV community delivery: lay health workers (home-based carers [HBCs]) deliver ARVs to the homes of patients who are clinically stable on ART, while nurses and physicians deliver standard facility-based care for patients who are clinically unstable. Specifically, the trial aimed to assess whether the ARV community delivery model performed at least equally well in averting virological failure as the standard of care (facility-based care for all ART patients).
The study took place from March 1, 2016, to October 27, 2017. All (48) healthcare facilities in Dar es Salaam that provided ART and had an affiliated team of public-sector HBCs were randomized 1:1 to either (i) ARV community delivery (intervention) or (ii) the standard of care (control). Our prespecified primary endpoint was the proportion of adult non-pregnant ART patients with virological failure at the end of the study period. The prespecified margin of non-inferiority was a risk ratio (RR) of 1.45. The mean follow-up period was 326 days. We obtained intent-to-treat (ITT) RRs using a log-binomial model adjusting standard errors for clustering at the level of the healthcare facility. A total of 2,172 patients were enrolled at intervention (1,163 patients) and control (1,009 patients) facilities. Of the 1,163 patients in the intervention arm, 516 (44.4%) were both clinically stable on ART and opted to receive ARVs in their homes or at another meeting point of their choosing in the community. At the end of the study period, 10.9% (95/872) of patients in the control arm and 9.7% (91/943) in the intervention arm were failing virologically. The ITT RR for virological failure demonstrated non-inferiority of the ARV community delivery model (RR 0.89 [1-sided 95% CI 0.00–1.18]). We observed no significant difference between study arms in self-reported patient healthcare expenditures over the last 6 months before study exit. Of those who received ARVs in the community, 97.2% (95% CI 94.7%–98.7%) reported being either “satisfied” or “very satisfied” with the program. Other than loss to follow-up (18.9% in the intervention and 13.6% in the control arm), the main limitation of this trial was that substantial decongestion of healthcare facilities was not achieved, thus making the logic for our preregistered ITT approach (which includes those ineligible to receive ARVs at home in the intervention sample) less compelling.
In this study, an ARV community delivery model performed at least as well as the standard of care regarding the critical health indicator of virological failure. The intervention did not significantly reduce patient healthcare expenditures, but satisfaction with the program was high and it is likely to save patients time. Policy-makers should consider piloting, evaluating, and scaling more ambitious ARV community delivery programs that can reach higher proportions of ART patients.
ClinicalTrials.gov NCT02711293.
| The number of individuals in sub-Saharan Africa needing antiretroviral therapy (ART) for HIV—and chronic disease care more broadly—is expected to increase over the coming decades, further straining already under-resourced health systems in the region.
Community delivery of antiretroviral drugs (ARVs) has the potential to decrease ART patient volume at healthcare facilities and reduce patient and government healthcare expenditures, while maintaining the positive effects of ART on health outcomes.
We randomized 48 ART facilities in Dar es Salaam, Tanzania, 1:1 to (i) ARV community delivery (lay health workers deliver ARVs to the homes of patients who are clinically stable on ART while nurses and physicians deliver standard facility-based care for patients who are clinically unstable on ART) or (ii) control (standard facility-based care for all patients).
We measured whether the risk of virological failure (i.e., poor control of one’s HIV infection) in the ARV community delivery arm was lower than, or equal to that in the control arm using a prespecified threshold above which we considered the intervention to be inferior to the standard of care.
As implemented in this study, in which roughly 40% of patients enrolled in the intervention arm received ARVs at home at least once and 60% remained in standard facility-based care, the new ARV community delivery model performed at least as well as the standard of care regarding the critical health indicator of virological failure.
While we did not observe any differences in patient healthcare expenditures between the 2 arms of the study, patients’ satisfaction with receiving ARVs at home was high.
While ARV community delivery appears to have been safe with respect to controlling patients’ HIV infection, some of the expected benefits of the program—decongestion of healthcare facilities and reduction in patients’ healthcare expenditures—were not realized.
Nonetheless, the program was popular with patients, presumably because it makes ART care more convenient and saves patients time.
Policy-makers should consider piloting, scaling, and evaluating more ambitious ARV community delivery programs that reach higher proportions of ART patients than reached in this trial.
| Chronic diseases are rapidly replacing acute infectious diseases as the leading cause of the disease burden in sub-Saharan Africa (SSA) [1]. While many of these chronic conditions are non-communicable, HIV has also become a chronic illness as effective therapy allows HIV-positive individuals to survive into old age [2,3]. Indeed, the HIV epidemic continues to take a considerable toll in SSA, where approximately 25.5 million people were living with HIV in 2016 [4]. This shift to chronic conditions poses new challenges to generally weak health systems in SSA.
HIV, in particular, is expected to place an increasing stress on health systems and patients in SSA in the coming years. The World Health Organization (WHO) eliminated all CD4 cell count treatment thresholds in its 2016 HIV treatment guidelines, recommending antiretroviral therapy (ART) for all people living with HIV [5]. As countries are gradually starting to implement ART for all, this will likely lead to a substantial rise in the number of people on ART in SSA in the coming years [6], particularly if guideline changes are coupled with an increased identification of individuals living with HIV who are currently unaware of their HIV status. Under former HIV care guidelines, HIV patients on ART generally attended facility-based care at least twice as frequently as HIV patients who were not yet on ART [7]. Given the high prevalence of HIV across SSA, the implementation of the new WHO treatment guidelines is thus expected to place further strain on nurses and physicians, as well as increase patient healthcare expenditures [8].
Community delivery of antiretroviral drugs (ARVs) through lay health workers is a promising ART delivery model, because it should increase the capacity of the health system to effectively provide care to rising ART patient numbers [5]. ARV community delivery may not only ease the workload of nurses and physicians (e.g., reducing the frequency with which ART patients have to attend healthcare facilities), but may also improve ART adherence and retention among patients (e.g., by increasing the convenience of remaining in care) [9]. In addition, it is plausible that ARV community delivery can reduce patient healthcare expenditures, because patients receive treatment in or close to their homes and thus save the time and financial costs of travel to healthcare facilities [10]. On the other hand, ARV community delivery may have adverse consequences for ART patients’ health by decreasing the frequency with which patients are seen by more highly trained healthcare workers.
This non-inferiority cluster-randomized pragmatic trial aimed to establish the effectiveness of ARV community delivery when implemented in the routine healthcare system of Dar es Salaam, the largest city in East Africa [11]. Specifically, this study aimed to determine whether an ARV community delivery model (lay health workers deliver ARVs to the homes of patients who are clinically stable on ART and nurses and physicians deliver standard facility-based care for patients who are clinically unstable on ART) leads to a lower or equal (“non-inferior”) risk of virological failure compared to the standard of care (standard facility-based care for all ART patients). A secondary aim of this study was to determine the impact of the ARV community delivery model on patient healthcare expenditures. This study was registered in the clinical trials registry of the United States National Library of Medicine at the National Institutes of Health, ClinicalTrials.gov (NCT02711293).
A detailed protocol of this study is available in the public domain [12].
This study took place in all 3 municipalities (Temeke, Kinondoni, and Ilala) of Dar es Salaam. Dar es Salaam is the most urbanized region of Tanzania, with an estimated 5.4 million inhabitants in 2016. The population of the city is expected to grow to 10.8 million by 2030 [11]. In 2012, when the latest HIV/AIDS and Malaria Indicator Survey was carried out in Tanzania, Dar es Salaam’s HIV prevalence was estimated to be 6.9% among adults aged 15–49 years, while the national prevalence was 5.1% [13]. The most recent HIV treatment guidelines in Tanzania were published in May 2015 (i.e., prior to trial start) and recommend initiation of ART for adults if the patient has a CD4 cell count < 500 cells/μl or is in WHO stage 3 or 4 [14].
Home-based carers (HBCs) are a lay health worker cadre in Tanzania’s public-sector health system. There are approximately 35,000 HBCs in Tanzania, and the program has existed since 1996. HBCs work in the neighborhoods in which they live; 1 to 3 HBCs serve one neighborhood. The HBC program exists in most, but not all, neighborhoods of Dar es Salaam. HBCs’ main responsibility consists of conducting regular (at least every 3 months) visits to HIV patients’ households in their assigned neighborhood. Their tasks have varied over the years but generally consist of household visits to provide counseling on ARV adherence, family planning, and nutrition; to promote the uptake of preventive healthcare services; and to refer ill clients to a healthcare facility. HBCs are affiliated with 1 healthcare facility in the vicinity of their neighborhood. A facility-based nurse (“community outreach nurse”) is responsible for supervising the healthcare facility’s team of HBCs. Dar es Salaam’s municipalities pay HBCs a monthly stipend of 50,000 Tanzanian Shillings (TZS) (72 purchasing power parity-adjusted dollars [PPP$]). As part of this trial, HBCs in the intervention arm received a further TZS 75,000 (PPP$ 109) flat payment per month to compensate them for the additional transport costs and workload. Because HBCs had a varying number of ART clients for home delivery of ARVs, this payment was changed to a payment of 10,000 Tanzanian Shillings (PPP$ 14) per community ARV delivery visit in January 2017.
In clusters randomized to ARV community delivery, patients who were clinically stable on ART could choose to receive ARVs and ART counseling in or close to their homes instead of having to return to the healthcare facility for a clinical checkup and to pick up their next ARV supplies. An HBC visited patients at home or at another meeting point in the community at which the patient wanted to receive his or her ARVs (such as somewhere close to the patient’s workplace). The HBC provided counseling, delivered a supply of ARVs, and performed an ARV pill count. HBC visits were conducted with the same frequency as the patient’s schedule for attending facility-based ART delivery prior to study enrollment, which was either a monthly or 2-monthly HIV care visit. Patients in the intervention arm did not have to return to a facility for HIV care until the study exit assessment. HBCs in the intervention arm received 3 days of training in the community delivery of ARVs and in counseling skills for this intervention prior to the start of the trial. Counseling focused on ART adherence, family planning, prevention of onward HIV transmission, and basic nutrition.
The eligibility criteria for enrollment into this trial were (i) age ≥ 18 years, (ii) having attended one of the participating healthcare facilities for ART delivery during the enrollment period, and (iii) residing in a neighborhood in the facility’s catchment area (ascertained through self-report). Exclusion criteria were pregnancy at the time of enrollment (ascertained through self-report) and inability to provide written informed consent. Pregnancy was an exclusion criterion because at many of the study’s healthcare facilities, pregnant women with HIV were seen in different sections of the healthcare facility than non-pregnant ART patients. Thus, enrollment of pregnant women would have required additional human resources, which were not available.
The random assignment to ARV community delivery versus standard facility-based care was at the facility level. After enrollment into the trial, patients who attended trial clinics in the control arm continued to receive the standard of care. In the ARV community delivery intervention arm, enrolled patients had to be clinically stable on ART to be eligible to receive ARVs in their homes or at other community meeting points. In consultation with Tanzania’s National AIDS Control Programme, the following criteria were established to define clinical stability on ART: (i) having taken ARVs for at least 6 months prior to study enrollment, (ii) having had a CD4 cell count >350 cells/μl or a suppressed viral load (VL) at 6 or more months after ART initiation (if both a CD4 cell count and VL measurement were taken 6 or more months after ART initiation, then at least 1 CD4 cell count had to be >350 cells/μl and 1 VL measurement had to show virological suppression [<1,000 copies/ml]), and (iii) the most recent VL was taken less than 12 months prior to study enrollment and showed virological suppression. For patients for whom a VL measurement was unavailable at the time of enrollment but for whom a CD4 cell count taken in the 12 months prior to enrollment was available, a CD4 cell count >350 cells/μl was used to replace criterion (iii). For patients for whom neither a VL nor a CD4 cell count taken in the 12 months prior to enrollment was available, a venous blood sample was taken for a VL measurement at the time of enrollment, and the result was used for the eligibility assessment. Type of ARV regimen and co-infection status were not exclusion criteria for receiving ARVs at home. While patients in the control arm underwent the same assessment for clinical stability as those in the intervention arm, all patients in the control arm received standard facility-based HIV care regardless of clinical stability on ART.
The enrollment process described below was the same in both control and intervention facilities. In each healthcare facility in this trial, 1 to 2 study team members were present full-time for the duration of the periods of enrollment and study exit assessment. During the follow-up period, most data collectors split their time between 3 facilities (1 in each of the 3 municipalities). The ART nurses asked all ART patients attending a participating healthcare facility during the enrollment period whether they resided within the neighborhoods that are part of the facility’s catchment area. Those who reported living in the catchment area were sent to the data collector, who was then responsible for introducing the study to the potential participants and obtaining written informed consent. Provided written informed consent was given, the data collector then administered a tablet-based baseline questionnaire (see the “data collection” section for more details). Next, the data collector noted down a description of the location of the patient’s residence (a “map cue”) and recorded the patient’s mobile phone number (and, with the permission of the patient, the mobile phone number of at least 1 household member). Lastly, for those patients who did not have a VL measurement taken in the 12 months preceding study enrollment, the data collector referred the patient back to the nurse for a blood sample that was then sent to the laboratory for a VL assessment. The only difference in the enrollment process between control and intervention facilities was that in intervention facilities the community outreach nurse was responsible for ensuring that the HBC received the map cue and phone numbers so that the patient could be visited at home.
This was a 2-arm non-inferiority cluster-randomized pragmatic trial. The unit of randomization was a healthcare facility with its catchment area (henceforth referred to as a “cluster”). All healthcare facilities that were located in Dar es Salaam Region, provided ART, and had an affiliated team of HBCs were eligible to be included in the trial. The study took place at all 50 healthcare facilities that fulfilled these eligibility criteria except 2 facilities (Amana Regional Referral Hospital and Mwananyamala Regional Referral Hospital), which were excluded because of an ongoing clinical trial at these sites. S1 Table describes the characteristics of each cluster.
Within each of the 3 municipalities in Dar es Salaam, we first matched clusters into pairs based on the number of patients currently on ART at the healthcare facility. For instance, the healthcare facility with the highest number of ART patients in the municipality of Kinondoni was paired with the facility with the second highest number of ART patients in Kinondoni, and so on. We expected this matched-pair design to increase the precision of our effect estimates because the complexity of implementing the intervention (and, thus, the expected probability of implementation failure) would tend to increase with a higher volume of eligible patients. Specifically, because each healthcare facility had only 1 community outreach nurse (except Kigamboni Health Centre and Mbezi Dispensary, which had 2 community outreach nurses), a higher number of ART patients at the healthcare facility resulted in a higher number of patients for whom the community outreach nurse had to supervise the delivery of ARVs into the community. A second advantage of matching on ART patient volume, and thus expected participant volume, was that it increased the probability of an approximately equal number of participants in each study arm, which in turn maximizes statistical power. The randomization was done (by PG) prior to study start using computer-generated random numbers. Allocation concealment in this trial was achieved because entire healthcare facilities (and thus automatically all eligible patients at these healthcare facilities) were randomized to each study arm, and the team member who randomized healthcare facilities did not enroll patients. It was not feasible to blind study participants, project managers, or data collectors to the intervention assignment. Neither was it possible to blind the data analysts (PG, GA, and TB) to the intervention assignment because they were also involved in the project management.
Enrollment into the trial took place at the 18 healthcare facilities in Temeke municipality from March 1, 2016, to July 29, 2016, at the 16 healthcare facilities in Kinondoni municipality from August 1, 2016, to October 31, 2016, and in the 14 healthcare facilities in Ilala municipality from November 1, 2016, to January 31, 2017. Study exit assessments started in Temeke in March 2017, in Kinondoni in May 2017, and in Ilala in June 2017. The study activities during the trial period are described in S2 Table. There was a delay of 11 calendar days between study start and registration of the study on ClinicalTrials.gov because the first 9 days of the trial were used to verify whether the enrollment process was feasible. This informal pilot period did not lead to any changes in any of the planned study processes and was thus considered to be part of the main trial period. No outcome data were collected before registration of the trial.
The prespecified primary endpoint for this trial was the proportion of patients with virological failure at the end of the study period. Virological failure was defined as a VL ≥ 1,000 copies/ml. The prespecified secondary endpoint was patient healthcare expenditures in the 6 months preceding study exit.
This study was designed as a non-inferiority trial. The non-inferiority design only applies to the primary endpoint (the proportion of patients with virological failure) [12]. The rationale for choosing a non-inferiority design for this study was that if the intervention results in equivalent (or better) control of one’s HIV infection among ART patients (as assessed through the VL measurement), then the intervention will be preferable to the standard of care because it has several important advantages beyond clinical effectiveness. First, HBCs earn lower salaries and are quicker to train than nurses and physicians. Shifting important components of ART from nurses and physicians to HBCs will thus likely reduce the per-patient costs of ART delivery and increase the capacity to quickly scale up treatment [15]. Second, shifting care from facilities to homes will decongest the facilities and allow nurses and physicians to concentrate on more complex and clinically unstable ART patients who require more intensive clinical workup and care. Finally, ARV community delivery should reduce the financial and time burdens on patients of having to attend a healthcare facility at frequent intervals [8,10].
Based on consultations with Tanzania’s National AIDS Control Programme, and in line with the margin of equivalence used by Jaffar et al. in their randomized trial of ARV home delivery in rural Uganda [10], we chose a margin of non-inferiority for the risk ratio (RR) of virological failure (comparing the intervention to the control arm) of 1.45. That is, if the RR of virological failure in the intervention group compared to the control group is statistically significantly lower than 1.45, then ARV community delivery will be considered to be non-inferior to standard facility-based care. We registered this prespecified margin of non-inferiority as part of our trial protocol in ClinicalTrials.gov. On the absolute scale, this non-inferiority margin corresponds to a higher absolute probability of virological failure in the intervention group of 9 percentage points, assuming (as done by Jaffar et al. [10]) that 20% of patients in the control arm of the study will be failing virologically at the end of the follow-up period.
This trial was implemented by Management and Development for Health (MDH). MDH is a local Tanzanian-led non-governmental organization based in Dar es Salaam, which works in close collaboration with the Tanzanian Ministry of Health, Community Development, Gender, Elderly and Children.
The study was approved by the research ethics committee of the National Institute for Medical Research (NIMR) in Tanzania on July 16, 2015 (NIMR/HQ/R.8a/Vol. IX/1989), and received an exemption by the institutional review board of the Harvard T.H. Chan School of Public Health in June 2015. During the study period, information on trial progress and analyses of the baseline data were shared at least once every 6 months with the National AIDS Control Programme within the Tanzanian Ministry of Health, Community Development, Gender, Elderly and Children. The Coordinator for Care, Treatment and Support in the Tanzanian National AIDS Control Programme (SL) served as a member of the core team that oversaw and managed this trial.
We calculated the sample size needed for this non-inferiority design under individual randomization (using the “ssi” package in Stata [16]), and then multiplied the sample size under individual randomization by the design effect to arrive at the sample size needed for the non-inferiority design under cluster randomization. The design effect was computed using the “clustersampsi” function in Stata [17], which implements a standard method for calculating power in cluster-randomized trials (but does not allow direct sample size calculation for non-inferiority designs). We assumed an intra-cluster correlation coefficient (ICC) of 0.03. This assumption was based on a study in Dar es Salaam [18], which found the healthcare facility ICC value for the 6-month cumulative incidence of non-adherence to ARVs (defined as a 50% drop in CD4 cell count from its peak value and return to pre-ART CD4 cell count or lower after 168 days on ART, or a VL greater than 10,000 copies/ml after 168 days on ART) to be 0.016. We took the upper bound of the 95% CI of this estimate (which was 0.03) as a conservative estimate of the ICC for our primary endpoint. To our knowledge, this was the best approximation of the expected ICC for our primary endpoint available in the extant literature.
Additionally, we made the following assumptions: 20% of patients in the standard of care arm will be failing virologically at study exit, the probability of a type I statistical error is 0.05, and the correlation coefficient between baseline and study exit VL measurement is 0.5. Under these assumptions and our prespecified margin of non-inferiority, this trial needed 398 patients per study arm to have 80% power to establish non-inferiority. We did not stop enrolling patients once the minimum sample size was reached because an important aim of this study was to investigate to what degree the ARV community delivery model could decongest ART facilities. Our actual sample size was thus substantially larger.
The primary analysis in this study was an intent-to-treat (ITT) analysis. All patients at a healthcare facility were in the ITT sample, regardless of whether they were clinically stable on ART and chose to receive ARVs in (or close to) their homes or not. In secondary analyses, we also examined the treatment effects among only those patients who had a suppressed VL (or, if no VL value was available, a CD4 cell count >350 cells/μl) at baseline (henceforth referred to as “suppressed VL at baseline” for simplicity). The effect of the intervention on our prespecified primary endpoint was examined using a log-binomial model, which generates a RR. In this cluster-randomized trial, the odds ratio (OR) and RR are equally valid relative measures of association, but the OR is sometimes misinterpreted as a RR while the reverse is rarely the case [19]. We thus preferred to express our results using a RR. Whether or not the RR was below the non-inferiority margin was assessed using the upper bound of a 1-sided 95% CI. If the upper bound of this CI for the RR comparing intervention to control was below 1.45, the intervention would be deemed non-inferior to the control. If the upper bound of this CI was greater than or equal to 1.45, then the null hypothesis that the intervention was inferior to the control could not be rejected (at the alpha equal to 0.05 level), and the results of the trial would thus be inconclusive. The CI was obtained from the log-binomial model adjusting standard errors for clustering at the level of the healthcare facility. In the primary analysis, we regressed virological failure at study exit on a binary variable indicating intervention versus control assignment. In secondary analyses, we first added VL at baseline as a covariate to our regression and then additionally controlled for, follow-up time, time between the baseline VL and the study exit VL, age, and sex. Our models did not include an indicator variable for each pair of healthcare facilities to adjust for the matched-pair design, because this is not necessary to obtain valid point estimates [20]. In expectation, including an indicator variable for each pair would lead to a lower (or equal) variance than when ignoring the matched-pair design [20]. However, in our case, this possible increase in statistical efficiency was offset by the fact that some pairs are dropped from the regression that includes all covariates if an indicator variable for each pair is used, because some covariate combinations have 0 observations. In addition to the ITT effect, we estimated the complier average causal effect, i.e., the effect of the ARV community delivery model on those patients who actually received ARVs at home. For this purpose, we used instrumental-variable regression with the randomly assigned intervention status of the healthcare facilities participating in this trial as the instrumental variable for patients' potentially endogenous receipt of ARVs in their homes.
The secondary endpoint (patient healthcare expenditures during the last 6 months) was analyzed using ordinary least squares regression (for inference based on the mean expenditure) and median regression (for inference based on the median expenditure). In these regressions, we determined statistical significance using randomization inference (as implemented in the “ritest” Stata package [21]). We chose to use randomization inference to assess statistical significance because it does not rely on asymptotic properties, which may not apply in cluster-randomized trials with a relatively small number of clusters, especially when there is substantial heterogeneity in cluster size [22]. By specifying the randomization scheme of the study, the randomization inference routine adjusted for clustering at the level of the healthcare facility as well as the matched-pair design.
Fig 1 shows the progression of healthcare facilities (clusters) and patients through the trial. Forty-eight healthcare facilities and 2,172 patients were enrolled into the trial. Twenty-four healthcare facilities with a total of 1,009 patients were enrolled in the control arm (standard facility-based care), and 24 healthcare facilities with a total of 1,163 patients in the intervention arm. The intervention, ARV community delivery, included HBCs delivering ARVs in or close to the homes of patients who were clinically stable on ART. Patients in the intervention arm who were either clinically unstable on ART or who opted not to have their ARVs delivered to their homes continued to receive standard facility-based ART. In all, 516 (44.4%) of the patients enrolled in the intervention arm received ARVs in or close to their homes. For 63 (12.2%) of the patients receiving ARVs in or close to their homes, no VL was available after enrollment (these patients were therefore considered lost to follow-up [LTFU]). For a further 69 (13.4%) of these patients, the only available VL after enrollment was taken prior to receiving the first ARV home visit. These patients were kept in the sample for the primary analysis because they may have indirectly benefited from other patients in their healthcare facility receiving ARVs in or close to their homes. However, we also show results when restricting the sample to those patients receiving ARVs in or close to their homes for at least 90 and 180 days. Among the 359 patients for whom we had a study exit VL taken after receiving the first ARV home visit, the mean duration of receiving ARVs in or close to the home was 226 days, with a standard deviation (SD) of 123 days. The median duration of receiving ARVs in or close to the home was 213 days, with an interquartile range (IQR) of 138 to 300 days.
Thirty-five patients (6 of whom were LTFU) received ARVs in or close to their homes but did not continue until the end of the trial period—4 patients transferred to a healthcare facility outside of Dar es Salaam; 8 informed the study team that they wanted to return to standard facility-based care; 3 were returned to standard facility-based care because they were enrolled based on a baseline CD4 cell count >350 cells/μl but the VL taken at enrollment came back from the laboratory as being non-suppressed; 1 died; 1 was imprisoned; 3 became pregnant and entered into the PMTCT program (without ARV community delivery); and the remainder could not be found again by the HBC. Among the 8 intervention recipients receiving ARVs in or close to their homes who wanted to return to facility-based care, 3 informed the study team that the HBC (the same HBC for all 3 patients) had not delivered their ARVs on time, and the other 5 wanted to return because they were attending the healthcare facility regularly for their children’s healthcare, which made it convenient to pick up ARVs while at the healthcare facility. The 1 death that occurred was a road traffic accident, which was unrelated to the intervention or the trial. Apart from the death, no adverse events were reported to, or detected by, the study team among those patients who received their ARVs in or close to their homes. Similarly, no adverse events among other participants enrolled in this trial (i.e., those in the control arm and those in the intervention arm who did not receive ARVs at home) were reported to the study team. In all, 13.6% (137/1,009) were LTFU in the control arm, and 18.9% (220/1,163) in the intervention arm, yielding a sample size for analysis of 872 patients in the control arm and 943 in the intervention arm.
The sample characteristics for clusters (a healthcare facility with its catchment area) are shown in S1 Table. Table 1 displays the sample characteristics of individuals at the time of the baseline assessment. Patients in the intervention arm were somewhat more likely to be male (22.2% versus 15.4%), to be married (44.3% versus 35.8%), and to self-report having been on ART for a longer time at baseline (mean of 1,407 versus 1,059 days). The percentage failing virologically at baseline was similar between the 2 study arms. The mean follow-up time was 326 days (SD 125) in the control arm and 327 days (SD 120) in the intervention arm. Median follow-up time was also similar between the study arms: 318 days in the control and 322 days in the intervention arm. S3 Table shows that the baseline characteristics of those who were LTFU were similar to those who were included in the analysis except that they were (i) less likely to have received ARVs at home (intervention arm: 28.6% versus 48.0%) and (ii) more likely to have been failing virologically (control arm: 28.2% versus 17.4%; intervention arm: 19.8% versus 15.4%).
Of patients who were offered to receive ARVs in their homes or at another community meeting point of their choosing (i.e., who were both clinically stable on ART and enrolled at a healthcare facility randomized to ARV community delivery), 87.4% decided to enroll in the program rather than remain in standard facility-based ART delivery. Over the course of the study period, a total of 151 HBCs (50 in Temeke, 45 in Kinondoni, and 56 in Ilala) conducted 3,039 household visits to 516 patients in or close to their homes. These patients received a mean of 5.9 (and a median of 6.0) home or community visits for ARV delivery and counseling during the trial period. Over the course of the study, 12 patients contacted the study team to inform them that the HBC had not delivered their ARV supply on time; these 12 patients were under the responsibility of a total of 4 HBCs.
At the end of the study period (defined by the time of measurement of the study exit VL), 10.9% (95/872) and 9.7% (91/943) of patients were failing virologically in the control and intervention arm, respectively. Among those who had a suppressed VL at baseline, 4.3% (27/626) and 4.6% (31/671) were failing virologically at study exit in the control and intervention arm, respectively. Among those who received ARVs in or close to their homes, 5.7% (26/453) were failing virologically at study exit. When restricting the sample to those who had received ARVs in or close to their homes for at least 90 days prior to the date of VL measurement at study exit, 7.0% (24/345) were failing virologically.
The RR for virological failure comparing the patients in the intervention arm to the patients in the control arm was 0.89 (95% CI 0.63–1.25) in the primary (the unadjusted) model (Table 2). The upper bound of the 1-sided 95% CI for this RR was 1.18, and therefore below the non-inferiority margin of 1.45. When the sample was restricted to those patients with a suppressed VL (<1,000 copies/ml) at baseline—57.6% (440/764) of whom received ARV community delivery (as opposed 48.0% when including all ART patients at intervention facilities)—the RR was above 1, and the upper bound of the 1-sided 95% CI above the non-inferiority margin, in all models (S4 Table). In supplementary files, we also show results (i) when adjusting for follow-up time and time between the baseline and study exit VL measurement (S5 Table), (ii) when restricting the sample to those for whom the study exit VL was taken at least 200 days after enrollment into the trial (S6 Table), (iii) when restricting the sample to those for whom the study exit VL was taken at least 200 days after the baseline VL (or the baseline CD4 cell count) (S7 Table), and (iv) when adjusting for time on ART at baseline (S8 Table). The results were similar when using a VL threshold of ≥200 copies/ml to define virological failure (S9 Table).
The complier average causal effect (i.e., the effect of the ARV community delivery program on those who received ARVs in their homes) was not significantly different from 0 in all models (Table 3). The regression coefficients in Table 3 can be interpreted as the absolute difference in the probability (between 0 and 1) of virological failure in the intervention arm compared to the control arm. In the unadjusted model, receiving ARVs at home led to a 2.6 percentage point lower probability of being in virological failure at the end of the study period compared to being in the control arm. As supplementary files, we also show the complier average causal effect under different model specifications and sample restrictions (S10 Table), and when restricting the sample to those who had a suppressed VL at baseline (S11 Table).
The percentage of all ART patients (regardless of whether they were eligible for enrollment into the trial or for receiving ARVs at home) at each intervention facility who received ARVs at home at least once varied from 0.3% to 19.0%, with a facility mean of 4.4% (Table 4).
In the study exit questionnaire, 83.1% (295/355—the denominator is all those who received ARVs and counseling in or close to their homes and for whom data from the study exit questionnaire were available) reported being “very satisfied” with the program of receiving ARVs at home (Fig 2). In all, 88.7% (315/355) reported that the HBC always delivered the ARVs on time (Fig 3), and 2.0% (7/355) reported that they missed a dose of ARVs because the HBC did not deliver ARVs on time. In all, 96.3% (342/355) of those who received ARVs in or close to their homes reported that they would like to continue with this delivery model (rather than returning to standard facility-based care), and 99.7% (354/355) said they would recommend it to other communities. In total, 0.9% (3/355) of patients who received ARVs in or close to their homes reported that the program led to an unintentional disclosure of their HIV status to a third person.
We experienced 3 major challenges during data collection. First, most participants did not have a VL or CD4 cell count taken in the preceding 12 months in their clinical records at the time of the baseline questionnaire. Thus, the study team had to send a blood sample for VL testing to the laboratory. Receiving the results from the laboratory on these VL measurements took between 4 and 12 weeks. As a result, for most participants, the study team was not able to assess eligibility for the intervention until 1 to 3 months after the baseline questionnaire administration. Second, 417 participants did not return to the facility for their study exit assessment (and a clinical checkup)—136 of these participants were at control facilities and 281 at intervention facilities. For some of these individuals (as well as many individuals who had missing VL results for other reasons), we were able to retrieve their latest VL from the central health system database housed at MDH because Tanzania started implementing a yearly VL for all ART patients during the study period. The central health system database records all VLs taken at any healthcare facility in Dar es Salaam. Loss to follow-up in this report thus refers to participants who did not return to the healthcare facility for the study exit assessment and for whom we were unable to retrieve their VL from the central health system database. Third, we experienced difficulties in linking participants across our different study databases. The databases used in this study were a study logbook, in which the data collection team kept a list of all participants in the trial (as well as age and sex of the participant), the baseline questionnaire data, the baseline laboratory data, and the study exit questionnaire data. Out of the 2,172 participants in this study, we had all questionnaire data for 1,348 participants; 193 had only logbook data, 94 only logbook and baseline questionnaire data, 139 only logbook and baseline laboratory data, 271 only logbook, baseline questionnaire, and baseline laboratory data, and 127 only logbook and study exit questionnaire data. The proportion with complete data was similar between the 2 study arms—64.7% in the control arm and 59.6% in the intervention arm. These participant linking issues were responsible for the relatively high level of missingness in socio-demographic variables other than age and sex, and the lower sample size for the analysis of healthcare expenditures as compared to VL measurements. In addition to a relatively high number of participants not returning to the healthcare facility for the study exit assessment, the main cause of unsuccessful linking was that the data collectors entered neither the health system patient identifying number nor the study identifying number correctly into the tablet.
In this randomized controlled trial in Dar es Salaam, we investigated the effect of an ARV community delivery model—lay health workers delivering ARVs directly to patients’ homes or other meeting points in the community if the patient was stable on ART, and patients receiving standard facility-based care if they were unstable on ART—on the probability of virological failure and patients’ healthcare expenditures. We found that the ARV community delivery model performed at least as well in averting virological failure as standard facility-based ART.
Patient satisfaction with the program was high, and receiving ARVs in the community through HBCs is likely to save patients substantial amounts of time. However, 2 other envisaged benefits of the program—decongestion of healthcare facilities and reductions in patient healthcare expenditures—were minimal. The ARV community delivery model shifted only a small proportion of all ART patients at a healthcare facility from facility- to community-based care, so that it is unlikely to have had a noticeable effect on clinicians’ workload and waiting times at healthcare facilities. Regarding patient ART expenditures, the median cost of attending 1 ART visit for patients was low, and thus the extrapolated savings to patients from receiving ARVs in the community were small.
Table 5 summarizes the key advantages and disadvantages of the ARV community delivery model tested in this trial. Regarding retention in care, ARV community delivery may improve long-term retention in care as only about 1 in 100 patients who received ARVs in or close to their homes were lost to care for reasons other than returning to standard facility-based care. However, because patients would continue to receive ARVs in or close to their homes regardless of whether they attended their once-a-year clinical checkup at the healthcare facility, they may miss these annual checkups, potentially reducing the clinical effectiveness of ART in the longer term. Implementation research needs to accompany any future scale-up of ARV community delivery to determine how the model performs over many years and how ART patients can be best motivated to attend their annual facility-based checkups. Moreover, in a routine scale-up of the ARV community delivery model, quality of care and clinical outcomes may be worse than those observed in this trial, because the intermittent presence of data collectors in the healthcare facilities that participated in this trial may have led to higher fidelity of implementation of the model and heightened attention to patient outcomes. The overall effect of ARV community delivery on per-patient costs is unclear at this point. Shifting ART from physicians and nurses to lay health workers will decrease the per-patient costs—if all other factors remain the same—because lay health workers earn less than physicians or nurses. However, some factors change when a health system moves from the facility- to the community-based ARV delivery model. Lay health workers may need more time to treat and care for ART patients because of the travel to the patient’s home and because each patient–health worker interaction may take longer in the home than in the facility-based setting—for instance, because patients need to find a private place before the interaction can take place, or because some of the 'rituals' that save time in the healthcare facility, such as minimal social conversation, may not be practiced in the home setting.
An important limitation of the ARV community delivery program, as implemented in this study, is that it allowed for only a small proportion of ART patients at the study’s healthcare facilities to receive ARVs at home. Table 6 outlines possible ways of increasing this proportion. Because many ART patients in Dar es Salaam do not attend the healthcare facility closest to where they live, the main reason for the low enrollment in receiving ARVs at home was the eligibility criterion that a patient must reside in the facility’s catchment area. While removing this eligibility restriction would likely greatly increase the proportion of eligible ART patients, an important drawback is that delivering ARVs to patients’ homes would become logistically more complex and possibly more costly to implement. The HBCs would either have to travel across the entire city to deliver ARVs or a mechanism would have to be established by which an HBC affiliated with a different healthcare facility than the one the patient is attending (but which is closer to where the patient resides) would be tasked with delivering ARVs to the patient’s home. The former would be costly to implement due to the much higher transport costs compared to the current model. The latter would be unlikely to add substantial costs to the model tested in this study, but it would require fast and reliable communication across healthcare facilities. At the time of study conception, the study team felt that establishing such a mechanism would be too complex logistically to be successful. However, before the trial, the team was not aware of the extent to which restricting eligibility to patients living in the healthcare facility’s catchment area would impact on the proportion of ART patients that could be enrolled in the trial. Retrospectively, we, therefore, believe that it would have been worthwhile to more intensively investigate and pilot ways of delivering ARVs to the homes of patients living outside of their healthcare facility’s catchment area. Future implementation research should identify, design and test approaches for ARV community delivery outside of facility catchment areas.
A second possibility to increase the number of patients enrolled in the ARV community delivery model is to expand the number of healthcare facilities in Dar es Salaam that have an affiliated team of lay health workers. At the time of study start, only about a third of the healthcare facilities that offered ART in Dar es Salaam also had a team of the local lay health worker cadre—the HBCs. A third possibility is to offer enrollment in the ARV community delivery program to all ART patients as long as they are able to meet up with the HBC within the facility’s catchment area. This would allow those who do not reside in the facility’s catchment area to still benefit from the program by foregoing the time spent waiting at the healthcare facility to pick up a new supply of ARVs. A fourth possibility, removing or relaxing the eligibility criterion that patients must be clinically stable on ART to receive ARVs in their homes, would only lead to a small increase in enrollment. In this study, removing the clinical stability criterion for eligibility would have led to an increase in enrollment of only 15.7%. As a fifth possibility, one could imagine a model that delivers ARVs to patients’ homes (or other meeting points in the community) but instead of HBCs uses another lay cadre for this purpose, such as HIV treatment supporters. This approach would, however, have the disadvantage that it would not harness the rapport that many HBCs have developed with members of the communities they serve. In addition, the non-HBC cadre may have less training or experience than the HBCs to recognize symptoms and signs that require referral to a nurse or physician. Lastly, enrolling pregnant women living with HIV would likely substantially increase enrollment numbers, but such a model would need to be carefully designed to avoid a reduction in antenatal care attendance.
This trial suffered from a number of implementation challenges. While we are certainly not the first to have faced such difficulties, and there are excellent resources on implementing randomized trials in resource-poor settings (e.g., Glennerster and Takavarasha [23]), it is our view that other researchers planning pragmatic health services trials in resource-poor settings might benefit from our experience. Specifically, we have drawn the following implementation lessons from this trial for our future work. First, relying on data collectors to correctly enter long unique identifying codes for patients into study registers and tablets should be avoided wherever possible, such as by trying to automate the process (for instance, with the use of bar codes). Second, collecting data in the minimum number of datasets needed to accomplish the task at hand will minimize linkage problems across datasets. Third, in future work, we will endeavor to devise measures early to reduce the possibility of bias from loss to follow-up. One approach in this regard could be to randomize the level of encouragement (e.g., the number of phone calls or level of monetary compensation) that patients receive to return to the healthcare facility for the study exit assessment, which would create an instrument for loss to follow-up that could be exploited to correct for attrition based on both observable and unobservable characteristics [24]. Lastly, in the case of non-inferiority trials, an extensive piloting period might be useful to verify to what degree the assumptions about the envisaged benefits of the intervention are likely to hold true, and thus whether a non-inferiority design is justified.
This study has several limitations. First, only a relatively small proportion of study participants at intervention facilities received ARVs in the community through the HBCs (48.0% [453/943] of participants at intervention facilities were enrolled in the program of receiving ARVs through the HBCs, and 41.3% [345/835] received ARVs through the HBCs for at least 90 days before the study exit VL measurement was taken). The primary analysis as per our study protocol, however, included all ART patients at a healthcare facility who resided in the facility’s catchment area because patients at intervention facilities who remain in facility-based care may indirectly benefit from some patients receiving ARVs at home through the envisaged decongestion of the healthcare facility. All else remaining equal, the lower the proportion of study participants at intervention facilities who receive ARVs at home, the more similar the ARV community delivery model is to the standard of care and thus the more likely the intervention is to appear to have no effect. Therefore, if patients receiving ARVs at home through HBCs had a worse virological outcome than if they had remained in standard facility-based care, this study may have found the ARV community delivery model to be non-inferior based on this “dilution” of the effect of receiving ARVs at home. To try to ascertain whether our non-inferiority conclusion is partly due to this dilution effect, we show the results when restricting the sample in secondary analyses to only those patients who had a suppressed VL at baseline—58.9% (395/671) of whom received ARVs at home (S4 Table). The RR in the unadjusted model among this sample was 1.07, but the upper bound of the 1-sided 95% CI (1.75) was above the non-inferiority margin (1.45), largely because this study was not powered to determine non-inferiority for this smaller sample of patients. Overall, it is important to note that dilution could only explain some part of our results, because dilution can never lead to an improvement in outcomes in the intervention arm, which is what we observe. We also calculated the complier average causal effect (i.e., the effect of the ARV community delivery model on only those who received ARVs at home) and found that the point estimates were generally close to 0 (although the CIs were fairly wide) (Table 3).
Second, the proportion of patients LTFU was not only relatively high, but also substantially larger in the intervention than in the control arm. This raises the concern that even if those LTFU in the intervention arm had the same probability of failing virologically at study exit as those LTFU in the control arm, this trial might have falsely concluded that ARV community delivery performs at least as well as the standard of care simply if those LTFU (regardless of study arm) had a higher probability of virological failure than those included in the analysis. This was the case at baseline (S3 Table) and, to the degree that baseline virological failure predicts virological failure at study exit, was thus likely also the case at study exit. However, even if we assumed the most extreme scenario, namely that 100% of those LTFU (regardless of study arm) were in virological failure at study exit, the unadjusted RR (i.e., our primary analysis) would have been 1.16, with the upper bound of a 1-sided 95% CI being 1.32, which is still below the margin of non-inferiority of 1.45 (S1 Text). Our conclusions are, therefore, robust to this source of bias. A more likely scenario in which bias from attrition could have changed our conclusions is that those LTFU in the intervention arm may have had a higher probability of failing virologically at study exit than those LTFU in the control arm. However, the fact that those who subsequently were LTFU in the intervention arm were less—not more—likely to be failing virologically at baseline than those LTFU in the control arm (S3 Table) suggests that (under the assumption that baseline virological failure predicts virological failure at study exit) this scenario is unlikely. Nonetheless, while our results appear to be robust to bias from attrition, it is possible that we might have reached a different conclusion if this study had had no attrition.
Third, it can be argued that the level at which the margin of non-inferiority was set (here, a RR of 1.45) is arbitrary. We have, however, adhered to best practices in setting this non-inferiority margin by involving relevant policy-makers in the decision, considering the margin set in similar studies conducted prior to this trial, and specifying the non-inferiority margin in our study registration. Nonetheless, the somewhat arbitrary nature of setting a non-inferiority margin is a limitation of non-inferiority trials in general [25], and thus it also applies to this trial.
Fourth, patients in the control arm appear to not have included ART visits when answering questions on health service utilization in the preceding 6 months. A more extensive piloting phase might have uncovered this misperception among respondents prior to the trial, which would have allowed us to word the relevant questions differently. In addition, qualitative work accompanying this trial could have shed further light on whether, and if so why, respondents did not include ART visits in these answers. However, given the low costs incurred from attending ART, it is unlikely that our estimates of patient healthcare expenditures during the preceding 6 months would have been substantially different had patients included ART visits in their response. In our view, a more important limitation of this study with regards to patient healthcare expenditures was that only those living close to the healthcare facility—and thus likely facing the lowest transport costs to get to the healthcare facility—were eligible for enrollment. It is therefore possible that the cost savings to patients would have been higher had all ART patients at a healthcare facility been eligible to enroll, regardless of the neighborhood in which they lived. Similarly, the cost savings would likely have been higher in more rural settings, where the average distance from a patient's home to the nearest healthcare facility is typically far longer than in urban settings.
Fifth, this study excluded pregnant women, and thus the trial results cannot be generalized to PMTCT care. Lastly, with patients receiving ARVs at home for an average of 226 days, we were unable to assess the longer-term safety of the ARV community delivery model.
As implemented in this trial, with roughly 40% of patients in the intervention arm receiving ARVs at home and 60% remaining in standard facility-based care, ARV community delivery performed at least as well as the standard of care regarding the critical health indicator of virological failure. ARV community delivery did not significantly reduce patient healthcare expenditures, but satisfaction with the program of receiving ARVs at home was high. In addition, receiving ARVs at home is likely to save patients substantial amounts of time and may reduce government health expenditures per ART patient. It is our view that with modifications to allow a larger proportion of ART patients at healthcare facilities to enroll, the ARV community delivery model can serve as an important alternative for ART delivery. The model holds particular promise for settings where—relative to demand—human and physical resources for ART are increasingly scarce. As the model is scaled up to serve increasingly large populations in the future, accompanying implementation research can ensure that issues arising due to the greater scale and longer time horizons than those in our trial are quickly detected and addressed.
|
10.1371/journal.pcbi.1000243 | Malleable Machines in Transcription Regulation: The Mediator Complex | The Mediator complex provides an interface between gene-specific regulatory proteins and the general transcription machinery including RNA polymerase II (RNAP II). The complex has a modular architecture (Head, Middle, and Tail) and cryoelectron microscopy analysis suggested that it undergoes dramatic conformational changes upon interactions with activators and RNAP II. These rearrangements have been proposed to play a role in the assembly of the preinitiation complex and also to contribute to the regulatory mechanism of Mediator. In analogy to many regulatory and transcriptional proteins, we reasoned that Mediator might also utilize intrinsically disordered regions (IDRs) to facilitate structural transitions and transmit transcriptional signals. Indeed, a high prevalence of IDRs was found in various subunits of Mediator from both Saccharomyces cerevisiae and Homo sapiens, especially in the Tail and the Middle modules. The level of disorder increases from yeast to man, although in both organisms it significantly exceeds that of multiprotein complexes of a similar size. IDRs can contribute to Mediator's function in three different ways: they can individually serve as target sites for multiple partners having distinctive structures; they can act as malleable linkers connecting globular domains that impart modular functionality on the complex; and they can also facilitate assembly and disassembly of complexes in response to regulatory signals. Short segments of IDRs, termed molecular recognition features (MoRFs) distinguished by a high protein–protein interaction propensity, were identified in 16 and 19 subunits of the yeast and human Mediator, respectively. In Saccharomyces cerevisiae, the functional roles of 11 MoRFs have been experimentally verified, and those in the Med8/Med18/Med20 and Med7/Med21 complexes were structurally confirmed. Although the Saccharomyces cerevisiae and Homo sapiens Mediator sequences are only weakly conserved, the arrangements of the disordered regions and their embedded interaction sites are quite similar in the two organisms. All of these data suggest an integral role for intrinsic disorder in Mediator's function.
| Intrinsically disordered proteins/regions do not adopt well-defined three dimensional structures; instead, they function as conformational ensembles. They are distinguished in molecular recognition and involved in various regulatory processes. Several components in the transcription machinery–for example, the transactivator domains of transcription factors–are disordered. Mediator, which is a large complex that transduces regulatory information from activators/repressors to the core apparatus, was found to contain a preponderance of intrinsically disordered regions in its various subunits. Such disordered regions are commonly involved in conformational changes coupled to functional transitions, in protein–protein interactions, or in posttranslational modifications. Several such predicted recognition sites were in good agreement with experimental data. Intrinsically disordered regions illuminate a novel aspect of Mediator's regulation and could explain its versatility and specificity in handling transcriptional signals. Their integral role in Mediator function is further underscored by the conserved arrangements of ordered/disordered segments and of the embedded interaction sites.
| The Mediator complex is a gigantic (1 MDa) multi-protein complex that plays a number of essential roles in eukaryotic gene regulation [1]. It functions as a co-activator, a co-repressor as well as a general transcription factor by transmitting information from the regulatory factors bound at enhancers to the RNAP II transcription machinery [1],[2]. Mediator is recruited by promoter- and/or enhancer-bound activators [3] followed by association of general transcription factors and RNAP II with the promoter in vivo [4],[5] (Figure 1). Mediator dissociates from RNAP II after initiation, and remains attached to the promoter [6],[7] providing a pre-formed scaffold for the reinitiation [8].
Interactions with RNAP II and regulatory proteins induce dramatic conformational changes in Mediator [9],[10]. Activator induced specific rearrangements in Mediator expose cryptic RNAP II binding site and modulate the assembly of the pre-initiation complex (PIC) [11],[12]. This suggests that activators/repressors regulate transcription by altering the structure of the RNAP II holoenzyme. These conformational changes were thus proposed to underlie the regulatory mechanism of Mediator [13].
Mediator consists of 20–30 subunits that are organized in a modular fashion, with Head, Middle, and Tail regions [14] (Figure 1). The Tail can serve as the main target for activators/repressors [15]. The Med9 submodule of the Middle may connect the regulatory signals to the Head [16], which could in turn interact directly with RNAP-TFIIF for pre-initiation complex formation [17]. The Middle also receives repression signals from the CDK module, which dissociates prior to transcription [18]. The functions of the individual subunits however, are rather obscure apart from the reported kinase activity of the Cdk8 [19] and the histone acetyltransferase activity of the Med5 [20], which are non-essential for Mediator's function. Mediator protein sequences are highly variable with the exception of a few subunits [21]. The majority of the subunits have no apparent domains, not even the expected domains for chromatin modification such as chromo [22] or bromo domains [23] (Y.T. unpublished data). Nevertheless, based on cryo-electron microscopy, the overall structural organisation of several eukaryotic Mediator complexes is similar [24].
The low sequence conservation of Mediator proteins and the absence of known globular domains suggest the presence of disordered regions in Mediator. Such disordered regions might be responsible for similar structural characteristics in different organisms observed in EM studies [24] despite the lack of sequence conservation. IDRs can contribute to Mediator's function in three different ways: they can provide flexible target sites that can adapt to different partners with variable architectures; they can act as malleable linkers connecting globular domains that impart modular functionality on the complex; and they can also facilitate assembly and disassembly of complexes in response to regulatory signals.
To understand whether IDRs play a role in transcription regulation of the Mediator, 340 sequences of 30 subunits were collected (Table S1) and their tendencies for intrinsic disorder were predicted using bioinformatics approaches [25],[26]. Out of the 27 eukaryotic organisms Saccharomyces cerevisiae and Homo sapiens sequences were analyzed in detail and the results were corroborated using all available sequences (shown in the Supporting Information, Figures S1, S2, S3, S4, S5 and S6). The estimated level of disorder increases from yeast to man and in both organisms the propensity of disordered regions substantially exceeds that of signaling proteins and also that of multi-protein complexes of similar size. Subunits that interact with activators/repressors or function in regulatory signal transfer, located mostly in the Tail and Middle modules, are most abundant in IDRs. Overall, 43 sites for protein-protein interactions were predicted in 16 subunits in Saccharomyces cerevisiae and 79 sites in 19 subunits in Homo sapiens Mediator. In yeast, 11 of the predicted molecular recognition features (MoRFs) overlap with experimentally detected binding sites or post-translational modification sites, out of which those in Med7/Med21 [27] and Med8/Med18/Med20 [28] complexes have been structurally confirmed. The arrangement of ordered/disordered regions and location of disordered interaction sites are similar in Saccharomyces cerevisiae and Homo sapiens, although sequences of IDRs are only weakly conserved. All these results suggest that Mediator functions as a malleable machine in transcription regulation with an integral role for intrinsically disordered regions for the gene-specific regulatory functions.
Preference of Mediator proteins for intrinsic disorder was assessed by two independent bioinformatics approaches: PONDR-VSL1 that is a support vector machine algorithm [25] and IUPred that utilizes statistical inter-residue potentials [26]. Disorder predictions for Mediator proteins were carried out by both techniques at the amino acid level using sequences of individual proteins and the disorder scores were averaged over the entire sequence. As the two prediction methods provided consensus results, in the following only those obtained by the IUPred algorithm will be detailed. A preponderance of intrinsic disorder (average disorder above the 0.5 threshold value) was found in 4 and 6 out of 25 subunits in Saccharomyces cerevisiae and Homo sapiens, respectively (Figure 2). In addition, Med9 (in yeast) and Med4 (in man) have a level of disorder that is comparable to the disordered proteins assembled in the DisProt database [29]. These proteins likely lack a well-defined tertiary structure in the free form, but can partly or fully fold upon interacting with their partners [30]. The inherent flexibility of these subunits however, can contribute to structural organisation and molecular interactions of the complex. Overall, the levels of disorder (as averaged over all subunits) are higher in man than in yeast, suggesting an increase in the propensity or length of disordered regions. In Saccharomyces cerevisiae the Tail is most enriched in subunits with preference for intrinsic disorder (Med2, Med3, Med15), while in Homo sapiens the Middle module appears to be most abundant in malleable proteins (Med1, Med9, Med19, Med26). In the Head only Med8 is predicted to be disordered in Homo sapiens. Disorder scores averaged over sequences from all available organisms also indicate large variations in some subunits (please note, that in this case the number of sequences/subunits differ; Figure S1). This might implicate functional changes of various Mediator proteins during evolution.
The amino acid compositions of Mediator proteins in Saccharomyces cerevisiae and Homo sapiens are also incompatible with a folded structure [31] (Figure 3), although they exhibit some variations. As compared to globular proteins, yeast and human Mediator proteins are depleted in hydrophobic (I, L, V), aromatic (W, Y, F) and C residues (designated as order-promoting); and enriched in polar (Q, N, T, S), charged (E, D) and structure-breaking (P) residues (designated as disorder-promoting). Such a composition resembles the general characteristics of intrinsically disordered proteins [32]. Various subunits, like the Med4 and Med15 are abundant in potential post-translational modification sites (S and T) that are preferably embedded in disordered regions [33]. Generally disordered polyQ and polyN regions frequently appear in various subunits, such as Med1, Med9, Med10, Med12 and Cdk8 (Figure S2). The Q-rich region in Med15 in Saccharomyces cerevisiae for example is involved in glucocorticoid receptor transactivity [34]. The propensity of Q-rich regions also increases from yeast to man. Repeat expansion may contribute to rapid evolutionary changes of Mediator proteins and may have created linkers between globular segments [35].
Intrinsically disordered regions of any length have been observed to be involved in biological functions, but those of 30 residues or longer have been especially well studied [36]. The function of these regions are diverse but are frequently related to molecular recognition [37]. IDRs are usually exploited for regulatory purposes as 66±5% of cell-signaling proteins [38], and 90% of transcription factors were predicted to contain IDRs (longer than 30 aa) [39],[40]. In Saccharomyces cerevisiae 80% of Mediator subunits have predicted IDRs equal to or longer than 30 residues, and 24% have IDRs above 100 residues in length [25] (Figure S3). In Homo sapiens, IDRs longer than 30 and 100 residues appear in 75% and 32% of Mediator proteins, respectively (Figure S3). This suggests that the length of IDRs increased from yeast to man. The number of disordered segments is also higher in the human complex than in the yeast complex (Figure 4). This is mostly due to the discrepancy in the number of IDRs in the Middle. This module is the most abundant in disordered regions in Homo sapiens. In the Head the propensity of IDRs is also slightly higher (below 70 residues in length) in man than in yeast. In Saccharomyces cerevisiae, disordered regions are preferably located in the Tail, some exceeding 100 residues in length. Along these lines, the longest IDRs in yeast are found in Med2 (334), Med3 (256), Med15 (263) of the Tail, whereas in human Mediator, Med1 (645), Med9 (241), Med26 (261) of the Middle are equipped with the longest IDRs (Figure 5 and Table S2). Med13 of the CDK appears to have a long IDR in both organisms: 226 and 162 in yeast and human, respectively.
Large multi-protein complexes generally take advantage of the plasticity of their components; i.e., the population of intrinsically disordered segments increases with complex size [41]. Multi-protein complexes of 11–100 proteins fulfilling various functions, have IDR propensity with median value of 12%, which estimates the percentage of disorder required to assemble a complex of a given size. The percentage of amino acids in IDRs is 32% and 33% in yeast and human Mediator, respectively (Figure S4), and these values considerably exceed those obtained for other complexes of similar size. One possibility is that the Mediator IDRs perform additional (eg., regulatory) tasks besides the self-assembly of the complex. Indeed, the level of disorder in Mediator is even higher than in signaling proteins (Figure S3).
Molecular recognition by IDRs is achieved by short, distinguishable segments, such as preformed elements [42], molecular recognition features [43], primary contact sites [44] and linear motifs [45],[46]. Preformed elements [42] and molecular recognition features [43] are predisposed to fold upon binding, and this reduces the entropy penalty of the recognition process. Primary contact sites [44] or linear motifs [45] are usually short, exposed segments that facilitate formation of highly specific interactions. In general all these recognition sites have higher local hydrophobicity than their environment and often exhibit transient secondary structure [46].
In Saccharomyces cerevisiae and Homo sapiens Mediators, we focused on those recognition sites that are biased for an α-helical conformation, termed α-MoRFs. These segments fold onto an α-helix in the bound form and can be predicted from the irregularities in computed disorder patterns using a neural network algorithm with 0.87±0.08 accuracy [47]. A prototypical example of an α-MoRF is the short α-helical segment in the disordered transactivator domain of p53 that mediates binding to Mdm2 [48],[49]. Multiple, tandem binding sites can be found in the BRCA1 protein that serve a scaffold function [50]. In yeast, predictions indicate the presence of 43 α-MoRFs in total, distributed over 16 subunits (Table 1). Some subunits have multiple α-MoRF regions, with Med15 of the Tail (11 α-MoRFs) and Med13 of the CDK module (6 α-MoRFs) in yeast having the largest numbers of these regions. In accord with the increased level of disorder, 79 interaction sites were identified in 19 subunits in Homo sapiens (Table S2). Most interaction sites were located in Med3 of the Tail (18 α-MoRFs) and Med1 of the Middle (14 α-MoRFs) and Med13 of the CDK (8 α-MoRFs).
The predicted α-MoRFs in Saccharomyces cerevisiae, which may serve as potential target sites for protein-protein interactions or for post-translational modifications, were compared to experimentally verified binding sites reported in literature or assembled in protein-protein interaction databases. So far 11 out of the of the 43 predicted α-MoRFs in yeast have been experimentally corroborated (Table 1). For example, the α-MoRF encompassing residues 333–350 of Med3 likely corresponds to the Gcn4 target site [51], while the α-MoRF 195–212 predicted in Med7 serves as a contact site with Med10 [52]. Specific mutation sites in Med17 at the interaction sites with the Middle and Tail modules [2] (and Y.T. unpublished data) also coincide with the identified MoRFs. The region 116–255 of Med15 that interacts with Gal4 [53] contains two predicted α-MoRFs. The 261–351 segment of Med15 that is responsible for transcriptional activation of glucocorticoid receptor also contains one α-MoRF that matches the observed interaction site [34]. The region 396–655 of Med13 contains 3 predicted α-MoRFs and has been observed to contact various partners: Caf1, Crc4, Not2 as well as Cdk8 [54]. The predicted phosphorylation site at T237 in Med4, which might play role in enhancement of RNAP CTD phosphorylation by TFIIH [55], matches the experimentally determined position.
In the case of Med7 and Med8, the available crystal structures of the Med7/Med21 [27] and the Med8/Med18/Med20 [28] complexes can be used for structural validation of α-MoRFs (Figure 6). The Med7/Med21 heterodimer serves as a hinge that was proposed to be responsible for large scale changes in the Mediator's structure [27]. In the complex three α -helices of Med7 were observed that constitute a coiled-coil. The predicted α-MoRF 195–212 is located at the C terminal end of α3 that makes contacts with α3 helical region of Med21. In accord with its predicted increase in flexibility, this segment has elevated B-factors in the bound form. Of course the elevated B-factor values might simply stem from its terminal location. The C-terminal fragment encompassing residues 193–210 of Med8, which was predicted as an α-MoRF, adopts an α-helical conformation in the Med8/Med18/Med20 complex [28]. While 27 residues of Med8 were used for crystallization, only 16 were observed in the complex, indicating the presistance of disorder even in bound form. This segment is embedded in a larger disordered region, encompassing the linker between the C and N terminal of Med8. This linker exhibits enhanced sensitivity to proteolytic digestion in the free protein corroborating its disordered state. This region was shown to be essential for transcription in vivo by harboring elongin B and C [56].
An independent argument for the functional importance of the predicted α-MoRFs in 6 subunits (Med7, Med9, Med10, Med11, Med15, Med17, cf. Table 1) is underscored by their overlap with helical regions that have been proposed to be highly conserved from yeast to man [21].
IDRs in homologous proteins often exhibit remote sequence relationships. The functioning of IDRs likely relies on their biased amino acid composition and their short motifs [43],[44],[46], the latter of which enables a rapid evolution of IDRs [57],[58]. Hence, the presence of IDRs might account for the weak sequence conservation of Mediator proteins despite their similar functions or architectures [14],[24]. As anticipated, a remarkable difference between the sequence conservation of disordered and ordered regions were also seen in Saccharomyces cerevisiae and Homo sapiens Mediators (Figure 7). This distinction can also be observed if Mediator subunits from all available organisms are aligned (Figure S5). In contrast to the sequence behaviors, the propensities of order and disorder promoting amino acids in IDRs were found to be highly conserved (Figure S5).
Recently we introduced a method to assess the conservation of IDRs based on the arrangements of ordered and disordered segments, as predicted by the IUPred algorithm, in different sequences [59]. This can be evaluated at the level of residues, i.e., by computing the percentage of residues designated as ordered or disordered at the same position in sequence alignments. On the average 74.5% of residues are located in regions with the same character (disordered or ordered) in Saccharomyces cerevisiae and Homo sapiens (Figure S6). Alternatively, the overlap between ordered and disordered segments in different sequences can be measured by adopting the accuracy measures of secondary structure predictions [59],[60]. In this case the arrangement of ordered/disordered segments in different sequences is compared to each other in terms of the persistence of their location in different organisms. The overlap between the patterns of ordered/disordered regions in yeast and human Mediator is 73.2%. This value significantly exceeds the corresponding value determined from randomized sequences with the same amino acid composition (Figure 8). Thus it appears that, in contrast to the sequences themselves, the arrangements (patterns) of disordered regions are conserved in different organisms, providing a further support for their functional importance.
Transcriptional control requires an intimate interplay between the enhancer- and repressor-bound factors and the basal transcription machinery. In eukaryotic organisms large co-activators, such as the Mediator complex [1] or CBP/p300 [61] are responsible for transducing regulatory information to the core apparatus and link chromatin remodeling to m-RNA synthesis. The mechanism by which these large assemblies impart versatility and specificity on transcription regulation however, remains to be uncovered. It has been proposed that dramatic conformational changes that occur upon interactions with regulatory proteins [10]–[13] as well as with RNAP II [9] could serve as a basis of the Mediator's control mechanism [13]. Such large-scale structural rearrangements could be facilitated by highly flexible/malleable segments that can serve as molecular “hinges” [10]. Furthermore, based on the abundance of intrinsically disordered proteins in signaling [36], we reason that the signal transducer function of Mediator is also intertwined with IDRs. IDRs mediating specific, transient interactions were observed at various checkpoints of transcription [62], like in histone tails [63], transactivator domains of transcription factors [64] and the C-terminal domain of RNAP II [65].
In this study, bioinformatics approaches were employed to assess the preference of Mediator proteins for intrinsic disorder, focusing on the comparison of Saccharomyces cerevisiae and Homo sapiens Mediator complexes. Various subunits, located mostly in the Middle (Med1, Med9, Med19, Med26) in human and in the Tail (Med2, Med3, Med15) in yeast are predicted to be enriched in disordered regions (Figure 2 and Figure 4). As the level of disorder in these proteins is higher than that of proteins assembling into other complexes of similar size, IDRs are likely exploited for additional, regulatory functions besides facilitating the self-assembly of the complex. Along these lines, the propensity of disordered regions in both yeast and human Mediator exceed that in signaling proteins. Results obtained on all available Mediator sequences (340) presented in Supporting Information (Figures S1, S2, S3, S4, S5 and S6) also corroborate the results obtained on the two organisms emphasized here.
Because the predictions were performed on individual sequences, we cannot exclude the possibility that regions predicted to be intrinsically disordered adopt a well-folded structure upon interacting with other Mediator subunits or with regulatory proteins. Electron microscopy results however indicate the pliability of the complex at low ionic strength (Francisco Asturias, private communication) that argues against the complete loss of disordered state in the Mediator complex. An independent argument comes from the structure-function analysis of complexes of intrinsically disordered proteins. In many cases IDRs were found to remain disordered even bound to their partners and yet critically affect binding affinity or specificity [66]. In these ‘fuzzy’ complexes IDRs interact via short segments, while the embedding regions may remain structurally variable.
To probe if IDRs are utilized for macromolecular communication, sites of protein-protein interactions were predicted in disordered regions and are biased for an α-helical conformation. In total 43 α-MoRFs were identified in yeast Mediator, with 79 α-MoRFs in human Mediator. The roles of α-MoRFs as protein-protein interaction sites is also suggested by the overlap of the predicted and experimentally observed binding regions. For example, in Saccharomyces cerevisiae 11 α-MoRFs were predicted in Med15 of the Tail that is likely to be the main sensor for regulatory proteins, while 6 α-MoRFs in Med13 of CDK is embedded in a region that hosts various trancriptional proteins (Table 1). Overall, the functional importance of 11 predicted α-MoRFs either as interaction sites or post-translational modification sites have been experimentally confirmed in yeast. In the cases of the Med7/Med21 [27] and the Med8/Med18/Med20 [28] complexes, structural data corroborate the role of the predicted α-MoRFs as recognition sites that adopt an α-helical structure in the bound state. Although less experimental data are available for human Mediator, 5 α-MoRFs predicted in Med1 fall into regions interacting with various transcriptional proteins (Table S2). For example, the N-terminal 306 residues of Med1 is involved in the transactivator function of BRCA1 [67], while the 433–803 region (with 4 predicted α-MoRFs) hosts the nuclear receptor LXRb and KIF1a [68].
So how does intrinsic disorder contribute to the function of Mediator? IDRs represent an ensemble of conformations [69] that imparts extreme flexibility onto the complex. In response to regulatory signals IDRs can adopt different conformations [70] and thereby induce functional transitions. In this way they could contribute to the observed pleomorphism of Mediator. IDRs with multiple binding sites indicated by the MoRFs may provide a scaffold-like function and thereby can be important to organize the complex. IDRs can also serve as malleable linkers between globular domains and may underlie modular functionality of the Mediator complex that enable it to interpret different combinations of transcriptional inputs [71]. IDRs can also facilitate assembly/disassembly of large complexes [37], for example association of Mediator with TFIID triggers assembly of the PIC. IDRs can be involved in complex signaling events [72] due to their adaptability. The same IDR can accommodate different partners [73] that may exert different, even opposite outcomes on transcription [74]. For example, the disordered N-terminal region of Med3 can host both Gcn4 and Tup1 proteins [51], or the C-terminal 100 residues of Med19 are involved in both transcriptional activation and repression [75]. IDRs are also preferred environments for post-translational modification sites [33] that provide a further regulatory tool for the Mediator complex (cf. T237 in Med4 [55]).
The presence of disordered regions also highlight an evolutionary aspect of Mediator's function. We observe that the propensity of disordered regions as well as the number of embedded interaction sites increases from yeast to man. This not only argues for an integral role of IDRs in Mediator's function, but may explain why the human Mediator is capable of processing a significantly larger number of regulatory signals (eg. the number of transcription factors increase by one order of magnitude from yeast to man [76]). Even if IDRs are conserved, as it was demonstrated by their similar arrangements in Saccharomyces cerevisiae and Homo sapiens their sequences are tolerant to substantial changes as long as the amino acid composition is biased for disorder [58],[66]. Only sequences of short segments that serve as recognition sites need to be restrained, as seen in case of 6 α-MoRFs [21]. On the other hand it is very easy to turn on and off the functionalities carried by these short motifs [45].
In conclusion, we propose that conserved intrinsically disordered regions contribute to the gene-specific regulatory function of the Mediator. IDRs with weak sequence restraints can provide an evolutionarily economic solution for the Mediator to handle a steadily increasing amount of complex regulatory signals. These results argue for the functional conservation of the Mediator and may account for the evolution of its regulation complexity.
Mediator protein sequences were extracted from the UniProt and NCBI databases using a large number of Mediator subunit names. Overall 556 sequences were identified out of which the redundant ones above 90% identity were removed by the CD-hit program [77]. In addition, a PSI-BLAST [78] search was performed using the 196 sequences from 10 organisms in the reference [21]. All resulting sequences were assembled in the MED_ALSEQ database that contained 340 sequences of 30 Mediator subunits derived from 27 eukaryotic organisms (Table S1). The corresponding randomized sequences (50 times each) were collected in the MED_ALRAN database. As a nomenclature for the Mediator subunits we adopted the unified convention proposed in reference [79]. Med19 and Med26 was assigned to the Middle module according to the reference [80].
Intrinsic disorder preferences of sequences in the MED_ALSEQ and MED_ALRAN databases were predicted at amino acid level using the IUPred (http://iupred.enzim.hu) [26] and PONDR VSL1 [25] algorithms. Intrinsically disordered segments were defined as regions with more than 30 subsequent residues with predicted disorder above 0.5, allowing a maximum of 3 residue long ordered gaps. MoRFs were computed using the reported algorithm [47]. Likely phosphorylation sites were identified using the DisPhos program [33].
The fractional difference is calculated as (CX−Cordered set)/Cordered set, where CX is the averaged content of a given amino acid in a protein set and Cordered set is the corresponding averaged content in a set of ordered proteins from the PDB.
Due to the presence of low-complexity regions, an iterative PSI-BLAST [78] based profile generation algorithm was performed to align full-length sequences of Mediator proteins [59]. Groups of homologous sequences were defined based on mutual sequence similarity (below the treshold of E = 10−5) between all members of the group. The final multiple alignment was generated by the CLUSTALW algorithm [81] using the BLAST profiles extracted from sequence groups. The performance of the alignment as compared to previous alignments [21],[27] are presented in Tables S3 and S4.
The sequence conservation of the Mediator proteins was evaluated comparing individual amino acid types (AAcons) using a simple Sum-of-Pairs (SP) score formula [82]. The score was 1 if identical residue was present in each positions of the alignment, otherwise it was 0 and these scores were averaged over the entire sequence.
Similarity between patterns of disordered and ordered regions was assessed using accuracy measures of secondary structure predictions [59],[60]. The overlap between ordered and disordered motifs (excluding gap positions) at residue level (Q) was characterized by the accuracy matrix defined as Q2 = 100 (MOO+MDD)/N, where MOO and MDD are the number of positions associated with the same motif type. Overlap between the segments were computed aswhere S1 and S2 stand for segments in two distinct sequences, respectively, minov(S1; S2) is the length of the overlap between S1 and S2, maxov(S1; S2) is the total extent of S1 and S2 in the given conformational state and len(S1) is the length of the segment in the reference sequence. δ(S1; S2) is the minimum of [(maxov(S1; S2)–minov(S1; S2); minov(S1; S2); int(len(S1)/2); int(len(S2)/2)]. The normalization factor N is given by the number of residues in conformational state i and the second summation runs over all M conformational states. Q and SOV values obtained for each possible pair within a given group of aligned sequences were averaged. The significance of the results was probed against the overlap values computed on the MED_ALRAN database.
|
10.1371/journal.pcbi.1000126 | Isolation-by-Distance and Outbreeding Depression Are Sufficient to Drive Parapatric Speciation in the Absence of Environmental Influences | A commonly held view in evolutionary biology is that speciation (the emergence of genetically distinct and reproductively incompatible subpopulations) is driven by external environmental constraints, such as localized barriers to dispersal or habitat-based variation in selection pressures. We have developed a spatially explicit model of a biological population to study the emergence of spatial and temporal patterns of genetic diversity in the absence of predetermined subpopulation boundaries. We propose a 2-D cellular automata model showing that an initially homogeneous population might spontaneously subdivide into reproductively incompatible species through sheer isolation-by-distance when the viability of offspring decreases as the genomes of parental gametes become increasingly different. This simple implementation of the Dobzhansky-Muller model provides the basis for assessing the process and completion of speciation, which is deemed to occur when there is complete postzygotic isolation between two subpopulations. The model shows an inherent tendency toward spatial self-organization, as has been the case with other spatially explicit models of evolution. A well-mixed version of the model exhibits a relatively stable and unimodal distribution of genetic differences as has been shown with previous models. A much more interesting pattern of temporal waves, however, emerges when the dispersal of individuals is limited to short distances. Each wave represents a subset of comparisons between members of emergent subpopulations diverging from one another, and a subset of these divergences proceeds to the point of speciation. The long-term persistence of diverging subpopulations is the essence of speciation in biological populations, so the rhythmic diversity waves that we have observed suggest an inherent disposition for a population experiencing isolation-by-distance to generate new species.
| A commonly held view in evolutionary biology is that new species form in response to environmental factors, such as habitat differences or barriers to individual movements that sever a population. We have developed a computer model, called EvoSpace, that illustrates how new species can emerge when a species range becomes very large compared with the dispersal distances of its individuals. This situation has been called isolation-by-distance because remote parts of the range can take different evolutionary paths even though there is no particular place where we would expect different populations to separate. When the extent of genetic difference between individuals is coupled with decreasing offspring viability (e.g., resulting from developmental problems), EvoSpace predicts that sharp spatial boundaries can emerge in arbitrary locations, separating subpopulations that occasionally persist long enough to become reproductively incompatible species. The model shows an inherent tendency toward spatial self-organization, in contrast with the traditional view of environmentally forced origins of new species. We think that isolation-by-distance is a common aspect of the evolutionary process and that spatial self-organization of gene pools may often facilitate the evolution of new species.
| The most common framework for understanding the process of biological speciation is geographical. For example, instances of speciation are typically allocated among three categories based on the extent of geographical separation between the daughter species. Allopatric speciation, in which a species range becomes severed and leads to population fragments that are not linked by gene flow, has been viewed as the most common means of speciation [1]. This process is easy to understand, because the independence of evolutionary processes (mutation, drift, selection) in populations that no longer communicate with one another would inevitably lead to reproductive incompatibility between such populations given enough time in isolation. Genetic incompatibilities are thought to accumulate in isolated subpopulations as first described by Dobzhansky [2] and Muller [3], and allopatric speciation has been modeled based on these ideas [4],[5]. Not only does genetic isolation between subpopulations simplify genetic modeling of speciation, it is also relatively easy to observe the “fingerprints” of allopatric speciation in many instances, such as the endemism of terrestrial species on islands (e.g., Darwin's finches [6]).
The two other geographical categories of speciation involve divergence between subpopulations in the face of gene flow, and it has been less clear what the compelling “fingerprints” of these processes might look like when observed after the fact. In the second category, parapatric speciation, one species becomes two where the daughter species occupy contiguous ranges. This has most often been modeled as a consequence of habitat variation and divergent local adaptation by subpopulations e.g., [7]–[10]. Sympatric speciation, in which the ranges of the daughter species overlap, has similarly been modeled as a consequence of microhabitat variability and niche-partitioning e.g., [10]–[12]. Models of sympatric speciation suggest that the efficiency of specializing in the exploitation of discretely different resources can favor the formation of two species over the maintenance of a single generalist species.
A common theme among all three of these categories is that speciation is induced by divisive, external factors and that the inherent tendency of biological populations is to remain unified in the absence of these factors. In other words, the conventional wisdom is that it is the environment that tears species apart, and that in the absence of local dispersal barriers or environmental heterogeneity, biological populations tend to sustain functional and genetic cohesion. One well-known, but rare, situation where this view breaks down is in the case of ring species [13]–[15]. The range of a ring species extends around some sort of environmental obstacle until the two ends of the range meet. If one were to sample the gene pool starting at one end of the distribution moving around the obstacle to the other end, the gene pool would become increasingly different from the starting point with distance traveled, as expected given relatively short dispersal distances within species with large ranges (isolation-by-distance [16]–[20]). Where the two ends meet, however, individuals with the greatest genetic differences within the species come into contact. If they are so different that they do not or cannot mate with one another, these local groups appear to be different species. This produces an enigma if local matings happen all the way around the obstacle, because the directly incompatible ends of the range still remain indirectly connected by gene flow. It is hard to say whether ring species represent one or two species, but these instances illustrate the potential for functional decoherence (speciation) under isolation-by-distance. This possibility, without the presence of obstacles to dispersal, is the focus of this study.
Isolation-by-distance can also lead to symmetry-breaking in the distribution of genetic variation across the species range, leading to the emergence of discretely different and spatially segregated subpopulations [21]–[28]. The formation of distinct, but reproductively compatible, subpopulations typically precedes speciation, so the inherent tendency of subpopulation emergence under isolation-by-distance further suggests the potential for autonomous speciation when dispersal distance is short relative to the species range. In this paper, we describe a model of a spatially extended biological population (isolation-by-distance) in the absence of both obstacles to dispersal and environmental heterogeneity, which suggests that biological populations inherently and regularly tend to tear themselves into reproductively incompatible daughter species.
Most previous spatial models of speciation have assumed predetermined subdivisions (e.g., island model or stepping-stone model), habitat variation-inducing localized selection differences, or both [4], [5], [29]–[31]. These external factors impinging on a population model constrain or determine the resulting spatial patterning of the gene pool. In contrast, subdivision resulting from isolation-by-distance alone is an organic consequence of the system's dynamics. Because the only evolutionary forces assumed by our model are mutation, recombination, dispersal, and outbreeding depression, it serves as a proof of concept that internal population dynamics can generate spatial subdivision of a gene pool, even to the extent of parapatric speciation.
We have implemented a generalized cellular automaton model of evolutionary processes, called EvoSpace. The simulated population was distributed across an N×N grid, where cells were either unoccupied or occupied by one individual. An individual contained genetic information in the form of a set of chromosomes and could migrate and mate with other individuals within a certain distance. A chromosome consisted of a string of characters from the set {A, C, G, T} representing the nucleotide bases. The number and length of chromosomes were the same for all individuals. During mating, an offspring was constructed by randomly selecting and combining two haploid genomes from the two diploid parents, possibly introducing random mutations in the process. Thus reproduction in our model was sexual, because genomes from two parents were combined to produce the offspring and the two genomes within a parent exhibited recombination through the independent assortment of chromosomes during gamete formation; but individuals were also hermaphrodites, as any adult could potentially mate with any other adult.
At every generation (time step in the model), migration, mating, and mutations created genetically distinct offspring, but resulted in only minor changes in the spatial structure of population genetic variation. First, individuals could randomly move on the grid world within bounds determined by a dispersal distance. Then, they could choose a mate within similar bounds. Finally, after mating, each offspring was placed in a random cell in the vicinity of one of the parents while the parents died, so that only one generation lived on the grid at each time step. The mean reproductive rate in the population was regulated each generation to buffer swings in population size resulting from a variety of factors, such as stochastic mortality (see below). This was achieved by randomly removing individuals or generating additional offspring, reflecting a constant carrying capacity of the environment. Simulations began with a population of genetically identical individuals, amounting to 80% of the grid cells (20% of locations remained empty space), and the system was allowed to evolve for hundreds to millions of generations.
For the experiments described in this paper, the habitat across the grid environment was homogeneous, so location did not influence the fitness of individuals. However, we introduced one important dependency: the offspring's survival probability was a decreasing function of the genetic difference between merging gametic genomes. Expressing the genetic difference between two chromosomes as a fraction between 0 (nucleotide identities at all positions of the DNA sequence were identical) and 1 (nucleotide identities at all positions of the DNA sequence were different), offspring resulting from gametes with a genetic difference greater than a threshold θ had zero survival probability (we used θ = 0.6 in this study; see Materials and Methods). Conversely, gametes with a genetic difference less than θ0 were 100% compatible (we used θ0 = 0.05 in this study). The negative relationship between gamete genomic difference and offspring viability was a simple representation of Dobzhansky-Muller reproductive incompatibility. Orr [4] reasoned that the Dobzhansky-Muller outbreeding depression function would decline exponentially, rather than linearly, but our linear function provided a conservative approximation in this context: it required diverging subpopulations to persist for a much longer period of time as they absorbed the demographic cost of decreasing viability of hybrid offspring. We also tested a nonlinear alternative in the form of a truncated Gaussian distribution with a peak at 1%, which roughly traced the decline in our linear function. Not only did this eliminate the potential for artifacts associated with the angles of our “broken stick” function, but it also imposed a slight amount of inbreeding depression for gametic genomes that were too similar. The results from the Gaussian outbreeding depression function were qualitatively the same as those reported here for the linear function.
The outbreeding depression function was central to the exploration of speciation in this model, as reproductive isolation between sexual species lies at the core of the concept of speciation. It was, in fact, the fundamental criterion embodied in the definition most commonly assumed in the context of evolutionary biology: the Biological Species Concept (BSC) [1]. Our rule for reproductive isolation between species was somewhat more restrictive than the BSC requires, because real species could be genetically compatible, but behaviorally or morphologically incompatible. For a comprehensive discussion of reproductive compatibility functions in speciation models, see [10].
Gene flow distances in the model resulted from a combination of factors: the dispersal of individual agents, the location of female mates, and the settlement of offspring. The rules governing these behaviors of agents (see Materials and Methods) yielded a distribution of single generation gene flow distances that rarely exceeded six cells in our 400×400 spatial matrix when δ = 1.5, and 20 cells when δ = 5 (Figure 1).
The shorter gene flow distances illustrated in Figure 1 generated a positive relationship between geographic and genetic distances, as described by the pattern view of isolation-by-distance (Figure 2). However, the distribution of points in Figure 2 seemed more informative than the slope of the regression line. Genetic surveys of real populations would not have the luxury of a sufficiently random sampling of such a large number of genomes, so it may be difficult to ascertain the distribution of points as effectively as was done here. Comparing the behavior of the model with and without the implementation of outbreeding depression showed very similar regression slopes, but very different patterns of point clustering, an important feature that could be easily missed with genetic survey data.
To assess the pattern of genetic diversity in our model system, we measured the genetic difference between two randomly selected, haploid gametes, as though these gametes were about to merge in fertilization and produce more or less viable offspring. We then analyzed this information with mismatch distribution histograms [32] to reveal the frequency distribution of genetic differences among the genomes in the population(s). The horizontal axis was the genetic difference, and the vertical axis showed the number of pairs of gametes found with that degree of genetic difference (Figure 3). In this plot, a population of genetically random individuals appeared as a single distinct peak at 0.75 (with a small standard deviation), because of the 25% probability that two bases were identical at any particular position of the DNA sequence. In the case of a population of genetically identical genomes—the starting condition for our simulations—the plot showed a single sharp peak at 0. As mutation led to genetic divergence, peaks traveled to the right in the mismatch distribution.
Observing time-sequence movies of these mismatch distributions under different model conditions illustrated the spatiotemporal patterns of gain and loss of genetic diversity, especially as it revealed the origin and existence of distinct subpopulations (traveling waves along the distribution). The series of snapshots in Figure 4 provides a glimpse into the dynamics of these systems. When the population was effectively well mixed (Figure 4C; which is achieved here with δ = 5), genetic differences within the population did not grow far beyond the θ0 = 0.05 threshold, where outbreeding depression began to impact offspring viability. Under these conditions, the population mixed across the grid rapidly enough to remain a single, genetically coherent population. This was represented as a distinct and stable peak in the histogram at a genetic difference level of 5%. No pair of genomes was found with a genetic difference greater than 10% when δ = 5. All aspects of the model's behavior described here were repeatable for different runs of the model under the same conditions.
With δ = 1.5, but in the absence of outbreeding depression (Figure 4B), the single initial peak on the histogram centered on 0 spread and moved to the right as mutation created genetic variation. When the space was large enough, this primary peak became centered on a genetic difference of 75%, the maximum expected under the Jukes-Cantor mutation model [33]. For certain combinations of grid size (sufficiently large), dispersal distance (sufficiently short), and mutation rate, however, additional dynamical patterns emerged. We were particularly interested in tracking the diversity waves described by Rogers and Harpending [32]. Indeed, small peaks arose at low levels of genetic difference (left side of the mismatch distribution), moved to the right, and often persisted long enough to merge with the primary peak. This observation is consistent with previous findings on spatial self-organization under isolation-by-distance in the absence of outbreeding depression [21], [22], [34]–[40].
The patterns that we detected in intraspecific dynamics were greatly enhanced by combining the outbreeding depression function with the shorter dispersal distance (Figure 4A). Along with sharpening the degree of spatial organization that emerged (see next section), outbreeding depression strongly increased the amplitude and separation of the secondary peaks appearing in the mismatch distributions. It also firmly established a large peak centered on θ0 (5% in this case), representing all within-subpopulation comparisons. In effect, each emergent subpopulation functioned like a single panmictic population. The traveling waves peeled off this peak as existing subpopulations divided. As these peaks moved to the right of θ0, the subpopulations being compared experienced an increasingly stringent demographic disadvantage, because when individuals from diverging subpopulations mated with each other, their offspring were decreasingly viable. Divergence continued as the demographic cost of outbreeding increased. Nevertheless, sometimes a peak became established to the right of θ. This represented a set of comparisons between gametes from reproductively isolated subpopulations, since their hybrid offspring could not be viable anymore. In summary, we interpret the traveling waves to reflect discrete genetic subpopulations, new species when they persisted past θ = 0.6, emerging through the spatial microevolutionary dynamics within a population.
We have further found that the development of new, stable peaks to the right of θ (new species, we argue) was quite sensitive to the interrelationships of the spatial scale of the simulation (grid size and dispersal distance), the mutation rate, and other factors. For example, if the mutation rate was too high, overall genetic diversity increased rapidly until it was too hard to find viable mating pairs within the mating neighborhood and the whole population went extinct. If the mutation rate was too low, the degree of genetic difference generated between subpopulations did not reach the threshold of speciation before at least one of the subpopulations went extinct. We are working now to examine systematically the likelihood that biological populations would evolve within the region of phase space associated with these interesting dynamics.
The mismatch distribution provided good insight into the existence of distinct, internally homogenous subpopulations, but it did not demonstrate whether clusters were spatially segregated on the lattice or show where they were located. Two other analytical tools, isolation-by-distance scatterplots (Figure 2) and genetic cluster plots (Figure 5), were useful in examining the spatial clustering of distinct subpopulations.
Genetic cluster plots were obtained by grouping sets of individuals for which all genetic distance relations were lower than a given level and displaying those groups in different colors (Figure 5). These plots clearly illustrated the spatial self-organization that emerged in our model. Visual examination indicated that the genetically homogeneous subpopulations revealed by the histogram plots also formed distinct spatial clusters that occupied coherent, non-overlapping regions of the lattice. The borders between these regions tended to be unoccupied, or occupied with hybrid genomes that were not genetically similar enough to any neighboring region to be classified with them.
Spatial plots such as isolation-by-distance and genetic clustering also demonstrated the strong sharpening effect of outbreeding depression. The intrinsic tendency of the grid world toward spatial order was greatly enhanced by introducing dependence of viability on similarity. It could be said that outbreeding depression played a negative feedback role analogous to long-range inhibition in morphogenetic reaction-diffusion processes [41],[42]. In this analogy, the combination of mating and gene flow played the positive feedback role of short-range activation. Together, these effects contributed to the spontaneous formation of “spots” by encouraging neighboring elements to be similar and, at the same time, distant elements to be different. In a sense, our model represents “evolutionary pattern formation” at the scale of populations of organisms, instead of morphogenetic pattern formation at the scale of tissues of cells.
Spatially explicit computational models of evolution are a relatively recent development made possible by the rapid rise in the power of computing hardware, although the earliest studies date back to the 1970s [21]. A pervasive, and perhaps universal, behavior exhibited by these models is the tendency for heritable variation to become spatially segregated in a process of self-organization e.g., [22], [27], [34]–[40]. Our model also exhibited this phenomenon, as we expected. It is important to recognize this inherent tendency for spatial diversification and the natural ways in which this would facilitate speciation, even if the model presented here does not represent a complete description of any particular speciation event. It should also be noted that the spatial dynamics of speciation in the absence of dispersal barriers is significantly more complicated than the issue of how Dobzhansky-Muller incompatibilities accumulate in isolated populations [4],[5], although they share the endpoint of completed speciation. Connected populations can become disjunct, creating the opportunity for allopatric speciation, but sexual recombination and dispersal make spatially-extended genetic networks the essence of sexual populations. This is also the basis of EvoSpace, which is used here to provide a spatially connected context for studying the evolution of reproductive incompatibilities among emergent (not assumed) subpopulations.
The dominant geographic paradigm for classifying modes of speciation recognizes three general categories: allopatric, parapatric, and sympatric speciation. The allopatric mode has long been widely thought to represent the most commonly realized mode [43]. The sympatric mode of speciation has also received much attention over the past 50 years or so, although it has long been considered controversial [43]. Recently, however, it seems that the potential for sympatric speciation has become more widely appreciated and some compelling empirical examples have come to light e.g., [44]. Parapatric speciation has received less attention (but see [10]), perhaps because it is something of a hybrid between the other two. It requires speciation in the face of gene flow, which is the same hurdle that must be overcome in achieving sympatric speciation, yet it involves subpopulations and sibling species that have essentially non-overlapping ranges, as in allopatric speciation. Models of parapatric speciation have typically involved environmental variation applying divergent selection pressures to different parts of the species range, with that reinforcing selection favoring positive assortative mating within the hybrid zone completes the speciation process [7]–[10]. In contrast, our model can illustrate a process of parapatric speciation in the absence of environmental variation and preferential mating. By including functional outbreeding depression, we have invoked an endogenous kind of selection that is independent of the external environment. Thus this is a model of speciation through the spatial self-organization of the gene pool.
Our model is similar to the neutral model developed by Gavrilets and colleagues, although there are fundamental differences [29]–[31]. Both approaches account for the entire process of speciation, from genetic homogeneity to reproductive isolation, but a key difference is represented in the geographic assumptions of the models. The neutral models of Gavrilets [29]–[31] assume a discretely subdivided population, connected by migration, at the start. Our model assumes a population with absolutely no predetermined subdivision or barriers that would divide subpopulations, yet it is able to self-organize into discrete subpopulations that can then evolve reproductive isolation. It is interesting that these two kinds of models behave in similar ways. For example, both models show that speciation is possible even in the presence of gene flow, and both models show that local adaptation is not necessary to generate reproductive isolation. Another important difference between these models regards the shape of the outbreeding depression function. Gavrilets assumes a step-shaped function where offspring viability is either 0 or 1, depending on the extent of genetic difference between gametes. In this way, reproductive isolation between individuals happens as a byproduct of a single mutation: the one that pushes the genetic difference between two individuals over the reproductive incompatibility threshold. In our model, reproductive incompatibility accrues by degree, requiring subpopulations gradually to take on the increasing demographic cost of more failed reproductive opportunities as the speciation process unfolds.
Several papers by Sayama, Bar-Yam, and colleagues [27],[28],[34],[35] have emphasized the role of spatial dynamics in the spatial patterning of gene pools, but the model that they have developed does not include a mutation process and assumes an artificial fitness function. Their model envisions two compatible sets of alleles across loci, and mixing alleles across these sets is assumed to result in decreased fitness. While there is heuristic value in observing how allelic incompatibilities sort themselves in space, it is hard to imagine how allelic variation with these features might evolve in the first place. This combination of assumptions results in a population with two distinct genotypes, each assumed to be internally compatible, with clumped distributions in space. While spatial self-organization is evident in this model, demographic stochasticity would ensure that it ultimately evolves to complete homogeneity in the absence of mutation. EvoSpace allows genetic incompatibilities to arise organically through mutation and lineage proliferation within the rules of the model. The occurrence of mutation within EvoSpace also allows the population at large to accumulate and organize genetic diversity to an extent that is sustainable under the model.
Under the set of parameter values explored here, our model suggests that mutation alone can drive genetic divergence between subpopulations to the point of overwhelming the constraints imposed by (limited) gene pool mixing and outbreeding depression under isolation-by-distance. The emergence of sharp boundaries between genetic subpopulations is consistent with previous computational models incorporating isolation-by-distance [22], [27], [34]–[40], although the natural tendency for spatial self-organization in population genetics has not yet been fully appreciated by the community of population geneticists. Indeed, there has been a long-standing confusion in the literature between the notions of isolation-by-distance as a process or as a pattern [45]. Sewall Wright [16] originally conceived of isolation-by-distance as a model of the evolutionary process in which
The presumed pattern of isolation-by-distance is a smooth, monotonically increasing relationship between geographic distance and genetic distance [17]–[20], which does not anticipate sharp transitional boundaries between internally homogeneous, divergent subpopulations. This expectation was based on mathematical models of Wright's [16] process view of isolation-by-distance that were not able to predict emergent population substructure because they relied on mean field approximations of spatial context that did not represent spatial configurations. Therefore, we advocate a return to Wright's original view of isolation-by-distance as part of the evolutionary process characterized by relatively short dispersal distances within an extensive population range and open-mindedness to the possibility that isolation-by-distance alone can result in the emergence of spatially bounded subpopulation structure.
The model of evolutionary genetics presented here is a very simple and generic one. It does not depend on idiosyncratic forms of selection or particular population structures. Instead, it is based on fundamental and common building blocks of biological populations (chromosomes and sexual individuals) that are stochastically affected by mutation, mortality, reproductive success, and dispersal. The interesting behavior of the model emerges dynamically due to the constraints of isolation-by-distance and outbreeding depression. Therefore, we conjecture that the tendency for spatial self-organization and parapatric speciation may occur universally in biological populations. This is not a claim that our model is the exclusively correct model of speciation; rather, we are suggesting that the dynamic of diversification illustrated by our model may exist even under conditions that suppress the realization of emergent population substructure, such as great dispersal distances in relatively small species ranges. We expect that the inherent tendency for spatial diversification amplifies the effects of environmental heterogeneities, and we plan to explore this interaction with further developments of EvoSpace. Thus we do not deny that habitat variation and dispersal barriers can play important roles in instances of speciation, but we think that the inherent dynamic identified here may generally drive diversification/speciation in a way that is molded to these external constraints.
In conclusion, our model reveals an aspect of intraspecific evolutionary dynamics that emerges when dispersal distances are sufficiently short relative to a species range. Localized subpopulations regularly form and diverge from one another while maintaining their identities as clusters of genetically similar individuals. Subpopulations that grow so large as to embody too much genetic diversity tend to subdivide through spatial segregation, just as the original population does in our model. If the degree of outbreeding depression grows as gametes' genomes become increasingly different, the pattern of genetic and spatial population subdivision becomes better defined, and some subpopulations can diverge to the point where they complete a process of parapatric speciation. The behavior of our model suggests that spatially extended populations regularly generate new subpopulations, each of which takes a path of genetic divergence with the potential of becoming a new species. Most of these embryonic subspecies become extinct before emerging as reproductively independent species, but the internal dynamic of this model constantly potentiates the production of new species.
The simulated world is an N×N grid that wraps around north-south and east-west, creating a torus. Typical grid sizes are 100×100 to 500×500 cells (see samples in Figure 5). Smaller grids run faster and consume less memory, allowing flexible parameter exploration, while larger grids allow investigation of larger-scale spatial isolation effects. Each grid cell is constrained to contain a maximum of m occupants. For the results reported here, we use m = 1 (i.e., 0 or 1 occupant per cell). Grid coordinates are Cartesian pairs r = (x, y), and distances between cells are simple Euclidean distances from the cell centers. Thus each cell has exactly four neighbors at distance 1, four more neighbors at distance √2, and so on. A simulation parameter δ correlates with the distance within which individuals can move, mate, or appear in a single generation. As the value of this parameter increases, so too does the mobility of the individuals within the grid world. Movements of individual genomes, carried by either diploid individuals or haploid gametes, occur in three stages: migration, mating, and offspring placement. (a) To begin with, each individual migrates to another site randomly determined by the combination of a Gaussian and a uniform probability distribution. Denoting by r the starting location of an individual, its new location r′ after migration is calculated in two steps. First, a distance d is drawn from a Gaussian distribution of mean and standard deviation δ: Gδ(d) = κ exp[−(d−δ)2/δ2], where κ is a normalization coefficient. Second, the location r′ is drawn from a uniform distribution within a disc D of radius d centered around r: Pmig(r′|r) = Dd(r′−r), where Dd(u) = 1/πd2 if ||u||≤d and 0 otherwise. (Additionally, r′ is corrected to fall on the nearest integer grid location.) (b) Then, each individual is considered in turn to act as the “father” and sends a gamete to a potential “mother” r″ at a location chosen uniformly randomly in a circle of radius δ centered around the father's location r′: Pmat(r″|r′) = Dδ(r″−r′). (c) Finally, a newly formed offspring (see next section) settles into a grid cell r′′′ drawn uniformly randomly in a circle of radius 2δ around its mother's position Poff(r′′′|r″) = D2δ(r′′′−r″).
Each individual contains its personal genetic information as a diploid set of chromosomes. A chromosome consists of a string of characters that can take one of four values representing the nucleotide bases (A, C, G, or T). The number n and length l (number of base pairs) of chromosomes are the same for all individuals and are both set at the start of a simulation; n = 2 and l = 200 for all results presented here. Reproduction in our current model is sexual, because genomes from two parents are combined to produce the offspring, but individuals are also hermaphrodites able to function as either the male or female in a sexual encounter. An offspring is constructed from two parents as follows. Each parent produces a haploid gamete by randomly selecting one of the chromosomes from each diploid pair (Figure 6). These haploid genomes are combined in the offspring to produce a diploid individual. Thus we have not allowed for crossing over within chromosomes, but non-homologous chromosomes assort independently during sexual recombination. During the transcription from each parent to its gamete, some random mutations in the form of single base substitutions are introduced according to a uniform and constant probability μ defined at the start of the simulation; μ = 0.00005 mutations/site for results given in this paper.
The genetic difference H between the two merging gametic chromosomes in a diploid pair is expressed as the fraction (between 0 and 1) of all the base pairs containing different bases: H = B/nl. Thus H = 0 means that all bases in the child's gametes #1 and #2 are identical at every position, and H = 1 means they all differ at every position. The offspring's survival probability S can be set to be dependent on this genetic difference, according to the curve in Figure 7, at the beginning of a run. Given two threshold values of genetic difference θ0 and θ, such that 0≤θ0≤θ≤1, we set S = 1 for H≤θ0, S = 0 for H≥θ, and S = (θ–H)/(θ–θ0) for θ0<H<θ. Thus offspring composed of gametes with a genetic difference H greater than θ are nonviable and immediately removed from the grid at birth. In general, two random genome sequences are expected to have a genetic difference H = 0.75, since each nucleotide position would have a 25% chance of being occupied by the same base [33].
The simulation begins at generation 0 with a population of genetically identical individuals. The initial population randomly fills a predetermined fraction of the maximum grid occupancy. For the experiments described in this paper, the grid occupancy fraction is 80%. For example, a grid of 400×400 = 160,000 cells starts with 128,000 individuals.
The creation of the next generation for the simulation is a three-phase process. (a) First, each individual migrates to a random cell r, according to the probability Gδ described above. (b) Then, individuals reproduce. The algorithm iteratively considers each member of the population once to be the “father” in a mating. For each father, a list of potential mates is created from those individuals in all cells within a radius δ of the father, and one of those individuals is randomly selected to be the “mother.” Mating then proceeds as described above. The number of potential offspring resulting from each mating is given by a Poisson distribution with mean 1. (c) Finally, each offspring is placed in a random cell within a radius 2δ of the mother. If no cell within this range has room for the offspring according to the maximum per-cell occupancy m, then some or all offspring of this mating could be lost. No parents survive into the next generation, so overcrowding can be the result only of offspring from matings that have already taken place during earlier processing of the current generation. Since each individual is the father in exactly one mating, and the mother in an average of one mating, and each mating has an average of one offspring, the population size should stay about the same from generation to generation. Certain factors, however, could result in the population size shifting from the target size, which is addressed in the next phase.
In a supplementary step, adjustment, the algorithm makes an effort to keep the population close to the initial target population size. This may require additional “make-up” births from randomly selected parent pairs, if offspring viability reduced the population size due to high genetic difference of parents, or if localized overcrowding resulted in the loss of some offspring. More rarely, if the population ends up above the target level, the algorithm randomly selects individuals for culling.
Given any two diploid individuals on the grid, we draw one haploid genome from each individual by randomly selecting one chromosome in each diploid pair they contain (a process identical to constructing a haploid gamete for that individual). The genetic distance (as plotted in Figure 1) is then defined by counting the number of mismatched base pairs between these two haploid genomes and dividing by the total number of base pairs. Due to the random assortment of chromosomes, this quantity is not necessarily the same every time it is computed for a given pair of individuals. We also use the term “distance” without verification that the triangle inequality holds.
A mismatch distribution reveals the shape of genetic diversity in a sample through pairwise comparisons of genetic differences, which are displayed in a frequency histogram [32] (Figure 3 and Figure 4). The number of peaks on the histogram does not directly correspond to the number of genetically self-similar subpopulations. For example, in the case of a population with three genetically distinct groups, A, B, and C, where A and B have an average genetic difference of 60%, A and C of 40%, and B and C of 20%, the histogram would show these three peaks plus one centered on θ0 (0.05 for the results presented here) representing the within subpopulation comparisons. However, if clusters B and C also happened to be 0.4 distant, two peaks would be superimposed and would obscure the number of subpopulations. A peak typically represents a set of comparisons between two subpopulations defining the degree of divergence between the subpopulations. The peaks move to the right in the mismatch distribution as long as both subpopulations persist and evolve along different trajectories, and the peaks stop moving when divergence hits the maximum value of 75% expected under the Jukes-Cantor mutation model [33].
Scatterplots with geographic distance on the x-axis and genetic distance on the y-axis (Figure 2) are commonly presented as a way to examine isolation-by-distance in data from spatial genetic surveys. It is expected that shorter dispersal distances will yield a steeper slope for this relationship, although the relationship may not be linear [46]. An effectively well-mixed population should show no relationship between geographic and genetic distances, because thorough mixing would randomize location with respect to genotype.
A genetic cluster plot (Figure 5) is constructed by randomly selecting pairs of haploid genomes and computing their genetic distance, similarly to the mismatch distributions, then only retaining pairs that are distant below a certain level set by the user. When a sufficient number of pairs with low distances have been gathered, we build a nondirected graph that contains the individuals as nodes and edges representing genetic distances smaller than the threshold. By analyzing this graph we can then identify regions of self-connected clusters. Since it is not computationally feasible to examine all pairs of individuals, the clustering might depend on the random selection of pairs of individuals (i.e., some clusters that were not connected in one graph construction might be connected in another). We have found, however, that repeated application of our clustering algorithm with different random seeds (leading to different pairs being examined) leads to qualitatively identical results. On the other hand, reducing the distance threshold has the expected effect of connecting formerly disconnected genetic clusters.
|
10.1371/journal.pntd.0003037 | Comparative Host Feeding Patterns of the Asian Tiger Mosquito, Aedes albopictus, in Urban and Suburban Northeastern USA and Implications for Disease Transmission | Aedes albopictus is an invasive species which continues expanding its geographic range and involvement in mosquito-borne diseases such as chikungunya and dengue. Host selection patterns by invasive mosquitoes are critically important because they increase endemic disease transmission and drive outbreaks of exotic pathogens. Traditionally, Ae. albopictus has been characterized as an opportunistic feeder, primarily feeding on mammalian hosts but occasionally acquiring blood from avian sources as well. However, limited information is available on their feeding patterns in temperate regions of their expanded range. Because of the increasing expansion and abundance of Ae. albopictus and the escalating diagnoses of exotic pathogens in travelers returning from endemic areas, we investigated the host feeding patterns of this species in newly invaded areas to further shed light on its role in disease ecology and assess the public health threat of an exotic arbovirus outbreak.
We identified the vertebrate source of 165 blood meals in Ae. albopictus collected between 2008 and 2011 from urban and suburban areas in northeastern USA. We used a network of Biogents Sentinel traps, which enhance Ae. albopictus capture counts, to conduct our collections of blooded mosquitoes. We also analyzed blooded Culex mosquitoes collected alongside Ae. albopictus in order to examine the composition of the community of blood sources. We found no evidence of bias since as expected Culex blood meals were predominantly from birds (n = 149, 93.7%) with only a small proportion feeding on mammals (n = 10, 6.3%). In contrast, Aedes albopictus fed exclusively on mammalian hosts with over 90% of their blood meals derived from humans (n = 96, 58.2%) and domesticated pets (n = 38, 23.0% cats; and n = 24, 14.6% dogs). Aedes albopictus fed from humans significantly more often in suburban than in urban areas (χ2, p = 0.004) and cat-derived blood meals were greater in urban habitats (χ2, p = 0.022). Avian-derived blood meals were not detected in any of the Ae. albopictus tested.
The high mammalian affinity of Ae. albopictus suggests that this species will be an efficient vector of mammal- and human-driven zoonoses such as La Crosse, dengue, and chikungunya viruses. The lack of blood meals obtained from birds by Ae. albopictus suggest that this species may have limited exposure to endemic avian zoonoses such as St. Louis encephalitis and West Nile virus, which already circulate in the USA. However, growing populations of Ae. albopictus in major metropolitan urban and suburban centers, make a large autochthonous outbreak of an arbovirus such as chikungunya or dengue viruses a clear and present danger. Given the difficulties of Ae. albopictus suppression, we recommend that public health practitioners and policy makers install proactive measures for the imminent mitigation of an exotic pathogen outbreak.
| Aedes albopictus is one of the most invasive and aggressive disease vectors in the world. The range of this species is currently still expanding, particularly into highly dense human population centers in temperate areas in the USA and Europe, raising the public health threat of emerging and re-emerging diseases such as chikungunya and dengue. The prominence of Ae. albopictus as a major vector was exposed during the global pandemic of chikungunya virus, primarily because of a virus adaptation which enhanced the transmission efficiency by this mosquito species and also because of the first locally-transmitted cases of chikungunya virus in temperate Europe. Blood feeding patterns by mosquitoes are a critical component of virus proliferation and determine the degree and intensity of disease epidemics, particularly in newly invaded areas. We examined the blood meal sources of invasive Ae. albopictus in the northernmost boundary of their range in temperate North America and found that the species fed exclusively on mammalian hosts, with over 90% of their blood meals derived from humans and their associated pets (cats and dogs). The high mammalian affinity of Ae. albopictus suggests that this species may be an efficient vector of mammal-driven zoonoses and human-driven anthroponoses such as dengue and chikungunya viruses in this region.
| Understanding the blood feeding patterns of mosquitoes is of paramount importance in determining their vector status in the maintenance and epidemic transmission of arboviruses. Blood feeding patterns of mosquito vectors provide insight into the ecological transmission cycles of pathogens and lead to more efficient disease and vector control measures for the benefit of animal and human health. For invasive mosquitoes with expanding geographic ranges, such as Aedes albopictus (Skuse), the specific blood-hosts impact endemic diseases and can lead to the epidemic transmission of exotic pathogens.
The Asian tiger mosquito, Ae. albopictus, has dispersed extensively from its native tropical range in Southeast Asia and is now found on every continent except Antarctica [1], [2]. The last decade has seen a dramatic expansion of Ae. albopictus into temperate regions of Europe and North America [3]–[5]. In many parts of its expanded range, this species is implicated as a significant vector of emerging and re-emerging arboviruses such as dengue (DENV) and chikungunya (CHIKV).
Although historically not an important vector of CHIKV, Ae. albopictus has become the principal driver of recent epidemics in Asia and islands in the Indian Ocean because of a mutation in the virus envelope protein enhanced transmission efficiency by this species [6], [7]. Autochthonous transmission of CHIKV has also been recorded in temperate regions of Italy and France [8], [9] where invasive Ae. albopictus have become abundant [3]. Aedes albopictus was also the sole vector in local epidemics of dengue in Hawai'i and other regions [10], [11] and is a competent laboratory vector for at least 22 arboviruses [12]. Due to the widespread and increasing distribution of Ae. albopictus in temperate regions and the escalating diagnoses of exotic pathogens in travelers returning from endemic or epidemic areas [13], [14], the risk of an outbreak in a new area is no longer hypothetical. Furthermore, because this species thrives in artificial containers found in close association with human peridomestic environments, it is essential to fully investigate the host feeding patterns of Ae. albopictus in order to completely understand its role in disease ecology and public health significance.
Surprisingly, given the vector potential and medical importance of Ae. albopictus, few studies have been conducted to investigate the host feeding patterns of this species in its native and expanding geographic range. This is likely because adult Ae. albopictus are a difficult species to collect efficiently in traps, and blood fed specimens are especially rare. From the few studies that have been conducted, the precise host feeding preferences of Ae. albopictus seem to vary considerably (Table 1). The species has been generally reported to feed on a wide range of mammals including humans, but will also feed on avian hosts at various proportions, and has even been incriminated to feed on amphibians and reptiles [15]–[34]. It has thus been considered an opportunistic feeder and a classic bridge vector candidate between zoonotic arboviruses and humans. However, caution should be taken in labeling Ae. albopictus as an efficient bridge vector because the large variation in the feeding plasticity of this species questions the exact role that it may play as an enzootic or epidemic vector of arboviruses. For example, in its native tropical range, Ae. albopictus feeds exclusively on humans in Indonesia [35], whereas in Singapore it feeds on humans, oxen, and dogs [15]. Additionally, studies conducted in Thailand [36] have reported that Ae. albopictus feed on humans, swine, buffalo, dogs, and chickens, while more recent investigations [26] report that Ae. albopictus feeds only on humans, with a few (<6%) double-host blood meals between humans and swine/cat/dog. In temperate Japan, Ae. albopictus primarily feed on mammals, with a high propensity for humans, but also on birds and amphibians/reptiles [29], [30] (Table 1).
In temperate locations of the expanding range of Ae. albopictus, the host preference of this species is also variable. Studies conducted at a tire dump in Missouri, USA, reported that Ae. albopictus will feed on birds (17%) but prefer mammals (64%), with 8.2% of those mammalian feedings obtained from humans [19]. A follow up study conducted in other tire yards and surrounding vegetation of rural and urban habitats in Missouri, Florida, Indiana, Illinois, and Louisiana, USA, concluded that Ae. albopictus showed a strong preference for mammals (>94%), with up to 8% human-derived blood meals, while also detecting avian (1%) and reptilian (5%) blood meals [20]. An additional study in suburban landscapes of North Carolina, USA, reported that Ae. albopictus feeds predominately on mammalian hosts (83%), but also on birds (7%), amphibians (2%), and reptiles (2%) [27]. In Europe, Italian populations of Ae. albopictus rarely feed on birds in urban settings, while 99% of specimens have been reported to feed on mammals, with 90% of those mammalian blood meals being derived from humans [31]. The same investigators report that in suburban settings of Italy, 7% of Ae. albopictus had fed on avian species, while the vast majority of the blood meals were mammalian-derived (95%), with 43% containing human blood [31]. Finally, in urban zones of Spain, Ae. albopictus obtained blood meals exclusively from humans (100%) [32] (Table 1).
Although it is apparent that Ae. albopictus feeds predominantly on mammals, the degree of mammalophagic or anthropophagic host feeding preferences of this species appear location specific. Because of the rapidly expanding range of Ae. albopictus, its abundance in metropolitan centers, and its close association with humans in peridomestic habits, combined with the emergence and resurgence of exotic pathogens for which Ae. albopictus is a capable vector, it is clear that assessing its host feeding preferences in newly invaded areas is critical to elucidate disease transmission cycles and develop strategies to reduce the local risk of an exotic arbovirus outbreak. However, the collection of Aedes (Stegomyia) spp., such as Ae. albopictus, has been difficult because standard vector surveillance traps are generally placed 1.5 m above the ground, are operated overnight, and utilize light as an attractant [37]. Since Ae. albopictus is diurnal and not attracted to light, host-seeks near the ground surface, and utilizes visual, in addition to olfactory cues for host location [18], [21], [38] these traps are not an effective way to collect this species. Consequently, most blood meal analyses to date were performed on specimens collected from areas where their densities are very high, such as tire yards and tire dumps (Table 1). The creation of newly developed vector surveillance traps, such as the Biogents Sentinel (BGS) trap, have only recently allowed the collection of large number of Ae. albopictus specimens from typical urban and suburban areas for ecological studies [39]. These traps simulate convection currents created by human body heat, utilize lures which mimic human odors, are operated during the day, placed at the ground level, and utilize contrasting black and white markings that provide additional visual cues that may be attractive to Ae. albopictus [37]–[41].
We investigated the host feeding patterns of Ae. albopictus in temperate North America, near the northernmost boundary of established populations in the eastern United States [4], [5]. We used an extensive network of BGS traps, which enhance Ae. albopictus capture counts, to conduct a multi-year collection of blooded mosquitoes (2008–2011) in urban and suburban sites as part of a larger area-wide project aimed at managing the Asian tiger mosquito [42], [43]. Additionally, we assayed blood meals from Culex mosquitoes collected in the same traps, locations, and dates as Ae. albopictus to determine the diversity of different blood meal sources obtained from the two vectors. We discuss the implications of our results on established and expanding populations of Ae. albopictus and the imminent outbreaks of exotic diseases such as chikungunya or dengue fevers in North America.
All studies were conducted within the jurisdictions of the authors' respective governance domain by professional mosquito control personnel. All entomological surveys and collections made on private lands or in private residences were conducted after acquisition of oral or written consent from residents. No specific permits were required for the mosquito collections. These studies did not involve endangered or protected species.
All collections were conducted within two counties (Mercer and Monmouth) located in central New Jersey, USA. Mercer County (40°13′N, 74°44′W) is highly urban, with 364,883 residents [44] and a population density of 630.2 inhabitants per square kilometer. Mercer County and the low-income City of Trenton, where the studies were conducted, have a population density of 4,286.5/km2 (USCB 2009a). The City of Trenton contains typical dense inner city housing, often built as adjoining row homes or duplexes [45]. Monmouth County (40°44′N, 74°17′W) is defined as primarily suburban and is located in east-central New Jersey with a population of 630,380 [46]. The boroughs on the Raritan Bayshore, within Monmouth County, where the studies were conducted, have an average population of 1,907.4/km2 [46]. The Raritan Bayshore primarily contains middle income coastal suburban homes which are often interspersed with forest and green space remnants [42]. Within each county, three predefined ∼1,000-parcel sites (a parcel is a combination of a house and its associated yard space), ranging in area from 1 km2 (Mercer) to 2 km2 (Monmouth) were chosen for our investigations. Although individual parcel sizes within the study sites in Mercer County were smaller (199.5±18.3 m2) than those in Monmouth County (571.1±31.2 m2), the number of residents within Mercer sites (19,494) were larger than within Monmouth sites (12,743). Every site, within each county, was previously selected to contain similar socioeconomic parameters, geography, human population density, and mosquito abundance. For a detailed description about site selection and the parameters of each individual site, please refer to [42], [43].
Mosquitoes were sampled on a weekly basis during 2008–2011 using a network of Biogents Sentinel (BGS) traps (Biogents AG, Regensburg, Germany). Specific details of surveillance protocols are outlined elsewhere [40]–[43], [47]; but briefly, trap locations were chosen by overlaying a grid of specific distance intervals. We used a 175–200 m distance between BGS traps for each site in Mercer County and 200–400 m distances in Monmouth County because of the larger site areas and limiting number of traps in inventory. These distances were based on current knowledge of Ae. albopictus flight range [21] and the available resources within each county. A total of 36 to 51 BGS traps, depending on the year, were deployed weekly in Mercer County, while 55 to 57 traps were deployed in Monmouth County. Each BGS trap was placed in residential backyards (near vegetation or shade) of each parcel selected, and was operated for 24 hours prior to collection. Each week, traps were placed in the same location within the backyards. The BGS trap was used with a solid BG-lure (Biogents AG, Regensburg, Germany) containing ammonia, lactic acid and fatty acids, components known to be attractive to Ae. albopictus [37]. Although the BGS trap was designed to capture host seeking (unfed) Aedes (Stegomyia) mosquitoes [39], the trap also captures other species such as Culex mosquitoes [37], [42] in addition to occasionally collecting female mosquitoes in varying gonotrophic stages (unengorged, blood fed, black blooded, and gravid). An unengorged or unfed mosquito does not contain visible evidence of blood in the abdomen, while a blood fed mosquito displays a distended abdomen with reddish blood clearly visible. A black blooded specimen has digested most of the blood meal and retains only a small portion of dark red or black blood visible near the ventral anterior of the abdomen, corresponding with Sella stage VI [48]. Gravid specimens have completely digested blood meals and contain visible eggs ready for oviposition.
Collections were placed on dry ice immediately and transported to the laboratory for identification and pooling. Species identification, enumeration, and gonotrophic stage determination was conducted under a dissecting microscope using a chill table to maintain a cold chain. Specimens were stored at −80°C for subsequent blood meal determination.
Abdomens of blooded Ae. albopictus were dissected over a chill table and then extracted using a Qiagen DNeasy Blood and Tissue Kit (Qiagen Sciences, Germantown, MD, USA). Specimens with very small blood remnants or those deemed poorly preserved (desiccated), were not utilized for DNA extraction because those samples rarely yield useful data [49]. To avoid contamination, forceps were flamed between extractions. To save time and reagents, we used a strategy that allows rapid identification of human-derived blood meals and mixes between human and non-human mammals [49]. This technique identifies human-derived blood meals based on the size of the PCR product on a gel without the need for extensive sequencing, thus drastically reducing costs. A mix between human and non-human blood is detected as two bands, and only the non-human band must be excised from the gel and purified with a QIAquick Gel Extraction Kit (Qiagen, Valencia, CA, USA) prior to sequencing [49]. Samples that did not amplify with the above assay were also tested with previously established primers designed for birds [50], reptiles/amphibians [51], and an additional primer set for mammals [52]. Approximately half of the specimens were tested with all bloodmeal identification methods above to legitimize the use of the rapid-assay [49]. To test for contamination, negative controls were employed in all reactions. The negative controls consisted of the PCR master mix with sterile water. Except for the short human-only band obtained with the Egizi et al. assay [49], and when the non-human band was excised from the agarose gel (see above), all PCR products were cleaned with Exo-Sap-IT (USB Products, Cleveland, OH, USA), cycle-sequenced with the forward primer of each pair, and run on capillary automated sequencers. Sequences were BLASTed in GenBank (http://www.ncbi.nlm.nih.gov/blast/Blast.cgi) to compare with sequences of known species. Only matches of >98% similarity were identified as the source of the blood meal [53].
A large number of blooded Culex mosquitoes, consisting primarily of Culex pipiens pipiens L. and Culex restuans Theobald, were also collected by the BGS traps. Because of the difficulty in accurate morphological identification of field-collected specimens due to age or damage [54]–[56] these specimens are often pooled as Culex spp. After using a molecular assay to identify all Culex mosquitoes to species [57], we tested blood fed Culex specimens from both counties collected in the same traps, locations, and dates as Ae. albopictus. Culex p. pipiens and Cx. restuans were the only Culex species collected in the BGS traps, and were assayed from Mercer County during 2009–2011 and from Monmouth County during 2008 and 2011. Blooded Culex specimens were extracted as described above for Ae. albopictus, amplified with the BM primer pair [58], then cleaned, sequenced, and identified as above. The BM primer pair targets a wide range of species, including mammals, birds, and reptiles, but it inadvertently amplifies in Ae. albopictus [49] and therefore cannot be used to identify blood meals in that species.
Spatial differences in the proportion of Ae. albopictus feeding on selected host species between the counties was compared by using Pearsons χ2 analysis for trend. All analyses were performed using IBM SPSS Statistics 21 (IBM, Armonk, NY, USA). Confidence intervals surrounding the estimated proportion of blood meals taken from a given species were calculated using the formula 95% CI = ±1.96×(square root p (1−p)/n), where p = the proportion of blood meals from a given source, and n = the total number of blood meals identified [59].
Our BGS trap surveillance during the active mosquito seasons of 2008–2011 collected 73,828 Ae. albopictus females in Mercer and Monmouth Counties (Table S1). A total of 33,392 Ae. albopictus were collected in Mercer County, 187 (0.56%) of which were visually determined to contain blood (blood fed or black blooded, hereafter “blooded”); while 40,436 Ae. albopictus were collected in Monmouth County, with 219 (0.54%) containing blood. In Mercer County, the number and proportion of blooded Ae. albopictus collected during each month was as follows: May (n = 1, 1.25% of monthly total), June (13, 0.82%), July (23, 0.42%), August (70, 0.57%), September (61, 0.57%), and October (19, 0.60%). Blooded Ae. albopictus in Monmouth County were collected during May (n = 4, 1.24% of monthly total), June (25, 1.11%), July (65, 0.99%), August (72, 0.45%), September (37, 0.33%), and October (16 (0.56%). We also captured 14,989 Culex mosquitoes (Cx. p. pipiens, Cx. restuans, and Cx. spp.) from both counties (Table S2). The BGS trap is highly specific for capturing host seeking Ae. albopictus females, as apparent by the nearly 74,000 specimens of this species that were captured versus the 15,000 specimens of Culex mosquitoes (Tables S1, S2). Interestingly, BGS traps were also capable of capturing blooded Ae. albopictus and Culex mosquitoes, as evidenced by the collection of over 406 blooded Ae. albopictus and 745 blooded Culex (Tables S1, S2).
Of the 406 blooded Ae. albopictus collected, 117 individuals were too desiccated and therefore only 289 specimens were suitable for dissection. Subsequently, the blood meal origin of 165 (57.10%) specimens was successfully determined (Table S1, 2). In Mercer County, 125 were tested for host blood meal origination with a successful identification from 86 (68.80%) specimens (Table 2). In Monmouth County, 164 Ae. albopictus were tested, with a successful host determination from 79 (48.17%) of those specimens (Table 2).
Aedes albopictus fed exclusively on mammalian hosts in Mercer and Monmouth Counties, with over 84% of all identified blood meals stemming from humans (52.12%), cats (20.61%), or dogs (11.52%) (Table 2). Blood meals were also detected from opossums (4.24%), gray squirrels (3.64%), cottontail rabbits (1.21%), and a white-footed mouse (0.61%). A small percentage (6.06%) of double blood meals (from two different host species) were detected in Ae. albopictus (4.65% of total in Mercer and 7.60% of total in Monmouth), and all included human blood (human+dog, n = 5; human+cat, n = 4; human+deer, n = 1). The number of Ae. albopictus feeding on humans was significantly higher in suburban Monmouth (62%) than in urban Mercer (43%) County locations (χ2 = 8.151; df = 1; p = 0.004), but significantly more Ae. albopictus fed on cats in Mercer than in Monmouth County (χ2 = 5.256; df = 1; p = 0.022). No significant difference was observed in the number of Ae. albopictus feeding on dogs between the two counties. No avian-derived blood meals were detected in any of the Ae. albopictus specimens tested.
Human- and cat-derived blood meals in Ae. albopictus were detected every month of our studies, while dog-derived blood meals were absent during May (Figure 1). Only 2.08% of all human-derived blood meals were detected in May, while the vast majority was detected during the month of August (38.54%). Four contiguous months (July, August, September, and October) accounted for over 87% of all blood meal collections (Figure 1).
We collected 745 blooded Culex (349 Cx. p. pipiens, 181 Cx. restuans, 215 Cx. spp.) mosquitoes during 2008–2011, and tested a subsample of 198 individuals identified as Cx. p. pipiens or Cx. restuans for blood meal source determination (Table 3). We selected 198 specimens to approximate the number of blood meals identified from Ae. albopictus and chose specimens from the same dates and traps as feasible. We were able to identify the blood meal source of 159 (80.30%) samples. Blooded Cx. p. pipiens were collected during April (n = 1, 0.79%), May (19, 15.08%), June (37, 29.37%), July (26 (20.63%), August (19, 15.08%), September (21, 16.67%), and October (3, 2.38%). Blooded Cx. restuans were collected during May (n = 10, 30.30%), June (12, 36.36%), July (6, 18.18%), August (2, 6.06%), September (2, 6.06%), and October (1, 3.03%). In Mercer County, specimens were tested from 2009–2011 and resulted in successful host determination from 61 Cx. p. pipiens (n = 74, 82.43%) and 7 Cx. restuans (n = 7, 100%). In Monmouth County, the blood meal hosts of 65 Cx. p. pipiens (n = 80, 81.25%) and 26 Cx. restuans (n = 37, 70.27%) were determined from 2008 and 2011 (Table 3).
Culex mosquitoes were predominately ornithophagic (n = 149, 93.71%) with only a small proportion feeding on mammalian hosts (n = 10, 6.29%) (Table 3). In Mercer County, the avian blood meal hosts of Cx. p. pipiens included 16 avian species (88.52%), while mammalian blood meals were obtained from only three species (11.48%). Mammalian blood was not detected in Cx. restuans from Mercer County, whereas avian blood meals were derived from four species (Table 3). In Monmouth County, avian hosts of Cx. p. pipiens included 12 species (95.39%), while mammalian blood meals were obtained from only two species (4.62%). No mammalian blood was detected in Cx. restuans from Monmouth County and avian-derived blood meals were obtained from ten species (Table 3).
Our investigations provide insight into the host associations of Ae. albopictus in the northernmost boundary of their established populations in eastern USA. Currently, about one-third of the human population of 55 million in this region reside in urban areas where Ae. albopictus is pervasive. This number is predicted to double under forthcoming climate change scenarios, encompassing all major urban centers and placing over 30 million people under the threat of dense Ae. albopictus infestations and potential public health threats from associated emerging mosquito-borne diseases [5]. Our analyses on the blood feeding behavior of Ae. albopictus demonstrate that this species is primarily mammalophagic in peridomestic environments of northeastern USA, and in some locations over 60% of their blood meals are derived from humans.
Host preference studies involving Ae. albopictus are often limited by the low sample numbers of blooded mosquitoes that are collected. This is because blooded Ae. albopictus have been difficult to collect [26], [32]. Previous sampling methods have often used combinations of aspirators, sweep nets, human baits, sticky traps, carbon dioxide-baited traps, and gravid traps in order to increase catch counts and as mentioned, often sampled exclusively in high density areas such as tire yards and dumps [17], [19], [20], [30], [31]. But trapping methods may bias results significantly [60], and Ae. albopictus is not readily attracted to traditional types of vector surveillance traps [26], [37]. A consistent sampling tool was not available for Ae. albopictus until the development of the BGS trap, which allowed us to sample populations of this species across a large geographic area over multiple years [42], [43]. Although we primarily utilized BGS traps for surveillance of host seeking Ae. albopictus, these traps also collected blooded specimens, which were subjected to molecular testing to characterize host feeding patterns of this species. However, unlike blooded or black blooded Culex mosquitoes which are easy to discern visually, blooded Ae. albopictus (unless fully engorged on fresh blood) are problematic to ascertain. This is because Ae. albopictus is a smaller species that imbibes smaller blood meals [18], [21] or on multiple hosts [61], [62], and contains a darker integument which hinders accurate detection of blood meals [32], particularly those in later Sella stages of development [26]. For example, parity studies conducted within our sampling sites on 166 Ae. albopictus visually determined as unengorged, detected blood meals or eggs in over 28% of those samples (Farajollahi et al. unpublished data). Our field investigations collected over 400 blooded Ae. albopictus during 2008–2011, 289 of which contained amplifiable blood for host determination analyses, with a successful amplification rate of close to 60%. In contrast, amplification rates were much higher for Culex mosquitoes (80%), likely because bird blood is nucleated and amplification of target DNA is easier for identification [53]. Interestingly, we collected twice as many blooded Culex mosquitoes than blooded Ae. albopictus, despite the demonstrable specificity of the BGS trap for the latter species. Amplification rates for Ae. albopictus also varied between the seasons and counties, as several abnormal weather patterns were experienced, threatening specimen handling and maintenance of the cold chain. The summers of 2010–2011 were particularly detrimental for blooded Ae. albopictus because the excessive heat (warmest and 3rd warmest summers on record) may have desiccated specimens much faster in the BGS traps and reduced amplifiable DNA through degradation (http://climate.rutgers.edu/stateclim_v1/data). Nonetheless, successful blood meal results from 165 Ae. albopictus across a consistent spatial/temporal span provides valuable insight into the host associations of this species in the northeastern USA.
Our investigations are consistent with previous studies that have shown a high mammalian affinity by invasive Ae. albopictus in temperate areas of USA and Europe [19], [20], [25], [27], [31], [32]. However, unlike most of these studies, we did not document avian-derived blood meals in any of our Ae. albopictus samples despite extensive testing with avian-specific primers. Our findings cannot be attributed to the method of collection, blood meal identification methodology, host availability, or spatial/temporal factors, since the Culex mosquitoes collected in the same traps at the same time, were found to feed predominately on birds within our study sites as expected [59], [63], [64]. The lack of blood meals obtained from birds by Ae. albopictus suggest that this species may have limited exposure to endemic avian arboviruses, such as West Nile virus (WNV), which is supported by the lack of WNV isolations in over 34,500 specimens assayed in a complementary study [65]. However, the high mammalian affinity of Ae. albopictus suggests that this species may be an efficient vector of mammal-driven zoonoses such as La Crosse virus, and human-driven anthroponoses such as DENV and CHIKV.
Another concern regarding the vectorial capacity of Ae. albopictus stems from detection of multiple blood meals from field populations. Previous studies have documented vertebrate blood from more than one host in Ae. albopictus throughout its endemic and invasive range (Table 1). Our studies detected double blood meals in 6% of the field-collected Ae. albopictus specimens, consistent with the 6% to 10% double blood meal proportion rates reported by others [22], [26], [27], [30], [31]. The capacity for Ae. albopictus to acquire multiple blood meals, particularly from human and other host species, increases the vector potential of this mosquito because of greater exposure to infected hosts during multiple feedings.
Large proportions of human-derived blood meals have been documented previously in Ae. albopictus and a few studies have reported that field populations feed exclusively on humans (Table 1), but the use of aspirators and human bait may bias these estimates. Additionally, recent investigations in temperate Italy have shown that Ae. albopictus feeding patterns differ between urban and rural habitats, with 90% of blood meals in urban areas from humans and only 20% being human-derived in rural habitats [31]. Our results report a significantly higher proportion of human blood meals in Ae. albopictus from suburban areas, rather than the densely populated urban areas. This was surprising, because of the higher (>2 times) human population density in urban Mercer County. However, suburban dwellers often spend more time outdoors gardening or undertaking leisure activities in backyards during daylight hours which will increase exposure. In addition, proportions of Ae. albopictus feeding on cats and dogs was higher in urban than suburban sites, likely reflecting large populations of feral cats in urban low income areas [66] and the fact that often dogs are kept in outside cages or yards for homeowner protection [40]. In contrast, suburban residents primarily keep their pets indoors and availability of these hosts for Ae. albopictus may be reduced. The significantly greater anthropophagic behavior of Ae. albopictus in more affluent suburban versus low-income urban habitats of northeastern USAindicates that a larger public health concern may exist within suburban landscapes, despite lower human population densities. Higher proportions of Ae. albopictus feeding on cats and dogs within urban environs may help fuel local mosquito populations but it may also afford zooprophylaxis protection for humans during epidemic outbreaks of anthroponoses such as DENV or CHIKV, because it will divert vector feeding to non-susceptible dead-end hosts.
Recent decades have witnessed a dramatic global expansion of Ae. albopictus into temperate areas and an increase in locally acquired autochthonous cases of tropical diseases such as DENV and CHIKV [9], [11], [67]. Because of the increasing abundance of Ae. albopictus and the escalating diagnoses of exotic pathogens in travelers returning from endemic or epidemic areas [14], the risk of a tropical disease outbreak in a new area is no longer speculative. We have shown that in urban and suburban areas of temperate northeastern USA, invasive populations of Ae. albopictus fed exclusively on mammalian hosts and that a large proportion (50–60%) fed on human hosts. Although we did not detect any avian-derived blood meals from Ae. albopictus during our investigations, the species has been traditionally classified as an opportunistic feeder whose host preference is greatly dependent on the abundance of available local hosts [18], [21]. Our studies indicate that Ae. albopictus may play a greater role in anthroponoses disease cycles, such as DENV and CHIKV, and a lesser role in zoonoses involving an avian animal reservoir. However, we cannot rule out the possibility that Ae. albopictus may occasionally act as a bridge vector for endemic pathogens such as St. Louis encephalitis virus and WNV by feeding on infected hosts when their abundance is great. Nonetheless, the large and growing populations of Ae. albopictus in major metropolitan urban and suburban centers, make a large autochthonous outbreak of an arbovirus such as CHIKV or DENV a clear and present danger. This may be particularly imminent in the case of CHIKV, as the virus is explosively spreading in the Caribbean region of the western hemisphere for the first time [68]. Given the difficulty in successful suppression of Ae. albopictus in areas where it has become firmly established [5], [43], we strongly recommend further ecological investigations on this species and caution public health practitioners and policy makers to install proactive measures for the imminent mitigation of an exotic pathogen outbreak.
|
10.1371/journal.ppat.1005561 | IL-12p40/IL-10 Producing preCD8α/Clec9A+ Dendritic Cells Are Induced in Neonates upon Listeria monocytogenes Infection | Infection by Listeria monocytogenes (Lm) causes serious sepsis and meningitis leading to mortality in neonates. This work explored the ability of CD11chigh lineage DCs to induce CD8+ T-cell immune protection against Lm in mice before 7 days of life, a period symbolized by the absence of murine IL-12p70-producing CD11chighCD8α+ dendritic cells (DCs). We characterized a dominant functional Batf3-dependent precursor of CD11chigh DCs that is Clec9A+CD205+CD24+ but CD8α- at 3 days of life. After Lm-OVA infection, these pre-DCs that cross-present Ag display the unique ability to produce high levels of IL-12p40 (not IL-12p70 nor IL-23), which enhances OVA-specific CD8+ T cell response, and regulatory IL-10 that limits OVA-specific CD8+ T cell response. Targeting these neonatal pre-DCs for the first time with a single treatment of anti-Clec9A-OVA antibody in combination with a DC activating agent such as poly(I:C) increased the protection against later exposure to the Lm-OVA strain. Poly(I:C) was shown to induce IL-12p40 production, but not IL-10 by neonatal pre-DCs. In conclusion, we identified a new biologically active precursor of Clec9A+ CD8α- DCs, endowed with regulatory properties in early life that represents a valuable target to augment memory responses to vaccines.
| Lm is a gram-positive food-borne pathogen that is the ethiological agent of listeriosis, a worldwide disease reported most frequently in developed countries. It can cause spontaneous septic abortions, fatal meningitis or encephalitis in immunocompromised and pregnant individuals. The murine model of systemic Lm infection has been demonstrated as a useful model to understand host resistance to intracellular pathogens. Neonates are highly susceptible to infections such as Lm, and display low responses to vaccines requiring IFN-γ producing T cells. In the present study, we characterized in murine neonates a precursor of conventional dendritic cells that is able to produce IL-12p40 and IL-10 cytokines and to modulate the development of the adaptive immune response, more particularly the CD8+ T cell response upon exposure to Lm. By targeting Lm-associated antigens to these conventional dendritic cell precursors in neonates, we succeeded to confer a partial protection to a lethal dose of Lm at the adult stage. Our study provides new insights into our understanding of the innate immune response to infections in early life and will help to design new vaccine strategies in newborns.
| Early life is a period of immune maturation characterized by a high susceptibility to infectious diseases. The underdeveloped immune system gives a Th2-biased response and has an impaired ability to develop long-lasting protective CD8+ T cell immunity [1, 2]. We are particularly interested in immune resistance to infections by Listeria monocytogenes (Lm). Lm is a gram-positive opportunistic food-borne bacteria with a facultative intracellular life cycle that commonly causes sepsis and/or meningitis, leading to mortality in neonates but is asymptomatic in immunocompetent Lm-infected adults [3].
DCs are the key components of the immune system, determining susceptibility to infections. The primary function of DCs is the detection of pathogens and the initiation of the adaptive immune response. Such a response requires the DCs to present an antigen (Ag) from a specific pathogen, as well as an innate signal from microbes or damaged cells allowing DCs to orchestrate the adaptive immune response. Conventional DCs (cDCs) in mice can be divided into two distinct populations, one with high expression of CD8α (CD8α+ cDCs) and the other with no expression of CD8α (CD8α- cDCs). These CD8α+ cDCs selectively express the C-type lectin receptor DNGR1, also called Clec9A [4]. The development of CD8α + cDCs depends on a common set of transcription factors including Irf8 [5], Batf3 [6], Id2 [7] and Nfil3 [8]. CD8α+ cDCs are particularly efficient at internalizing and cross-presenting exogenous Ag on MHC class I molecules, especially from dead or dying cells [9–14]. cDCs represent a key subset which initiates cell-mediated immunity against tumors, viruses and bacteria [15, 16]. Upon Lm infection, adult CD8α + DCs phagocytize the bacteria in the marginal zone of the spleen, and migrate to the T-cell zone in order to present the bacterial antigens to CD8+ T cells [17]. The resultant response involves the up-regulation of co-stimulatory molecules, the production of cytokines like IFN-γ and the generation of cytotoxic T-cell immunity. Finally, CD8α+ cDCs have been identified as professional IL-12p70 producers priming the adaptive immune cells towards Th1 differentiation [18–21].
In murine neonates, CD8α+ cDCs have been shown to be defective in the first 6 days of life. Beyond this time, the CD8α+ cDCs producing IL-12p70 induces the downregulation of the IL-4Rα/IL-13Rα1 on T cells, favoring a Th1 response [2]. Since the study by Lee H. et al. [2], the immune neonatal period has been redefined. As a result, some of the previous reports on the quantitative and qualitative shortcoming of neonatal DCs have to be revisited. For example, it was demonstrated that at 7 days of life the Flt3 ligand-treated “neonatal” mice showed an increase in DCs lineage development and an increased in IL-12-dependent innate resistance against Lm [22]. Another study reported that one-day-old DCs were able to produce adult level of IL-12p70, but only after IL-4, a maturating cytokine, was added to GM-CSF and CpG in the culture [23].
Neonatal induction of Th1/Tc1 memory is still controversial. Neonates have shown to be more susceptible to intracellular pathogens due to a suboptimal capacity to mount an efficient cell-mediated immunity, particularly the memory CD8+ T cells. For instance, qualitative defect in neonatal Batf3-dependent CD103+ lung DCs were recently reported to influence the CD8+ T cell response, following respiratory syncytial virus (RSV) infection [24]. However, other studies have demonstrated that neonates could mount an adult-like CD8+ T cell immune response against human CMV or Trypanosoma cruzi [25, 26]. Concerning Lm infection in early life, a previous study demonstrated that 5- to 7-days old neonates are able to develop robust primary and secondary CD4+ and CD8+ Th1-type responses against Lm without characterizing the antigen presenting cells that were involved [27].
The objectives of this study were to describe the phenotype of Batf3-dependent CD11chigh DC subset and to explore their abilities to induce a CD8+ T cell immune protection against Lm at 3 days of life. First, we characterized the splenic DC subset bearing DNGR1/Clec9A but not CD8α, a precursor of CD8α + DCs. This DNGR1/Clec9A bearing DC is the predominant lineage before 6 days of life. Next, we demonstrated the role of these early DCs in taking up and presenting exogenous Lm Ag to prime a CD8+ T-cell response. Additionally, we defined the role of IL-12p40 and IL-10 uniquely produced by these neonatal pre-DCs in the establishment of an adaptive response. Finally, we assessed vaccination strategies, directly targeting neonatal DCs using OVA coupled to anti-Clec9A in the presence of poly(I:C). This study clarifies the function of pre-CD8α + DCs in early life and highlights the advantages for human neonatal vaccination strategies.
To determine the type of DC involved in the adaptive immune response against Lm at 3 days of life, we employed Batf3-/- mice known to lack the conventional CD8α+ type of DC [6]. We compared the OVA-specific primary immune response to the attenuated strain Lm actA-/--OVA in 3-day-old and adult C57BL/6 and Batf3-/- mice. As seen in Fig 1A, the production of IFN-γ, following restimulation with MHC I-restricted OVA peptides, was drastically reduced in splenic cultures from Batf3-/- mice compared with C57BL/6 mice at both ages. This suggested that Batf3-dependent DCs are required to trigger an IFN-γ T cell response against Lm infection at 3 days of life as well as at the adult stage.
To characterize the DCs likely to generate an adaptive CD8+ T-cell response against Lm in neonates, we focused on the Batf3-dependent CD8α+ DC lineage. In naïve adult C57BL/6 spleen, the majority (69%) of CD11b-CD11chigh cells were CD8α+ (Fig 1B). In contrast, in neonates, the majority (85%) of splenic CD11b-CD11chigh cells lacked CD8α expression but were positive for CD205 and DNGR1/Clec9A, which are features of CD8α+ DC family [4], as confirmed in adult splenic CD8α+ DCs (Fig 1B). We further observed that CD11chigh CD11b-CD205+DNGR1+ DCs in mesenteric lymph nodes collected from 5 day-old C57BL/6 mice (S1 Fig) expressed the same intermediate level of CD8α as their splenic counterpart. The neonatal splenic CD11chigh CD11blow CD205+CD8α- DCs also expressed CD24 and cKIT at similar levels but expressed MHCII, CD80 and CD86 molecules at a lower level, compared to adult CD8α + DCs (S2A Fig). Interestingly, neonatal CD11chighCD11blowCD205+CD8α + or CD8α - displayed similar levels of aforementioned molecules suggesting that they were closely related phenotypically (S2A Fig). No expression of CD207, CD4 or B220 was detected in neonatal CD8α-CD11b-CD205+ DCs nor in adult CD8α+CD11b-CD205+DCs (S2A Fig) and the high expression of DNGR1/Clec9A was restricted to neonatal CD8α -CD11b-CD205+ DCs and not to neonatal pDCs or cDC2 (S2B Fig). These results demonstrated that neonatal CD8α -CD11b- DCs display a surface phenotype closely similar to adult CD8α + DCs except for the expression of CD8α, what suggests that they are earlier forms of the same DC lineage. This was supported by analysis of Batf3-/- neonates in which no CD11c+CD11b-CD205+DNGR1+ CD8α - DCs could be detected (Fig 1C and S3A Fig).
Next we examined the developmental relationship between neonatal CD11b-CD8α - DCs and adult CD8α + DCs. A time course of the frequency of CD11c+CD11b-CD205+CD8α - DCs versus CD11c+CD11b-CD205+CD8α + DCs showed that the latter were dominant for the first 5 days of life. After day 7, the proportions were totally reversed as CD11c+CD11b-CD205+ DCs expressing CD8α became dominant (Fig 1D). We thus confirmed the appearance and accumulation of the CD8α + DCs at day 6 as previously reported [2]. Finally, to test if neonatal spleen contained precursors of the CD8α + DC type found in adults, we transferred, into adult C57BL/6 CD45.1+ mice, neonatal C57BL/6 CD45.2+ spleen cells that were depleted of CD3+, CD19+ and Gr1+ cells and that contained mostly CD11c+CD11b-CD24+ CD205+ CD8α - cells. The number of DCs of donor origin expressing CD8α among CD45.2+CD11c+ CD11b- CD24+ CD205+ spleen cells showed a steady increase to adult levels by day-6 post transfer (Fig 1E). Taken together, these results suggested that earlier forms of the CD8α + DCs, that we will call preCD8α Clec9A+ DCs, are predominant among splenic and mesenteric lymph nodes CD11chigh DCs from birth to 5 days of life.
We further determined that C57BL/6 neonates that were submitted to a 3 day treatment with the dendritic cells growth factor, Flt3L [28] (instead of 7 days as previously reported [22]), expanded preferentially the absolute number of preCD8α Clec9A+ DCs (Fig 1F) without affecting the number of neonatal plasmacytoid DCs or conventional DC2 (S3C Fig) and significantly enhanced their defense against Lm (Fig 1G) as previously described [22].
As presented in Fig 2A, we first demonstrated that neonatal preCD8α Clec9A+ DCs were able to phagocytize Lm-GFP (62%±2.6) as did adult CD8α + DCs (84.3%±3). To assess capacity to cross-present cell-associated Ag during Lm infection, we purified splenic C57BL/6 neonatal preCD8α Clec9A+ DCs or adult CD8α + DCs, incubated them with OVA257-264 peptides or Lm-OVA and co-cultured them with OVA-specific OT-I T cells, with or without GM-CSF known to enhance cross-presentation capacity of newly formed DC [29]; and monitored the IFN-γ production (Fig 2B). Neonatal preCD8α Clec9A+ DCs were able to cross-present OVA as efficiently as adult CD8α + DCs in the presence of GM-CSF. We evaluated at the RNA level the cDNA relative expression of a large number of genes involved in the cross-presentation machinery (such as Rac2, Ergic 1, 2 and 3, Rab14, Erap1, Sec22b, Syntaxin 4, TAP1 and 2) and concluded that they were similar in sorted preCD8α DCs and adult CD8α DCs except for a slight difference with H2-K1, β2m and Rab27a (S4 Fig). This cross-presentation ability was confirmed ex vivo with neonates that were injected with Lm actA-/--OVA, from which preCD8α Clec9A+ DCs were sorted and co-cultured with OT-I T cells, with or without GM-CSF (Fig 2C).
We then assessed the capacity of neonatal preCD8α Clec9A+ to produce cytokines in several settings such as Lm exposure or TLR ligand encounter and compared it to fully competent adult CD8α + DCs. As an endosomial TLR receptor, TLR3 binds to ligand such as dsRNA, namely polyinosinic-polycytidylic acid (Poly(I:C)) which mimic the replication intermediate of virus. TLR3 agonist was shown to promote the cross-presentation of antigens that transit in the endosome as it increases MHC class I and costimulatory molecules of DCs and stimulate IL-12 secretion. It is considered as an efficient manner to optimally activate DCs to promote CD8+ T cell activation. Moreover, TLR3 is only expressed in CD8α + DC subset. Poly(I:C) was first injected into C57BL/6 adults and neonates and intracellular staining for IL-12p40 was performed on CD11c+CD11b-CD205+CD8α + and CD11c+CD11b-CD205+CD8α - at different time points (Fig 3A and S5 Fig). As expected, the overall number and relative frequency of IL-12p40 producing CD8α + DCs was higher in adults than in neonates. Interestingly, 2h after poly(I:C) treatment, a high proportion of CD8α - DCs were producing IL-12p40 in neonates but not in adults (Fig 3A).
To assess cytokine responses by the DC to Lm infection, cytokine secretion of cultured cells and mRNA synthesis was utilized. For mRNA measurement, we sorted DC subsets from neonates infected with 5 x 105 Lm actA-/- or adults infected with 5 x 105 or 5 x 106 CFU Lm actA-/-. IL-12p35 mRNA synthesis was induced in adult CD8α + DCs but not in CD8α - DCs at both ages (Fig 3B). A significant induction of IL-12p40 mRNA synthesis was observed in neonatal and adult CD8α- DCs and in adult CD8α + DCs upon Lm actA-/- infection (Fig 3B). In addition, sorted neonatal preCD8α Clec9A+ DCs secreted quite similar protein levels of IL-12p40 as adult CD8α + DCs after in vitro stimulation with poly(I:C) or Lm (Fig 3C). However, neither IL-12p70 (Fig 3D) nor IL-23 (S6 Fig) were significantly produced by these neonatal preCD8α Clec9A+ DCs after such stimulation, even in the presence of GM-CSF, IL-4 and IFN-γ; by contrast, CpG did induce IL-12p70 secretion in such maturating conditions (Fig 3D) as previously described [23]. IL-23p19 mRNA synthesis was not induced in all the tested DC subsets in response to Lm infection (Fig 3B). Of particular interest, IL-10 transcripts were strongly induced only in neonatal preCD8α Clec9A+ DCs after Lm actA-/- infection whereas adult preCD8α Clec9A+ DC and adult CD8α + DCs did not (Fig 3E). The exclusive production of IL-10 by neonatal preCD8α Clec9A+ DCs after Lm actA-/-stimulation was confirmed in vitro at the protein level (Fig 3F). Finally, the TLR expression was compared between preCD8α DCs and CD8α + DCs to potentially explain their distinct functionality. As presented in S7 Fig, TLR9 RNA expressions were similar and TLR3 RNA was expressed only 1.14 times more in adult CD8α + DCs. Altogether, these results indicate that, upon Lm infection, splenic neonatal preCD8α Clec9A+ DCs have the unique ability to produce IL-12p40 and IL-10 but no IL-12p70 and IL-23.
We investigated the impact of the IL-12p40 subunit and of IL-10 in the protective immune response against Lm infection. Neutralizing anti-IL-12p40 mAb was administered to neonates exposed to Lm. The survival rate of neonates was significantly reduced compared to control isotype-treated infected mice (Fig 4A). The impact of IL-12p40 on the primary CD8+ T-cell response during Lm infection was then assessed. IL-12p40 neutralization during Lm actA-/--OVA infection strongly inhibited the frequency of IFN-γ-producing CD8+ T cells (Fig 4B) as well as the secretion of IFN-γ (Fig 4C) in response to OVA257-264 peptides. Furthermore, when anti-IL-12p40 mAb was added during co-culture of sorted neonatal preCD8α Clec9A+DCs with Lm-OVA (MOI 1) in the presence of GM-CSF and OT-I T cells, an inhibition of IFN-γ production was obtained (Fig 4D). We further demonstrated that the in vivo proliferative response of transferred OT-I T cells in C57BL/6 neonates induced by Lm actA-/--OVA infection was significantly inhibited in IL-12p40 -/- neonates but not in IL-12p35-/- or IL-23p19 -/- neonates (Fig 4E). Finally, the role played by IL-10 in CD8+ T-cell activation induced by Lm was assessed by blocking the IL-10 receptor. The IFN-γ production of OT-I T cells in response to Lm-OVA presentation in vitro by sorted neonatal preCD8α Clec9A+ DCs was enhanced in the presence of anti-IL-10R mAb (Fig 4F). These results demonstrated that the IL-12p40 subunit produced by neonatal preCD8α Clec9A+ DCs is functional in inducing an efficient primary CD8+ T cell response whereas the IL-10 secreted by these precursor DCs moderates the CD8+ T cell activation.
We then investigated whether targeting antigens to Clec9A on the neonatal preCD8α Clec9A+ DCs could be an effective strategy for immunization against Lm infection. We first tested whether the neonatal preCD8α Clec9A+ DCs are able to cross-present to CD8+ T cells in vivo through Clec9A targeting. C57BL/6 neonates or Batf3-/- neonates were i.v. injected with either the construct of OVA protein linked to anti-Clec9A antibody (anti-Clec9A-OVA) or with Lm actA-/--OVA, along with CFSE-labelled OT-I T cells that were monitored for proliferation. Constructs of OVA linked to an isotype-matched control Ab (GL117-OVA) was used as a control. As seen in Fig 5, anti-Clec9A-OVA administration induced OT-I T cell proliferation in vivo, similar to that obtained with Lm actA-/--OVA (Fig 5A). Furthermore, the proliferation induced by the Clec9A targeting was exclusively Batf3-dependent, since no proliferation was observed in the Baft3-/- mice compared to GL117-OVA treated mice (Fig 5B) whereas the Lm actA-/--OVA-induced OT-I cell proliferation was only partially dependent on Batf3 (Fig 5B). We therefore conclude that neonatal preCD8α Clec9A+ DCs serve as effective targets for cross-presentation through Clec9A and for CD8+ T cell activation, such as described for MHC-I restricted CD8+ T-cell responses in adult mice [30–32].
In a next step we proceeded to determine whether targeting antigen to preCD8α Clec9A+ DCs in neonates by anti-Clec9A-OVA constructs leads to enhanced protection against later Lm infection, and whether this requires DC activating agents. Neonates were vaccinated with anti-Clec9A-OVA or GL117-OVA constructs, with or without poly(I:C) as adjuvant. The mice were challenged by infection with Lm-OVA 60 days later. A group of unimmunized control mice was infected at day 60 to provide a primary response comparison (Fig 6A). Mice vaccinated at the neonatal stage with anti-Clec9A-OVA and poly(I:C) were partially protected, 55% surviving compared to 90–100% death in control groups (Fig 6B). Immunization with OVA linked to a non-targeting isotype control antibody, or with the targeted construct alone without adjuvant, did not allow survival. The survival of mice that were vaccinated at 3 days of life and challenged 60 days later with Lm-OVA was correlated with the decrease of bacterial burden. Neonatal immunization with anti-Clec9A-OVA and poly(I:C) reduced the amount of Lm in the adult spleen compared to GL117-OVA and GL117-OVA + poly(I:C) (Fig 6C). However, it was surprising to observe a significant early reduction of bacterial load in mice immunized with anti-Clec9A-OVA alone.
We next assessed the efficiency of the secondary T-cell response after vaccination. Five days after infection with Lm-OVA, the frequency of splenic IFN-γ producing CD8+ T cells in Clec9A-OVA and poly(I:C) immunized group after infection was strikingly higher than in all the other groups (Fig 6D). The use of poly(I:C) during immunization did increase the production of IFN-γ by the T cells after infection. These results with the OVA model antigen system raise the possibility of selectively delivering Lm antigens to neonatal preCD8α Clec9A+ DCs to trigger a protective immunity against Lm infections.
Quantitative and qualitative deficits in the neonatal innate immune response were proposed as causal factors to account for their inability to mount a protective IFN-γ-dependent CD8+ T-cell response against various viral pathogens such as respiratory syncytial virus, influenza virus, hepatitis B virus, herpes simplex virus as well as intracellular bacterial pathogens such as Lm. Indeed, enhanced neonatal protection against Lm infection can be restored through administering recombinant IFN-γ [33], through Flt3L treatment [22] or through administering CpG oligonucleotides [34], both last treatments inducing an IL-12-dependent innate resistance. However, the cellular elements of the neonatal innate immune system able to mount protective type 1 T-cell activation against Lm were not clearly identified, as the IL-12p70-producing CD8α + DC subset was reported to be defective in murine spleen before 7 days of age [2]. It has been well established that CD8α + DCs play a critical role in mounting an effective cytotoxic CD8+ T cell response to Lm infection in adult mice [15, 35]. CD8α+ DCs are crucial for both efficient bacterial entry into the spleen and induction of the immune response [17, 36].
In this work, we first demonstrated that a murine splenic neonatal CD8α-type DC precursor subset is predominant in the Batf3-dependent DC lineage before 6 days of life. These preCD8α Clec9A+ DCs express CD11chigh, DNGR1/Clec9A, CD24, CD205, and MHCII, without expression of CD8α, B220, CD4 or CD207. They are Batf3-dependent and expand preferentially after a limited Flt3L-treatment (3 days instead of 7 days), compared to pDC or cDC2, which are respectively the transcription factor and the growth factor known to be involved in CD8α+ DC subset development [6, 28]. They share the lineage features of CD8α+ DCs [37–39] and are most likely converted in vivo into CD8α + DC. These splenic DC precursors are still present in adults but constitute a minor population of the Batf3 lineage as reported by Bedoui et al. who described a similar immediate splenic CD11chigh,CD24+, MHCII+, CD205+, CD8α- precursor subset capable of cross presenting Ag but poorly addressed their ability to produce cytokines leading to Th1/Tc1 response [40]. We further characterized these adult splenic DC precursors by demonstrating that they were high producers of IL-12p40 exclusively. We also demonstrated that in addition to phagocytosing Lm, neonatal preCD8α Clec9A+ DCs possess the ability to efficiently cross-present Ag, both in vitro and in vivo, an ability well described to be restricted to CD8α+ DCs [6, 11, 12, 14]. Furthermore, GM-CSF enhances and allows full expression of their capacity to cross-present Ag, similar to the newly formed CD8α lineage DCs in Flt3 ligand stimulated bone-marrow cultures that require a maturation step promoted by GM-CSF to acquire the capacity to cross-present Ag [29]. We therefore conclude that the only CD11chigh DCs subset potentially able to cross-present Ag in neonatal spleen are the preCD8α Clec9A+ DCs, in contrast to adult spleen in which both the few CD24+CD8α- DCs and the predominant CD24+CD8α+ DCs have this capacity [40].
We clearly demonstrated that neonatal preCD8α Clec9A+ DCs are able to produce optimal levels of IL12p40 in response to poly(I:C) and Lm without IL-12p70 and IL-23 secretion. This is in accordance with the study which previously described the IL-12p35 gene expression defect of newborn monocyte-derived DC [41]. However, the secretion of IL-12p70 during early life has also been shown to be environment dependent. Indeed, one-day-old purified CD11c+ DCs were reported to be able to produce IL-12p70 in response to CpG if a cocktail of maturation agents like GM-CSF and IL-4 were added [23]. Here, we determined that the preCD8α Clec9A+ DCs are the subset able to produce IL-12p70 in response to CpG in combination with GM-CSF, IL-4 and IFN-γ. The fact that neonatal preCD8α Clec9A+ DCs did not produce IL-12p70 upon Lm and poly(I:C) stimulation, even in the presence of GM-CSF, IL-4 and IFN-γ, indicates that they are refractory to maturation through these pathways.
Interestingly, we have shown that IL-12p40 secreted by neonatal preCD8α Clec9A+ DCs plays a role in neonatal T cell immunity. The inhibition of IL-12p40 during Lm infection increased the neonatal mortality and reduced significantly the Ag-specific CD8+ T-cell expansion and activation. Furthermore, the ability of neonatal preCD8α Clec9A+ DCs to cross-present OVA Ag in vitro or ex vivo after Lm-actA-/-OVA incubation was significantly inhibited when IL-12p40 was neutralized but not IL-12p35 or IL-23p19. Some studies support an independent role for IL-12p40 and more precisely for the IL-12p(40)2 homodimeric form. It has been shown that IL-12p40 could act negatively by competitively binding to the IL-12 receptor in an IL-12 mediated shock [42] or by inhibiting IL-23 functions [43]. In contrast, it has also been suggested that IL-12p40 could have a positive role in inducing immune responses [44]. It has been shown that IL-12p40 promotes macrophage inflammation, DC migration and has a protective function in Mycobacterial infection [45, 46]. In line with our findings, it was also demonstrated that IL-12p(40)2 was involved in activation of naive T cells and in the induction of IFN-γ production by CD8+ T cells [47, 48]. Therefore, IL-12p70 but also IL-12p40 may act as a feed-back loop on costimulatory molecules and MHC molecule expression on dendritic cells to increase naive T cell activation and IFN-γ production [49]. We may therefore conclude that the CD8+ T cell immunity induced in early life against Lm is IL-12p40 dependent and IL-12p70 or IL-23 independent. These data are in contrast to previous studies showing the requirement of IL-23 in the protection against Lm in adults although this role was mostly associated to the activation of IL-17A/IL-17F producing γδ T cells [50] without identifying the cellular source of IL-23. Concerning the role of IL-23 in the CD8+ T cell response to Lm, it seems to be minor in adult mice [51].
A surprising finding was that, in addition to secreting IL-12p40, neonatal preCD8α Clec9A+ DCs produce IL-10. It was already known that neonatal mice display an increased production of IL-10 early in the course of infection with Lm and after CpG stimulation, but the source of IL-10 in these neonatal studies had been shown to be macrophages and CD5+ B cells [52, 53]. This is the first demonstration that Batf3-dependent DC precursors produce IL-10, suggesting a new mechanism responsible for suboptimal activation of neonatal CD8+ T cells. Indeed, blocking IL-10R during cross-presentation after Lm stimulation enhanced the production of IFN-γ by Ag-specific CD8+ T cells. Numerous publications indicate a regulatory role of IL-10 in DC activation and in the Th1/Th2 polarization in both adults and neonates [52–54]. The direct impact of IL-10 secreted by neonatal preCD8α Clec9A+ DCs in T-cell polarization will be analyzed in future investigations.
We demonstrated that these neonatal preCD8α Clec9A+ DCs could be used as a cellular target for direct delivery of Lm Ags in order to induce efficient immunization when poly(I:C) was co-administrated, allowing a later effective secondary immune response against Lm infection. This is in line with previous studies showing that delivering Ags into the cytoplasm of APCs, with for instance synthetic microspheres, was the key for a better induction of neonatal CD8+ T cell response [55–57].
Previous studies have shown that DNGR1/Clec9A excels as a target for enhancing CD8+ T-cell response and to generating follicular helper T cells in the presence of poly(I:C), in part due to its restricted expression, predominantly in the CD8α+ DC lineage and at a lower level in PDC [30, 58, 59]. We observed that injection of anti-Clec9A-OVA construct with poly(I:C) in 3-day-old neonates enhances the frequency of OVA-specific IFN-γ producing CD8+ T cells. Furthermore, the absence of CD8+ T-cell proliferation in Batf3-/- mice following anti-Clec9A-OVA construct injection confirms a specific involvement of neonatal CD8α- DCs in this process, and argues against a potential role for pDCs. Specifically, we showed here for the first time that a single treatment with anti-Clec9A-OVA construct and poly(I:C) at 3 days of life is enough to significantly enhance the protection of mice against later exposure to the Lm-OVA strain. This protective secondary response was associated with a control of the bacterial burden and a memory CD8+ T cell response involving Ag-specific IFN-γ producing CD8+ T cells. The poly(I:C) treatment was shown to induce in vitro the IL-12p40 but not the IL-10 secretion by isolated preCD8α Clec9A+ DCs, inhibiting their regulatory properties.
In summary, we have characterized a preCD8α Clec9A+ DC subset that is predominant in mouse spleen during the neonatal period. Compared with their adult counterpart or to the adult CD8α+ DCs, this neonatal Batf3-dependent DC precursors, that cross-present Ag, display the unique abilities to be high producers of IL-12p40 but also of IL-10. Upon infection with Lm, we demonstrated that these preCD8α Clec9A+ DCs are endowed with regulatory properties that control the CD8+ T cell response through IL-10. The capacity of these neonatal preCD8α Clec9A+DCs to induce a protective type 1 T-cell immune response against intracellular pathogens was allowed with anti-Clec9A construct and poly(I:C) treatment through a mechanism involving only the IL-12p40 subunit production with no IL-10.
This discovery opens new strategies for future human vaccine development. It requires the investigation of the ontogeny of the human equivalent of these neonatal preCD8α Clec9A+ DCs.
C57BL/6 CD45.2, C57BL/6 CD45.1, C57BL/6 IL-12p40-/-, C57BL/6 IL-12p35-/-, Batf3-/- [6] and OT-I TCR transgenic (OT-I) mice were purchased from Jackson Laboratory (Bar Harbor, USA). C57BL/ 6 IL-23p19-/- mice, with EGFP reporter gene, were kindly provided by E. Muraille (Université Libre de Bruxelles, Belgium). Mice were bred and housed in our specific pathogen-free animal facility. For all experiments, neonatal mice are defined as 3-day-old and adults as sex-matched 8-to 12-week-old mice. They were kept in sterile confinement in a P2 animal unit during infections. All animal studies were approved by the institutional Animal Care and Local Use committee.
The animal handling and procedures of this study were in accordance with the current European legislation (directive 86/609/EEC) and in agreement with the corresponding Belgian law “Arrêté royal relatif à la protection des animaux d'expérience du 6 avril 2010 publié le 14 mai 2010”. The complete protocol was reviewed and approved by the Animal Welfare Committee of the Institute of Biology and molecular medicine (IBMM) from the Université Libre de Bruxelles (ULB, Belgium) (Permit Number: 2014–43).
Lm-EGD strain (Lm), Lm-EGD strain deficient for actA ((Lm actA-/-) and Lm-GFP strain were kindly provided by Prof. P. Cossart (Pasteur Institute, Paris, France). Lm-OVA and Lm actA-/--OVA were purchased from DMX incorporated (Philadelphia, PA) [60]. Bacteria were cultured in BBL Brain Heart Infusion (BHI) medium (BD Diagnostics, USA) and stored at -80°C in 10% DMSO.
For survival studies, mice were injected i.p. for 7-day-old mice and i.v. for 3-day-old neonates or adults with different doses of Lm diluted in sterile-PBS. To determine the median lethal dose (LD50) of neonates, 3 or 7-day-old and comparative adult C57BL/6 mice were injected with 4 doses (50, 100, 1000 and 10000 CFU) of Lm WT and survival rates were observed for 15 days (S8 Fig). All 3-day-old mice died 2 days following 10000 CFU and 5 days following 1000 CFU of Lm inoculation; only 20% survived after 100 CFU infections and 60% after 50 CFU. In contrast, adults survived every Lm dose administered. The 7 day-old mice showed intermediate sensitivity, dying 8 days after the highest dose of Lm, 50% surviving 1000 CFU and all surviving 100 CFU and 50 CFU.
To quantify Lm burden, spleens were harvested 3 days after Lm infection and homogenized in LPS-free PBS using gentleMACS Dissociator (Miltenyi Biotec, Leiden The Netherlands). Serial dilutions of homogenates were plated on BHI-Agar for 24h at 37°C and bacterial CFUs were assessed.
For primary responses and/or vaccination, neonates mice were i.p. injected with 5 x 105 CFU of Lm actA-/- or Lm actA-/--OVA and adult mice were injected with 5 x 105 or 5 x 106 CFU Lm actA-/-. Neonates were vaccinated i.v. with 0,1μg of anti-Clec9A/OVA mAb or GL117/OVA control mAb. For secondary responses, 60 days after the first immunization, mice were i.v. challenged with 5 x 105 CFU of Lm-OVA.
When indicated, neonates were injected with neutralizing purified NA/LE Rat anti-Mouse IL-12p40/p70 (clone C17.8; 25μg/neonate 6 hours before, and 1 and 3 days after Lm injection) (BD Biosciences) or isotype-matched Ab (BD Biosciences) and with poly(I:C) (1 mg/kg; Sigma-Aldrich). Mice were eventually s.c. injected with 50μl of saline buffer (NaCl 0,9%) or Flt3L (20μg/ml) at day 0, 1 and 2 of life (Celldex, Phillipsburg, New Jersey).
Seven days after primary response or 5 days after Lm-OVA challenge, spleen cells were harvested and cultured with OVA257-264 peptide (1μg/ml, Polypetides Laboratories, Strasbourg, France) in complete culture RPMI 1640 medium (Lonza Research Products, Switzerland) as described (58). When indicated, IL-2 was added to the culture (10ng/ml, R&D Systems, Minneapolis, USA). Production of IFN-γ by CD8+ T-cells was measured by cytometry and ELISA.
OT-I cells were isolated from lymph nodes of OT-I TCR transgenic mice using the Dynabead untouched mouse CD8 cell protocols (Invitrogen, Life Technologies Europe B.V, Ghent, Belgium). CFSE-labeling (CellTrace CFSE Cell Proliferation Kit, Invitrogen, Life Technologies Europe B.V, Ghent, Belgium) or VPD450-labeling (Violet Proliferation Dye 450; BD Biosciences) were done following the manufacturer’s protocol.
C57BL/6 and Batf3-/- neonates were injected i.v. with 3x105 unlabeled or CFSE-/VPD450-labeled OT-I cells and with anti-Clec9A/OVA or GL117/OVA construct (0,1μg/mouse) or Lm actA-/--OVA (5x105 CFU/mouse). Spleen cells were harvested 60h later. Proliferation of OT-I T cells was assessed on a Cyan ADP cytometer (Dako Cytomation, Everlee, Belgium) by the dilution of CFSE staining.
All the following fluorochrome-conjugated mAbs (B220 (RA3-6B2), CD3ε (500A2), CD4 (GK1.5), CD8 (53–6.7), CD11b (M1/70), CD11c (HL3), CD19 (1D3), CD24 (M1/69), CD44 (IM7), CD62L (MEL-14), CD80 (16-10A1), CD86 (GL-1), CD117 (2B8), MHCII (M5/144.15.2) and Sirpα (P84) were purchased from BD Biosciences, fluorochrome-conjugated anti-CD205 mAb (NLDC-45) was purchased from BioLegend and anti-PDCA1 (eBio129c) and CD207 mAb (eBioL31) from ebiosciences (San Diego, CA, USA). Anti-DNGR1 mAb was a kind gift from Dr. Caetano Reis e Sousa (Immunobiology Laboratory, Cancer Research UK’s London Research Institute). For DC characterization, neonatal spleens and lymph nodes from C57BL/6 or Batf3-/- mice were harvested and disrupted using a Pyrex Potter tissue homogenizer (VWR). Red blood cells were lysed by Ammonium-Chloride-Potassium (ACK) Lysing Buffer. Cells (2-5x106) were stained in FACS buffer (PBS/0,5% BSA/2mM EDTA) at 4°C in the dark for 20 min. After fixation in 1% paraformaldehyde (Sigma-Aldrich BVBA, Diegem, Belgium), analysis was performed on a Cyan ADP (Dako Cytomation, Everlee, Belgium).
For DC and CD8+ T-cell intracellular staining, splenocytes were incubated at 37°C for 4h in complete culture medium with Golgiplug (1μl/ml; BD Biosciences). Cells were then harvested and stained with extracellular mAbs. Intracellular staining (IFN-γ, clone XMG1.2 and IL-12p40, C15.6) was then done following the manufacturer’s protocol (Cytofix/Cytoperm; BD Biosciences).
For DC sorting and adoptive transfer, spleen cells were first labeled with anti-CD3, -CD19, -B220 and -Gr1 biotinylated mAbs for negative selection using BD IMag Streptavidin Particles Plus DM (BD Biosciences), following the manufacturer’s protocol. Enriched splenocytes were then injected or stained to sort CD11chighCD11b-CD205+CD24+ CD8α- DCs (termed CD8α- DCs) or CD8α+ DCs (termed CD8α+ DCs) on a BD FACSAria II Cell sorter (BD Biosciences).
For phagocytosis assays, spleen cells from neonates were cultured with Lm-GFP (MOI 1:5) for 2 hours and stained with CD11c, CD11b, CD205 and CD8α specific mAbs for FACS analysis.
30x106 pre-purified C57BL/6 CD45.2+ neonatal splenocytes, collected from 20 to 30 neonates by negative selection as described above were i.v. injected in C57BL/6 CD45.1+ adults. Recipient spleen cells were harvested at different time points and were stained to follow CD8α expression on CD45.2+ CD11c+CD11b-CD205+CD24+ cells by flow cytometry.
2x104/well of indicated sorted DCs (collected from 40 to 60 neonates and 5 to 7 adults) were cultured for 24h. with CpG (2μg/mL), poly (I:C) (10μg/mL) and Lm (MOI 1:1) with or without GM-CSF (20 ng/mL, R&D), IL-4 (20 ng/mL, R&D Systems) and IFN-γ (20 ng/mL, R&D). Supernatants were harvested for cytokine measurements by ELISA.
For antigen presentation assay, sorted CD8α-DCs were cultured at 5000 cells/well in complete RPMI-1640 medium at 37°C in the presence of OVA257-264-peptide (1μg/mL, Polypeptide Laboratories, Strasbourg, France) or Lm-OVA (MOI 1:5) with or without GM-CSF (20ng/ml; R&D Systems), and purified NA/LE Rat Anti-Mouse CD210 (20ng/mL) (IL10R, clone 1B1.3a) or Anti-Mouse IL-12p40/p70 (600ng/mL) (clone C17.8) or isotype matched rat Ig (BD Biosciences). After 4h, isolated OT-I T cells were added to the cultures at a ratio of 1:5 (sorted DC/T). After 48h, IFN-γ was measured by ELISA.
For ex-vivo cross-presentation assay, neonates were injected i.p. with 106 CFU of Lm-actA-/- OVA. Neonatal CD8α- DCs were sorted 24h later and co-cultured at 2x104 cells/well with OT-I T cells at a ratio 1:5, with or without GM-CSF (20ng/ml). IFN-γ was measured by ELISA after 48h.
For total RNA Quantitative real time PCR and microArrays, neonates and adults were injected or not with 5 x 105 or 5 x 106 CFU of Lm-actA-/- for 24h.
Total RNA from 40,000–60,000 sorted neonatal or adult CD8α- DCs or adult CD8α+ DCs (collected from 40 to 60 neonates and 5 to 7 adults) was extracted with phenol/chloroform and purified with the RNeasy microkit (Qiagen) according to manufacturer’s instructions. For quantification of transcripts, reverse transcription and quantitative real-time PCR were performed in a single step using the TaqMan RNA Amplification (Roche Diagnostics) on a Lightcycler 480 apparatus (Roche Diagnostics). For individual samples, mRNA levels were normalized to those of β-actin. Sequence of primers and probes are available on request. For microArrays, total RNA was amplified using the Ovation PicoSL WTA System V2 (NuGen), labeled with biotin using the Encore BiotinIL Module (NuGen), and applied on Illumina HT12 bead arrays at the GIGA-GenoTranscriptomics platform (Liège, Belgium). Microarray data (derived from Affymetrix GeneChip arrays HG-U133 plus 2.0) from CD8α- neonatal and CD8+ adult DCs samples were obtained from the National Center for Biotechnology Information Gene Expression Omnibus.
IL-12p40, IL-10 and IFN-γ Duoset ELISA kits and Mouse IL-23 Quantikine ELISA Kit (R&D, Minneapolis, USA) were measured in culture supernatant according to the manufacturer’s instructions. For IL-12p70 ELISA assays, culture supernatant were measured as previously described (23).
Data are expressed as mean ± SEM. Statistical comparison between experimental groups was analyzed using a two-tailed nonparametric Mann-Whitney test for the CFUs, absolute number/% of cells and cytokine levels or with the logrank test for survival curves (GraphPad Prism, GraphPad Software, Inc.). p values less than or equal to 0.05 were considered significant. * = p<0,05, ** = p<0,01, *** = p<0,001.
|
10.1371/journal.pmed.1002311 | Reducing US cardiovascular disease burden and disparities through national and targeted dietary policies: A modelling study | Large socio-economic disparities exist in US dietary habits and cardiovascular disease (CVD) mortality. While economic incentives have demonstrated success in improving dietary choices, the quantitative impact of different dietary policies on CVD disparities is not well established. We aimed to quantify and compare the potential effects on total CVD mortality and disparities of specific dietary policies to increase fruit and vegetable (F&V) consumption and reduce sugar-sweetened beverage (SSB) consumption in the US.
Using the US IMPACT Food Policy Model and probabilistic sensitivity analyses, we estimated and compared the reductions in CVD mortality and socio-economic disparities in the US population potentially achievable from 2015 to 2030 with specific dietary policy scenarios: (a) a national mass media campaign (MMC) aimed to increase consumption of F&Vs and reduce consumption of SSBs, (b) a national fiscal policy to tax SSBs to increase prices by 10%, (c) a national fiscal policy to subsidise F&Vs to reduce prices by 10%, and (d) a targeted policy to subsidise F&Vs to reduce prices by 30% among Supplemental Nutrition Assistance Program (SNAP) participants only. We also evaluated a combined policy approach, combining all of the above policies. Data sources included the Surveillance, Epidemiology, and End Results Program, National Vital Statistics System, National Health and Nutrition Examination Survey, and published meta-analyses.
Among the individual policy scenarios, a national 10% F&V subsidy was projected to be most beneficial, potentially resulting in approximately 150,500 (95% uncertainty interval [UI] 141,400–158,500) CVD deaths prevented or postponed (DPPs) by 2030 in the US. This far exceeds the approximately 35,100 (95% UI 31,700–37,500) DPPs potentially attributable to a 30% F&V subsidy targeting SNAP participants, the approximately 25,800 (95% UI 24,300–28,500) DPPs for a 1-y MMC, or the approximately 31,000 (95% UI 26,800–35,300) DPPs for a 10% SSB tax.
Neither the MMC nor the individual national economic policies would significantly reduce CVD socio-economic disparities. However, the SNAP-targeted intervention might potentially reduce CVD disparities between SNAP participants and SNAP-ineligible individuals, by approximately 8% (10 DPPs per 100,000 population). The combined policy approach might save more lives than any single policy studied (approximately 230,000 DPPs by 2030) while also significantly reducing disparities, by approximately 6% (7 DPPs per 100,000 population).
Limitations include our effect estimates in the model; these estimates use interventional and prospective observational studies (not exclusively randomised controlled trials). They are thus imperfect and should be interpreted as the best available evidence. Another key limitation is that we considered only CVD outcomes; the policies we explored would undoubtedly have additional beneficial effects upon other diseases. Further, we did not model or compare the cost-effectiveness of each proposed policy.
Fiscal strategies targeting diet might substantially reduce CVD burdens. A national 10% F&V subsidy would save by far the most lives, while a 30% F&V subsidy targeting SNAP participants would most reduce socio-economic disparities. A combined policy would have the greatest overall impact on both mortality and socio-economic disparities.
| Suboptimal diet is a leading cause of cardiovascular disease, death, and health disparities. Dietary policies have the potential to reduce this burden.
However, the potential benefits of policies targeting fruit, vegetable, and sugar-sweetened beverage consumption in the whole US population and among those participating in the Supplemental Nutrition Assistance Program (SNAP) have not been quantified.
We modelled and compared the potential benefits of several dietary policies targeting fruit, vegetable, and sugar-sweetened beverage consumption.
We found that a modest universal reduction in fruit and vegetable prices (10% subsidy) was most likely to reduce cardiovascular disease mortality, whilst a 30% fruit and vegetable subsidy offered to SNAP participants appeared most promising for reducing disparities in cardiovascular disease mortality.
Finally, we found that a combination of all these policies potentially offered the biggest benefits in terms of reducing cardiovascular disease burden and also reducing disparities.
Our findings highlight the potentially powerful effects of fiscal measures targeting diet in the US. Dietary policies could potentially reduce cardiovascular disease, death, and associated disparities.
A modest subsidy of fruits and vegetables for all, accompanied by a larger subsidy for SNAP participants, might be most beneficial in terms of reducing the disease burden and disparities.
| Cardiovascular disease (CVD) is declining in the US [1–3]. However, CVD remains the leading cause of mortality, generating approximately 800,000 deaths and 6 million hospital admissions annually [2]. Crucially, these burdens are highly unequal across the population, in particular according to socio-economic status (SES) [4].
Among modifiable risk factors [5], insufficient consumption of fruits and vegetables (F&Vs) [6–8] and excess intake of sugar-sweetened beverages (SSBs) [9] are important contributors to CVD. Furthermore, dietary patterns and intakes of these foods are worse among low-SES groups [10], making them important dietary targets for policy makers wishing to reduce CVD and also decrease disparities [11]. Policies to reduce F&V prices and increase SSB prices are effective measures for altering consumption, and may be especially effective among individuals with lower SES [12], who have worse CVD health [13]. However, the quantitative impact of different dietary policies on CVD mortality and socio-economic disparities is not well established.
To address these gaps, we quantified and compared the potential effects on total CVD mortality and CVD socio-economic disparities of specific dietary policies to increase F&V consumption and/or decrease SSB consumption in the US population up to 2030. Using empirical estimates of policy and food consumption effects, we evaluated both national policies and targeted policies for the Supplemental Nutrition Assistance Program (SNAP), the largest federal feeding programme, which serves approximately 46 million low-income Americans. For comparison, we also evaluated the potential impact of a national mass media campaign (MMC) on CVD deaths and socio-economic disparities.
We modelled the potential effects of specific US dietary policies targeting F&Vs and SSBs from 2015 to 2030 using the previously validated US IMPACT Food Policy Model. We quantified and compared the associated change in dietary intake for a national MMC, national fiscal policies targeting F&Vs or SSBs, and a F&V targeted policy in SNAP participants, all by age, sex, and SNAP participation status. We also evaluated the joint associated change in dietary intake of combining these policies. We modelled the comparative effects upon coronary heart disease (CHD), stroke, and total CVD mortality, as well as the effect on CVD socio-economic disparities, comparing SNAP participants and SNAP-ineligible individuals.
The US population was estimated using data from the Surveillance, Epidemiology, and End Results Program single-year population estimates, stratified by sex and age in 10-y age groups (25–34, 35–44,…, 85+) [14]. Population projections to 2030 were sourced from the US Census Bureau 2012 National Population Projections [15]. Based on the number of annual CHD, stroke, and total CVD deaths (ICD codes I20–I51, I60–69) from 1979 to 2012 from the National Vital Statistics System [16], we projected mortality trends, by age and sex, to 2030 as previously described [17]. This allowed us to incorporate estimates of continuing declining trends in CVD mortality, rather than utilising only current CVD mortality rates as in other similar studies [18,19], a major advantage to avoid overestimating the benefits of any CVD intervention.
Data on age, sex, and SNAP participation status were acquired from the National Health Interview Survey (NHIS), 2000–2009. NHIS data were subsequently linked to mortality data from the Public-Use Linked Mortality Files (2000–2011), which provide follow-up for the NHIS sample through December 31, 2011. In total, we evaluated 499,740 adults aged ≥25 y who provided information on age, sex, and SNAP participation. Further details are published elsewhere [13]. This led to stratification by SNAP participation and eligibility: SNAP participants, SNAP-eligible non-participants, SNAP-ineligible individuals. Data on current consumption levels of F&Vs and SSBs, by age, sex, and SNAP status, were obtained from the nationally representative National Health and Nutrition Examination Survey (NHANES) 2009–2012 [20] using two consecutive 24-h dietary recalls, further incorporating projected intake forecasts derived from NHANES data from 1999–2012 [10] (S1 Appendix).
The policy scenarios modelled were as follows: (a) a national MMC to increase F&V consumption and reduce SSB consumption targeting all American adults aged 25+ y, (b) a national tax to increase SSB prices by 10% (10% SSB tax), (c) a national subsidy to reduce F&V prices by 10% (10% F&V subsidy), (d) a SNAP-specific subsidy to reduce F&V prices by 30%, similar to the successful US Department of Agriculture (USDA) Healthy Incentives Pilot (HIP) in Massachusetts [21] (SNAP 30% F&V subsidy), and (e) a combination of all four policies above: a 10% SSB price increase for all, a 30% price reduction in total F&Vs for SNAP participants and a 10% price reduction for SNAP-eligible non-participants and SNAP-ineligible individuals, and a national F&V and SSB MMC for all (combined policy).
The baseline scenario of “no intervention” assumes that recent and current trends in consumption [10] simply continue. The combined policy scenario considers the associated change in dietary intake of each individual scenario as additive. However, the associated change in F&V intake for the 30% subsidy for SNAP participants is lower in the combined policy scenario than in isolation owing to different elasticities used for the universal (first 10% of the subsidy) and targeted (remaining 20% of the subsidy) aspects. This is explained further below.
Extensive data from cross-sectional price elasticity and intervention studies support the association of changes in food or beverage prices with changes in consumption. We derived estimates from a recent meta-analysis of interventional and prospective longitudinal studies that directly evaluated the changes in consumption associated with changes in price [22]. In these studies, a 10% reduction in the price of F&Vs increased consumption by approximately 14% (95% CI 11%–17%), and a 10% increase in the price of SSBs reduced consumption by approximately 7% (95% CI 4%–10%). These pooled findings are broadly consistent with additional recent empirical evidence [12,23]. A recent evaluation of the USDA HIP showed that among SNAP-participating or -eligible populations, a 30% subsidy on the price of F&Vs purchased exclusively through the SNAP Electronic Benefits Transfer (EBT) card increased consumption by some 27% [21].
We recognised consistent evidence for differential associations by SES. For our national policies, we modelled a price elasticity gradient between SNAP participants and the SNAP-ineligible population of 50%, based on published estimates of the differential social associations of prices with food consumption [12,24]. We recognised that to achieve a 10% increase in national SSB prices, the actual implemented policy, e.g., an excise tax, might need to be modestly larger to offset any part of the tax that might be absorbed by industry. Thus, our findings should be interpreted as the likely association for the final retail price change of any tax or subsidy policy, and the specific policy formulation to achieve this price change could vary (e.g., agricultural subsidy, retailer subsidy, excise tax, sales tax). Methodology and sources are detailed further in S1 Appendix and S3 Table.
We estimated the associated outcomes of a national MMC based on a meta-analysis of the association of national MMCs [25], including the US national 5 A Day campaign [26], with dietary habits, including F&V and SSB consumption. That pooled analysis indicated that a typical national MMC increases F&V intake and reduces SSB intake by approximately 7% each (95% CI 4%–9%). We further accounted for potentially varying coverage by age and sex using data from the 5 A Day campaign [26] (S1 Table). We modelled the associated outcomes of a 1-y MMC in 2015, assuming a 7% national association with increased F&V consumption at 1 y, which then fell linearly to a 20% residual association at year 5 (minimum estimate 5%, maximum estimate 40%), which then persisted to year 15 (i.e., 2030).
The effects of changes in consumption of F&Vs and SSBs upon CVD mortality were obtained from meta-analyses of prospective cohort studies and randomised trials [27] of the direct etiologic effects of fruit, vegetable, and SSB intake on CHD and stroke. These effect estimates are based on interventional and prospective observational studies (not exclusively randomised controlled trials [RCTs]). They are thus imperfect and should be interpreted as the best available evidence. The association sizes for the association of the policy components below (price changes and MMCs) with fruit, vegetable, and SSB consumption are aggregate estimates only, as provided in each respective meta-analysis, although we do include MMC coverage estimates stratified by age and sex taken from the evaluation of the 5 A Day campaign [26].
The US IMPACT Food Policy Model is an adaptation of the CHD IMPACT Model [3,28], and the IMPACT Food Policy Model has previously quantified potential health gains from healthier food policies in the UK [29–31] and Ireland [32]. The validated IMPACT methodology to calculate deaths prevented or postponed (DPPs) has been described [33] and is detailed in S1 Appendix. Briefly, using mortality trends (1979–2012), we estimated baseline mortality projections (no intervention) for each year from 2015–2030 for CHD and stroke [17] to provide the number of stroke and CHD deaths expected each year, stratified by age and sex. The crucial estimation quantifies the steadily declining CVD mortality rates in the US, and thus avoids substantial overestimation of the potential benefits of any preventive intervention [17]. To stratify these projections by SNAP status, we maintained the mortality rate ratio between each group throughout the 15-y projected period for baseline number of deaths.
We estimated the change in intake of F&Vs and SSBs by applying the appropriate intervention association measure to the baseline intake data for each stratum. We ran the model to calculate DPPs. We assumed equal coverage of price change policies among the relevant population in each scenario. The time lag from a price change policy being implemented to the subsequent change in F&V or SSB consumption was assumed to be less than a year; hence, no time lag was modelled. Finally, we assumed a sustained impact of the policy throughout the 15-y period, i.e., no attrition.
The US IMPACT Food Policy Model was used to calculate the expected change in numbers of CHD and stroke deaths attributable to changes in diet intake in our analysis. We first estimated the effect of each given policy scenario upon F&V and SSB intake. We then used the best evidence of effect size upon CHD and stroke for fruits, vegetables [27], and SSBs separately, stratified by age and sex. This provided the policy scenario association with CHD and stroke mortality, hence the “intervention expected number of deaths” following the intervention. The difference between the baseline and intervention expected deaths provided the cumulative DPPs from 2015 to 2030.
We used probabilistic sensitivity analysis to estimate the effect of uncertainty in key model parameters with given input probability distributions. For each policy scenario, we performed 10,000 iterations of the full model in R, version 3.2.2 [34], providing 95% UIs. The key parameters included the associated outcomes of the MMC, the association of price reduction with F&V intake and of price increase with SSB intake, consumption of F&Vs and SSBs, the effect size for the effect of F&V and SSB consumption upon CVD mortality individually, baseline CVD mortality, and ratio of ischaemic:haemorrhagic strokes. We varied each parameter by one standard deviation above and below the mean to assess the sensitivity of DPPs to each parameter in each scenario (S2 Fig). The F&V intake projection was the parameter generating the largest variation in DPPs when varied. Details on parameters and chosen distributions for the Monte Carlo simulation are available in S3 Table.
There are approximately 21.6 million SNAP participant households (involving 44.5 million individual adults and children) nationwide and a further 20.9 million eligible but not participating in SNAP. The non-eligible population is approximately eight times larger, comprising approximately 173 million adults. Significant disparities in CVD mortality exist, with baseline aggregate age-standardised mortality more than 50% higher in SNAP participants compared to SNAP-ineligible adults (330/100,000 versus 206/100,000; using 2015 total population as reference) (Table 1), and even higher when only men are considered. There are also differences in the baseline intake of fruits, vegetables, and SSBs, with SNAP participants having consistently lower intake of F&Vs, and higher consumption of SSBs (Table 2). These differences are projected to persist to 2030.
The change potentially achievable in consumption of F&Vs varies substantially across the different policy scenarios, and across the SNAP groups. By 2030, the national 10% F&V subsidy would increase aggregate fruit consumption by 15% compared to baseline (129 g/d versus 112 g/d), some 17 times more than the national MMC (Table 2), with a similar association for vegetable consumption. The national 10% SSB tax would reduce SSB consumption by 8% on average, whilst the national 10% F&V subsidy would increase F&V consumption considerably in all groups. Consumption of fruits among SNAP participants in 2030 is projected to remain 28% lower than among SNAP-ineligible individuals (107 g/d versus 149 g/d). This difference is projected to be similar for consumption of vegetables. With the 30% F&V targeted subsidy for SNAP participants, the projected difference in fruit consumption is reduced to just 6% (Table 2).
The 10% F&V subsidy is the most effective single policy in reducing CVD mortality over the period 2015–2030. This policy could yield approximately 150,500 DPPs (95% UI 141,400–158,500). This mortality reduction comprises some 78,100 DPPs from CHD, and 72,400 from stroke, reducing the overall mortality rate by approximately 4/100,000 against baseline (Table 3; Fig 1). These mortality reductions are some seven times higher than those projected for the other national policies. The 10% SSB tax could reduce CVD mortality by approximately 31,000 DPPs (95% UI 26,800–35,300). This would reduce the CVD mortality rate by 0.8/100,000. Finally, the MMC, targeting both F&V and SSB consumption, would have slightly less effectiveness, representing some 25,800 (95% UI 24,300–28,500) DPPs coming equally from CHD and stroke DPPs. The MMC policy could reduce CVD mortality by approximately 0.7/100,000. However, the majority of DPPs would be gained in the first year of the media policy, after which the benefit would fall substantially. The 30% F&V targeted subsidy for SNAP participants could reduce CVD mortality by approximately 35,100 DPPs (95% UI 31,700–37,500), representing a 0.9/100,000 decrease in mortality rate. The combined policy would be more effective than any single policy, generating approximately 230,000 DPPs (95% UI 215,800–237,100) over the 15-y period. This would comprise approximately 90,700 stroke DPPs and 137,300 CHD DPPs, reducing the CVD mortality rate by approximately 6.1/100,000 (Table 3).
In absolute terms, of the individual policies, the targeted SNAP 30% F&V subsidy would yield the largest number of averted deaths in SNAP participants, yielding approximately 35,100 DPPs (95% UI 31,700–37,500) (Table 4; Fig 2). This represents a reduction in CVD mortality of 9.5/100,000 in this group compared to those ineligible for SNAP. This is more than two times, five times, and 11 times more DPPs that would be generated in the SNAP group compared with the universal 10% F&V subsidy, 10% SSB tax, and media campaign, respectively. In relative terms, the targeted SNAP 30% F&V subsidy would also be the most effective in reducing CVD disparities. All of the national policies would reduce CVD disparities between SNAP participants and the SNAP-ineligible population, but less than the targeted 30% F&V subsidy. The national 10% F&V subsidy and 10% SSB tax were approximately three and five times more effective at reducing disparities than the media campaign (Table 4). The combined policy is effective at reducing disparities: this scenario generates 11.6 DPPs per 100,000 in SNAP participants, some 7 DPPs per 100,000 more than in the SNAP-ineligible population.
The combined policy might best achieve the goal of reducing both the overall mortality burden and CVD disparities (Table 4; Fig 3). This approach could potentially generate approximately 12,800 DPPs in 2030 and thus reduce CVD disparities by approximately 6.0 deaths per 100,000. Whilst the national 10% F&V subsidy might generate almost 8,900 DPPs in the year 2030 alone, its effectiveness in reducing CVD disparities would be substantially lower than that of the combined policy. In contrast, although the targeted 30% F&V subsidy could generate substantially fewer DPPs than the national 10% F&V subsidy (approximately 2,100) in 2030, this might be the most effective policy for reducing CVD disparities (by approximately 8.5 deaths per 100,000 in 1 y) (Table 4; Figs 3–5).
Our study suggests that reducing the price of healthy foods for SNAP participants whilst additionally reducing SSB consumption nationally through taxes and a media campaign could potentially reduce socio-economic disparities in CVD mortality and powerfully improve dietary quality, the leading risk factor for CVD.
This is the first US study to our knowledge to compare the likely effects of national policies targeting F&V and SSB consumption and a policy targeting F&V consumption in the SNAP population upon CVD mortality and socio-economic disparities. Policies effectively increasing F&V consumption or reducing SSB consumption might powerfully reduce CVD mortality and disparities. Further, a combination of these policies could be even more powerful. Whilst all four individual policies would result in reductions in deaths by 2030, the magnitude and rate of such reductions differed substantially. The 10% F&V subsidy might reduce CVD mortality by some 2.1%, saving approximately 150,500 deaths from 2015 to 2030. It might thus be approximately five times more effective than the MMC. The SNAP 30% F&V subsidy might be four times less effective at reducing total US CVD mortality than the national 10% F&V subsidy, but could be the most effective approach in reducing CVD socio-economic disparities.
The findings of this study have important implications for crafting specific price and incentive policy approaches to optimise access to SSBs and F&Vs, respectively. F&Vs have high production costs because certain crops are especially susceptible to adverse weather, have limited storage time, often have to be transported with temperature control, or typically have to be hand-picked or hand-sorted [35]. The ultimate price that consumers pay for F&Vs is affected by policies and practices that have impact across the entire food production system [36], including international trade agreements, immigration law, import/export policies, as well as the technology used to harvest and transport fragile crops across the globe. Embedding pricing incentives systematically within government feeding programmes such as SNAP could increase the purchase and consumption of F&Vs within low-income populations. These benefits could be extended if Electronic Benefits Transfer was integrated into all farmers markets, allowing recipients to authorise transfer of their government benefits to the retailers in local markets. Retail outlets where consumers make their final purchase are playing an increasingly important role in food pricing. Non-traditional retailers/discount stores are making F&Vs more affordable. There are also efforts underway by large retailers to encourage local sustainable agriculture to support the availability and affordability of fresh fruits, vegetables, and other specialty crops in their stores. Other pricing-related policy approaches could extend to growers, providing them with more accessible crop insurance, agricultural subsidies for growing specialty crops, or government incentives to diversify crops across base acres of land. Such programmes could be financed through revenue raised by the modelled 10% SSB price increase or other public health taxes. Thus, a system-wide approach to price strategies might be particularly effective for improving diet [37].
The differing effectiveness in reducing aggregate CVD mortality across the four individual policies can be attributed to several factors. F&V consumption at baseline was higher than SSB consumption, and the elasticity of a F&V subsidy is greater than that of a SSB price increase; with these factors coupled together, a 10% price reduction for F&Vs results in a greater change in consumption than a 10% price increase for SSBs. Further, SSB consumption is highest in younger age groups, where CVD mortality is low. Whilst the 10% SSB price increase has an equally large proportional association with consumption across age groups, the differences in baseline consumption across age groups result in smaller absolute dietary reductions in the middle and older age groups, which experience the majority of the CVD burden. The apparent effectiveness of each policy also varied by SNAP group. Baseline CVD mortality is approximately 60% higher in SNAP participants [13] compared to those not eligible for SNAP. SNAP participants are also more price sensitive. Despite this, the F&V subsidy resulted in a larger absolute increase in consumption of F&Vs in SNAP-eligible non-participants and SNAP-ineligible individuals. This was due to the consistently large relative association with consumption (14% increase) and baseline F&V intake being much higher in these two groups.
Low intake of F&Vs is a risk factor for CVD as well as certain cancers, and intake is often lowest in the most deprived groups in society, thus widening disparities. In the US between 1999 and 2012, F&V consumption remained [10] substantially below the recommended amounts of 2.5 cups of vegetables and 2 cups of fruit per day [38], while disparities by income, education, and race/ethnicity did not improve. SSB intake declined by only about half a serving per day during this period, thus remaining high and with significant remaining disparities [10].
Public health strategies can aim to improve the environment (“structural policies”) or facilitate behaviour change in individuals using their personal resources (“agentic policies”) or do both [39]. Most national governments currently favour agentic policies for dietary change, rather than population-level structural policies [40]. This can be contrasted to many other existing government health- and safety-focused policies and standards that strongly favour structural approaches, such as biosafety and food contaminants, water and air safety, and toy, motor vehicle, housing, and occupational safety. Structural interventions, which are not dependent on individual responses, are generally the most effective and sustainable, as well as being the most equitable [39,40]. One additional important policy, not modelled here, would be to consider removing SSBs from SNAP-eligible items. Our findings suggested that a combined national fiscal intervention, enhanced by additional intervention among those with lower resources and worse diet and health, would be most effective in reducing the dual burdens of CVD mortality and socio-economic disparities. These results are consistent with the notion of “proportionate universalism” [41,42], and lend support to the idea that a combination of structural and agentic policies appears more effective for reducing the unequal burden of CVD in populations [40]. We did not model policy costs, nor the cost-effectiveness of the policy scenarios. However, not accounting for potential savings, the subsidies are likely to cost more than the media campaign to implement and run, whilst the SSB tax would be revenue-raising; hence, combining this tax with a subsidy might make the policy fiscally neutral. These novel results might be useful to inform policy makers in the US, such as those developing the new USDA Farm Bill, which includes SNAP, as well as leaders in professional advocacy associations such as the American Heart Association.
This study has several strengths. We used nationally representative datasets encompassing the US adult population [14,16]. Further, we used comprehensive meta-analyses for effect sizes for the effect of F&V [27] and SSB consumption upon CVD mortality and for the associations of each policy with F&V intake within the US population [22,25]. The effect size for the effect of SSB consumption upon CVD accounts for direct effects upon CHD only and hence may underestimate the total effects, which include potential effects upon stroke mediated by change in body mass index. Using the HIP elasticity [21] for a targeted F&V price reduction was helpful, as was stratifying potential policy associations by SNAP group. Further, our health outcomes analysis sensibly assumes that the recent declines in CVD mortality will likely continue [17], unlike more conventional methods, which simply use a static baseline. If, in future, mortality rates plateau (as already seen in young adults [43]) or even increase [43,44], mortality savings from the modelled policies would be even greater. Despite our baseline mortality projections accounting for recent trends in CVD mortality, we do not account for competing risks of other diseases such as cancers over the 15-y period modelled, nor additional change in the policy environment over this period.
This study also has limitations. The effect of F&V and SSB consumption upon CHD and stroke mortality, and the association of each policy with consumption, are taken from comprehensive meta-analyses of interventional and prospective observational studies; they are thus imperfect estimates [22,25,27,45] compared to dietary RCT-level evidence such as that from the PREDIMED trial [46]. Further, an increase in consumption of F&Vs would have further benefits upon incidence of diabetes and some cancers not modelled here. We also assumed a short lag time between policy implementation and reductions in CVD mortality. However, evidence consistently supports this assumption [47]. We did not account for additional (new) dietary policies being implemented over the 15-y period such as those targeting salt consumption, an important CVD risk factor. Dietary policies targeting salt have been effectively implemented in several countries and the US Food and Drug Administration has proposed a voluntary reformulation strategy. Modelling the potential health effects of the proposed reformulation would be of real use to policy makers. Our model analyses potential effects of food policies upon CVD mortality and disparities in CVD mortality; however, such policies would undoubtedly have effects upon CVD incidence and CVD health care costs as well. Whilst we incorporated coverage estimations by age and sex for the MMC using data from the nationwide 5 A Day campaign [26], less information exists regarding the decaying impact of media campaigns. We therefore assumed an approximate 20% residual effect after 5 y. However, this assumption was tested robustly, using wide uncertainty parameters (5%–40%). We estimated a gradient of price elasticity for SSBs and F&Vs between SNAP participants and the population not participating in SNAP using the best available evidence [12,21]. However, we lacked the detailed data needed to stratify for differing price elasticities of demand by age or sex. This could therefore lead to underestimation of the association of changes in consumption from such policies and reduction in socio-economic disparities. Similarly, we did not report life years gained and hence may have under-reported the associated outcomes of these policies for disparities, given the younger average age of CVD incidence in lower income groups [13]. We did not explicitly account for any substitution effects when increasing F&V consumption; however, the meta-analyses from which model parameters were derived used observed effects, thus accounting for average actual population substitutes and compliments. Furthermore, focused efforts to encourage specific substitutions could make such interventions even more effective. Future research addressing the cost-effectiveness of such specific food policies is also warranted.
Fiscal strategies targeting diet might substantially help to reduce the unequal CVD mortality burden in the US. All four individual dietary policies could be effective, whilst a combination of national and targeted policies might be even more powerful in reducing both CVD mortality and socio-economic disparities.
|
10.1371/journal.pcbi.1000968 | Infectious Disease Modeling of Social Contagion in Networks | Many behavioral phenomena have been found to spread interpersonally through social networks, in a manner similar to infectious diseases. An important difference between social contagion and traditional infectious diseases, however, is that behavioral phenomena can be acquired by non-social mechanisms as well as through social transmission. We introduce a novel theoretical framework for studying these phenomena (the SISa model) by adapting a classic disease model to include the possibility for ‘automatic’ (or ‘spontaneous’) non-social infection. We provide an example of the use of this framework by examining the spread of obesity in the Framingham Heart Study Network. The interaction assumptions of the model are validated using longitudinal network transmission data. We find that the current rate of becoming obese is 2 per year and increases by 0.5 percentage points for each obese social contact. The rate of recovering from obesity is 4 per year, and does not depend on the number of non-obese contacts. The model predicts a long-term obesity prevalence of approximately 42, and can be used to evaluate the effect of different interventions on steady-state obesity. Model predictions quantitatively reproduce the actual historical time course for the prevalence of obesity. We find that since the 1970s, the rate of recovery from obesity has remained relatively constant, while the rates of both spontaneous infection and transmission have steadily increased over time. This suggests that the obesity epidemic may be driven by increasing rates of becoming obese, both spontaneously and transmissively, rather than by decreasing rates of losing weight. A key feature of the SISa model is its ability to characterize the relative importance of social transmission by quantitatively comparing rates of spontaneous versus contagious infection. It provides a theoretical framework for studying the interpersonal spread of any state that may also arise spontaneously, such as emotions, behaviors, health states, ideas or diseases with reservoirs.
| Information, trends, behaviors and even health states may spread between contacts in a social network, similar to disease transmission. However, a major difference is that as well as being spread infectiously, it is possible to acquire this state spontaneously. For example, you can gain knowledge of a particular piece of information either by being told about it, or by discovering it yourself. In this paper we introduce a mathematical modeling framework that allows us to compare the dynamics of these social contagions to traditional infectious diseases. We can also extract and compare the rates of spontaneous versus contagious acquisition of a behavior from longitudinal data and can use this to predict the implications for future prevalence and control strategies. As an example, we study the spread of obesity, and find that the current rate of becoming obese is about 2 per year and increases by 0.5 percentage points for each obese social contact, while the rate of recovering from obesity is 4 per year. The rates of spontaneous infection and transmission have steadily increased over time since 1970, driving the increase in obesity prevalence. Our model thus provides a quantitative way to analyze the strength and implications of social contagions.
| Social network effects are of great importance for understanding human behavior. People interact with a varying number of individuals and with some individuals more than others, and this affects behavior in fundamental ways. Sociologists have long studied social influence through networks, and networks now routinely appear in investigations from other fields, including economics [1], physics [2], public health [3] and scientific publishing [4], [5]. Extensive reviews of social networks analysis, including investigations of their structure and their effect on social dynamics, include Mitchell [6], Wasserman [7],Watts [2], Rogers [8], Jackson [1], and Smith [9]. Networks have also long been known to be important in many areas of biology (reviewed by [10]), including ecological food webs and the evolution of cooperation [11]–[14]. Social networks have also been studied as determinants of health (reviewed by Smith [9]), ranging from determining the patterns of infectious disease spread [15] to the propagation of phenomena such as emotions [16]–[18], smoking cessation [19], obesity [20], suicide [21], altruism [22], anti-social behavior [23], and online health forum participation [24]. These studies suggest that on top of the physical environment, the social environment can also be an important contributor to health. They have lead to suggestions that public health interventions must be designed that work with the network structure and that the network can be exploited to spread health related information [9], [25].
Within network studies, much work has focused on how information, trends, behaviors and other entities spread between the individuals in social networks. These processes are generally referred to as ‘contagion’. Such suggestions of contagious dynamics and the possible relevance of network structure can be rigorously examined using mathematical models of contagious processes. These can then be used to obtain accurate measures of expected prevalences, interventional efficacy, and optimized information flow. Many previous models have been proposed to study influential interactions between individuals. Most of these have considered well-mixed populations, although more recent work has focused on network-structured populations. The most well studied are classic epidemiological models (like SIS and SIR) for the spread of microbial infectious diseases [26], including spread in network-structured populations [27]–[30], [31], [15]. Various related processes have been used to model social influence, with important contributions including the same epidemiological models [32], [33], diffusion models [8], [34]–[38], statistical mechanics type interactions [39], [40], and threshold models [41](reviewed by Jackson [1] and Newman et al. [42]).
Each of these models, however, has one or more properties that are problematic for studying social contagion. Many do not capture the probabilistic nature of contagion, or the asymmetry inherent in traditional infectious disease (where the infected state spreads through social contagion whereas the non-infected state does not). Others only consider well-mixed populations, where everyone is influenced by everyone else, ignoring the effect of network structure. Most models inspired by epidemiology are not directly applicable to the social spread of other phenomenon, because many phenomena that spread by social contagion may also arise spontaneously. That is, it is possible to adopt a trend or behavior, or obtain information, from an outside source, without directly ‘catching’ it from a contact in the network. In other words, on top of the probability of obtaining the infection from each infected contact, there is also a non-zero probability of ‘automatically’ obtaining the infection, independent of the local network. This ‘automatic’ non-social infection is not included in traditional infectious disease models. Economic models for the diffusion of innovations, based on early work by Bass [34], do take into account ‘automatic’ infection. Individuals move from ‘susceptible’ (non-adopter) to an infected (adopter) state by adopting a new product or idea, influenced by both social and non-social factors. However, these models do not allow for recovery; because the innovation adoptions are assumed to be permanent changes in behavior, individuals never move back to a susceptible state. This results in the entire population becoming adopters at equilibrium. This does not reflect the dynamics of many phenomena that spread socially, which may be repeatedly acquired and lost (for example, happiness or obesity). Through a balance of infection and recovery, a steady-state with multiple states of individuals coexisting can be reached. Finally, most previous models make assumptions about the type of interaction between individuals, the particulars of which are not usually validated with real data. Yet, long term behavior of a model and the prevention strategies it suggests can depend critically on the specifics of the interaction assumptions.
Here, we introduce a new model to study the spread of entities in a social network which has all of the important properties listed above. We then analyze its characteristics and show how it can be applied in different contexts. This model is an extension of the classical infectious disease model, combining features from other models mentioned above. It describes infections that can be contracted both spontaneously and through social (network-structured) transmission, and allows for recovery from infection. As an example, we focus on the spread of obesity in the Framingham Heart Study (FHS) network. The interaction assumptions of the model will be validated using longitudinal network transmission data. We show how we can quantitatively assess the values for the rate of adopting a trend spontaneously versus by contagion to determine the extent to which social transmission is important. We use it to predict prevalences and intervention effectiveness (i.e. get quantitative output, not just qualitative behavior). The results of this model are very different from models with other interaction assumptions, such as the ‘majority rules’ models. We will show that transmissive components are often small compared to the automatic component, but may still contribute materially to prevalence levels. Lastly, we will use pair-wise approximations to generate analytic results for infections in network-structured populations, as well as presenting simulations using a real social network.
In the simplest infectious disease models [26], individuals are classified as occupying one of two states: ‘susceptible’, meaning they do not have the disease, and ‘infected’, meaning they do have the disease. The disease can be transmitted to a susceptible person when they come into contact with an infected person. The rate of this disease transmission from infected to susceptible is defined as , the transmission rate. Once an individual is infected, they recover from the disease at a constant rate , regardless of their contacts with susceptibles or infecteds. In one class of disease models (susceptible-infected-recovered, or SIR), recovered individuals become immune to further infection and enter a ‘recovered’ state. However, behaviors, trends, health states, etc, can occur many times over an individual's life, and therefore we assume infected individuals return to the susceptible state after recovering. This form of susceptible-infected-susceptible (SIS) model is used to model infectious diseases that do not confer immunity, like many STDs.
In the standard SIS model, infection can only be transmitted by having a contact between an infected and a susceptible individual. Social ‘infections’, however, can also arise due to spontaneous factors other than transmission. Therefore, we extend the SIS model by adding a term whereby uninfected individuals spontaneously (or ‘automatically’) become infected at a constant rate , independent of infected contacts. A diagrammatic representation of our modified SIS model, which we will call SISa, is shown in Figure 1. The corresponding differential equations for a well-mixed population are described in Eq. 1(1)where is the number of infected individuals, is the number of susceptible individuals, is the population size, is the transmission rate, is the recovery rate, and is the rate of spontaneous infection. This model assumes a constant population size and neglects birth and death. The SISa model is related to infectious disease models with ‘imports’ (migration of infecteds into the population), although here the rate of spontaneous infection is proportional to the number of susceptibles, while in import models it is a constant or proportional to the total population size.
In the infectious disease literature, a disease is said to be ‘endemic’ if a stable, non-zero fraction of the population is infected at steady state. If a single infected individual is introduced to a totally susceptible population, then the average number of secondary infections they cause before recovery is called the basic reproductive ratio, . For the regular SIS model in a well-mixed population of N individuals, . An epidemic, leading to an endemic equilibrium, only occurs for , and hence is a fundamental quantity used to describe and compare infectious diseases. For the SISa model, an epidemic occurs for all parameter values, due to the spontaneous infection term. Thus, social behaviors that can be adopted independently of neighbors mean that there is no longer a threshold for the behavior to become prevalent in a population, and even in the absence of contagion there would be a non-zero steady state prevalence. Because of this, there is not an obvious definition for in the SISa model. The steady state fraction of infected individuals in a well-mixed population is given by Eq. 2.(2)
Traditional models of infection assume that the population is well-mixed. However, this assumption is unrealistic for many diseases, and also for the social spread of trends and behaviors. To account for the population structure, the infectious process can be constrained to take place on a social network. An infected individual can only pass their infection on to the suspectibles to whom they are connected. Properties of the infectious process thus depend on both the epidemiological parameters and the network structure, and there are often no longer simple analytic formulas to describe the reproductive ratio or steady state level of infection. For example, a property of disease spread on networks are spatial correlations (in the network sense) that arise between individuals in the same state. This correlation is defined as the ratio of the observed number of connections between two types of individuals to the number of connections expected if the positioning of individuals in the network was random. Spatial correlations of like individuals can be caused by an infective process spreading within a network [29], but may also be caused by confounding environmental factors which similarly influence the behavior of connected individuals, or the formation of contacts based on similar behavior (also called homophily). For a network of N individuals with a total of E connections between them, the correlation between two states X and Y is defined by:(3)
The correlation between infected individuals, , rises above one as the epidemic proceeds, due to cluster formation as infected individuals transmit to their contacts. Similarly, the correlation between infected and susceptible individuals, , drops below one. The deviation of these correlations from 1 increases with (i) the ratio of transmissive infection () to spontaneous infection () in our model (there are no correlations without a transmissive component), and (ii) the inter-connectivity (transitivity) of the network. As a result of these spatial correlations, diseases on networks can progress more slowly than their well-mixed counterparts, leading to lower basic reproductive ratios. However, heterogeneity in the number of contacts per individual acts to increase . For two networks with the same average degree, if one has a larger variance in degree, then will be increased. Thus, it is possible for diseases on networks to have lower (or nonexistent) thresholds for endemic epidemics.
There are no analytic methods to solve SIS-type dynamics on arbitrary networks without making approximations. Thus, simulations are a more accurate tool to explore theoretical disease dynamics in structured populations without making simplifying assumptions about the network structure. For scaled, well-mixed populations, the formulas given in the previous sections for and are good approximations if is replaced with , the average contacts at a given time, while fixed networks, especially if non-uniform and highly inter-connected, can deviate from these values significantly. We can use a pair-wise approximation [29], [43], [44] to formulate the infectious process on a network structure in terms of differential equations. The fundamental variables are numbers of individuals of each type, and also the pairs of individuals, [XY] (where the edges are not directional). Because [XY] = [YX], and the total individuals and total edges is constant, the system can be reduced to three equations.(4)
Here [XYZ] represents the number of situations where and X individual is connected to a Y individual who in turn is connected to a Z individual. We can approximate all these triples in terms of pairs, using a moment closure approximation ([43], Text S1), which then reduces the number of variables to three also. Then these equations can be simplified to(5)with(6)where is the number of contacts each individual has and is the transitivity of the network (the ratio of triangles to triples). Having a simplified set of equations is very useful for understanding contagion dynamics in structured populations. Integrating equations is much faster than running simulations on large networks, and from them analytic results can be derived which allows determination of parameter dependence. These equations assume that the local neighborhood for each individual is identical, that is, everyone has the same number of contacts () and the same . They thus take into account the effects of fixed network structure but not heterogeneities between individuals. In the Supplementary Information (Text S1) we have included the extension of these equations to include heterogeneities. These equations can be used to easily simulate disease spread and get expected steady state prevalences and correlations, which are very useful approximations and give insight into parameter dependence. Later, we will compare these equations to results from full simulations on realistic networks. When (which is approximately the case for most random graphs) we can get a closed-form solution for the prevalence at steady state:(7)(8)(9)
The result of a network structure is that the number of partnerships between susceptible and infected individuals quickly becomes less than if random, and so . We can compare Eq. 7 to the well mixed result (Eq. 2), and see that the effect of the network is to lower the effective transmission rate by a factor of , and hence lower the prevalence, due to these correlations that build up locally. The larger is compared to , the more network effects are important. If infection is mostly automatic (when 0), the network no longer matters. Equation 7 actually holds generally (for any homogeneous network and any value), while Equations 8 and 9 are only applicable with = 0.
Analyzing the n-regular pair-wise equations allows us to get analytic results and determine how and under what conditions network structure affects the spread of behaviors which are both spontaneously acquired and spread interpersonally. Although simple closed-form solutions do not exist when is non-zero, these equations can easily be integrated or numerically solved to get solutions. These equations ignore heterogeneities in the number of edges for different individuals, which can facilitate spread under some conditions (see supplement Text S1 for extension). Full stochastic simulations on large networks can be carried out to determine how and when the results differ.
The SISa model provides a formal way for assessing the social contagion of trends and behaviors that may be repeatedly caught and recovered from. Using data from the Framingham Heart Study (FHS) [45] we tested the validity of this model and estimated transmission parameters for various health related behaviors, though the focus here is on obesity as an example. To both demonstrate that obesity can display infectious-disease-like dynamics, and to estimate values for the model parameter , and , we use dynamic information about transitions between states based on our multiple time points of data. For data points separated by time intervals () smaller than the average time between transitions, the transition probabilities can be linearized. The probability of a transition from susceptible to infected after a time can be given by , and the probability of transition from infected to susceptible after time , by . It is necessary for the time between measurements to not be comparable to or greater than the average lifetime of a state to keep the probability of double transitions within a time interval low.
This epidemiological approach to social contagion has important differences from other models which look at correlations in present and past states of connected individuals. Here, similar to others [16], [19], [20], [46] we look at how contacts influence the transitions between states, which better captures the nature of contagion. Since we use pre-existing social ties, we do not see effects from selection bias in choosing friends with similar states. Additionally, time invariant confounding events that lead to concurrent changes in connected individuals will not show up as contagion effects in this model.
The dataset we use is a subset of individuals from the Framingham Heart Study [45]. This study was initiated in 1948 in Framingham, Massachusetts and has continued enrolling subjects through the present. We examined individuals in the Offspring Cohort, enrolled starting in 1971. Subjects come to a central facility at regular intervals (approximately every 4 years) for medical examination and collection of other survey data. Body mass index (BMI) was measured at each exam, and obesity was defined as BMI30 [47]. All other, lower, weights, which include underweight, normal range weight and over-weight, were classified as ‘not obese’. In additional to information on mental and physical health, subjects were asked to name at least one close friend at each exam, and were also connected to all first-order relatives, as well as coworkers and residential neighbors. For each subject, the following social connection data is available: (i) each other person to whom they were connected, (ii) the dates of initiation and termination of that relationship, (iii) the type of relationship (neighbour, coworker, first-degree relative, or friend), and (v) the geographic distance between the two subjects. The social network for each exam was constructed by creating a network matrix G, where if subject nominated subject as a connection before or during the time that subject was administered exam . All relationship types are mutual except for friendships, which are self-nominated, such that is possible for friendships.
To study the transmission of obesity, we examine changes in BMI between sequential exams. Seven exams were administered to the Offspring Cohort between 1971 to 2001, with network data collected for each. We examine transitions occurring between each exam. The average fraction of the network that was classified as obese increased between these seven exams, suggesting the transmission process is not yet at steady state (Exam 1: 14 obese; Exam 7: 29 obese). Each set of exams were closely and consistently spaced ( (exam 1), (exam 7)). In general when modeling an infectious process, the rates of infection and recovery are assumed to be constant over time, with the prevalence changing as the infectious process begins and finally reaches equilibrium or is eliminated. When examining the spread of obesity using longitudinal data on transitions between exams, we can actually test this assumption and detect changes in the rates themselves.
A given state is considered infectious if having more contacts in state makes you more likely to switch to state . That is, a positive relationship between the number of contacts in state and the probability to transition from state to state indicates that state is infectious with respect to state . Therefore, to test whether a given state is infectious with respect to another state , we perform an ordinary least squares (OLS) linear regression as follows. Each subject in state in exam N is coded as either having transitioned to state (transition = 1) or not (transition = 0) in exam N+1. We then regress this binary transition variable for each subject against the number of contacts in state that subject had during exam N. A significant positive correlation indicates that having more friends in state at the earlier exam makes you more likely to switch to state in the later exam. If state is infectious (a significant positive correlation exists), then the value of can be calculated from the slope of the regression line, and the value of can be calculated from the intercept. If state is not infectious (no significant correlation exists), then the value of can be calculated from the intercept. was taken as the average time between examinations, which varied between exams from 3 to 8 years. Using logistic regression as opposed to OLS regression gives very similar results, as the datapoint line is within the linear range of the logistic model.
The structure of the Framingham Heart Study social network varies over the course of time, ranging from 7500 individuals with an average of 5.3 connections each at the first exam, to 3500 individuals with 2.8 connections on average at the seventh exam. Summary statistics are presented in the supplement (Table S1). These changes in population size and average degree occur because individuals may die or drop out of the study but new individuals are not added. The network is approximately Poisson distributed (see Figure 2), although with some subjects having no connections. The transitivity is consistent over time at approximately 0.64. While neighbors were included as contacts in the study, like Fowler and Christakis [20] we find no significant trends when including neighbors, and so did not include these contacts. For friendships, we only consider the contacts of an individual to be those other individuals whom they nominated (other relationships are all mutual), and so the network is directional.
The results of infectiousness analysis for the spread of obesity between exams 4 and 5 are shown in Figure 3 as an example. Consistent with the SISa model formulation, we find a significant positive correlation between the probability of transitioning from ‘not obese’ to ‘obese’ and the number of ‘obese’ contacts (Figure 3A, coeff = 0.016, p = 0.0001), and no significant relationship between the transition from ‘obese’ to ‘not obese’ and the number of ‘not obese’ contacts (Figure 3D, coeff = 0.006, p = 0.15). Additionally we find no significant relationship between the probability of transitioning from ‘not obese’ to obese and the number of ‘not obese’ contacts (Figure 3B, coeff = −0.0005, p = 0.75), or the probability of transitioning from ‘obese’ to ‘not obese’ and the number of obese contacts (Figure 3C, coeff = −0.002, p = 0.85). The same analysis was repeated for each interval between sequential exams and very similar results were found. The full results from the regression analysis are presented in the supplement (Table S2). This suggests that obesity can indeed be modeled as an infectious process in the SISa framework, with ‘not obese’ susceptibles becoming ‘obese’ infecteds, and transmitting obesity to other susceptibles. The parameters for the SISa model can be calculated from the transition probabilities mentioned earlier, by dividing slope and intercept values by , the average time between exams. These values are reported for each exam in Figure 4, and the values at the latest exam interval are summarized in Table 1. For most recent exam, the transmission rate, , is found to be /year. The spontaneous transmission parameter is found to be /year. The recovery parameter is found to be /year. From these SISa model parameters, other values of interest can be calculated. The ‘average lifetime’ of a state is the average length of time and individual spends in this state before recovering, which was found to be 24 years for this time period. The ‘influence’ of a state is the cumulative probability that the infection will be passed from an infected to a susceptible connection before the infected individual recovers, and is observed here to be 13. The ‘cycle length’ is the average length of time between spontaneous infections, and is 56 years. The basic reproductive ratio is approximately 0.35, which implies that without spontaneous appearance, the obesity epidemic would not be self-sustaining based on transmission alone. However this calculation is an approximation since uses the formula for a population that is well-mixed but only effectively contacting a fraction of the total population at each time ( contacts), so does not factor in fixed network structure (there is no analytic formula for this situation). We observed a correlation in the positioning of obese and non-obese individuals of = 1.3 and = 0.9.
Since these rates were measured for 6 different inter-exam transitions over 30 years, we can look at how the value of these rates changes over time. Figure 4 shows the measured automatic infection (a), transmission (), and recovery rates (g) for each exam interval. Error bars are 95 confidence intervals on measurements from analyses like Figure 3. While the rate of recovery () has remained relatively constant since the 1970s, the rate of spontaneous infection () has steadily increased over time. The transmission rate, , also appears to have increased over time. These trends were tested using weighted regression (to include the different errors for each measurement) and found to be significant for and but constant for . For the rest of the study we used the time-averaged value of , . This suggests that the obesity epidemic may be driven by increasing rates of becoming obese, both spontaneously and transmissively, but not by decreasing rates of losing weight.
We also found that both happiness and depression fit the SISa model, both being contagious from a neutral emotional state [18], that smoking cessation, though not smoking itself, also fit, and that both alcohol consumption and abstinence were contagious from the opposite state (data not shown). For all of the above cases, we tested if the transition probability depended instead on the fraction of contacts in a state, instead of the number, and found no significant dependence. We also tested for dependence on other personal attributes such as age, sex and education, and found no dependence in most cases. For obesity, the transition probability from not obese to obese decreased slightly with age (coeff = −0.0012, p = 0.04). Our results show that many models of social influence make assumptions about interpersonal interactions that are not supported by this longitudinal data. One of these assumptions is the ‘majority rules’ interaction, which assumes that people will be most likely to switch to the state most of their contacts are in [40]. Here, transitions depend on the number of contacts, and only certain states (those we class as ‘infectious’) actually influence transitions (in other words, contagion is only in one direction). This has significant effects on the predictions for epidemic progression. For example, ‘majority rules’ models predict 100 infected at steady state, and that weight loss behavior spreads and so an effective intervention is to ‘pin’ certain individuals at low weights. Also, many models assume that the probability of transitioning to a state is zero if no contacts are in that state, but these results show that there is a constant probability of spontaneously becoming ‘infected’. Finally, using this framework, we can get rates for transitions, and hence have an idea for the time-course of the progression, not just the final outcome.
In this section, we will use the SISa model to make predictions and evaluate interventions for the obesity epidemic, using the parameters observed in the FHS data. For simplicity and generality, we will keep the parameters and constant at the values observed for the most recent exams, and use the time-averaged value of . Since we are mostly interested in predicting future trends, and the parameters seem to have relatively constant values over the final decade, this simplification should not affect these predictions. We also keep the network fixed at the structure observed at Exam 6, except when we compare to historic data. While the simplified pair-wise equations we present are designed for symmetrical networks, they can be approximately adapted to directional networks by letting represent the average out-degree (average number of influential contacts) instead of the total number of contacts. In the Framingham data, greater than 90 of contacts are symmetrical, and so there is little error in this approximation. For hypothetical networks were the contacts formed by out-degree and in-degree are very different sets of individuals, deviations are expected. Figure 5 shows the results of both the n-regular pair-wise equations and a full simulation on the FHS network for the spread of the obesity epidemic. The parameters used were those measured from FHS as discussed earlier. One of the important properties of the SISa model is that it always leads to a stable coexistence of both infected and susceptible individuals, with infecteds becoming 100 prevalent only in the limit as or approaches infinity. This is very different from statistical-physics-based interaction models where the population always ‘coarsens’ to everyone in a single state [40]. These results show that for the parameters measured for obesity, the pair-wise equations are not significantly different from the full simulations for predicting prevalence, and hence provide a good substitute. The reason is that the spontaneous rate () is significantly larger than the transmissive component (). For larger values of , there is a noticeable difference (shown in the next section).
This model predicts that, assuming the rates do not further change over time, the steady state proportion of obese individuals will be 42. While not great, this is a much more optimistic estimate than 100 [40]. However, all of the parameters observed in this study have an error associated with them, and so there is some uncertainty in this prediction. Figure 4 shows the ranges of the 95 confidence intervals for these values. We can estimate the uncertainty in this prediction by using first the values of these parameters, within the range of one standard deviation, that would give the highest prevalence () and then those that would give the lowest (). We used = 0.05, = 0.015 and = 0.002 to get the minimum and = 0.03, = 0.023 and = 0.008 to get the maximum. These simulations suggest the confidence interval for the expected prevalence can be approximated as 25 to 54. This model also allows us to estimate the time-course of the epidemic, and suggests it would take around 40 more years for the obesity prevalence to be within 1 of this maximum value. At the first time point in our data (1970), we measured the rates to be = 0.008, = 0.03 and = 0.001, and the prevalence to be 14. These parameters would have led to a steady state prevalence of 24, which suggests that the rates of becoming obese must have originally been much lower than this.
We can also compare historical data on the obesity prevalence (from both national studies [47] and the FHS data) to the predicted time course shown here. To generate the model prediction, we simulated an epidemic with the pair-wise equations but allowed the rate values and network parameters to change as measured from the data (see Figure 4 and Table S1). We kept constant at the average value observed, 0.035, and varied and as observed. The value for parameter measured for the transition between exam and () was used in the simulation for times (years) between the average examination dates of exams and , and then increased to for the next time interval. The same was done for . For times before the earliest data points in FHS for which we have measured rate constants(pre 1970), we assumed the epidemic was at a steady state of 14. This could be achieved, for example, with and . Figure 6 shows that there is a good match in the time course of the model with reality after 1970, with similar rates of increase in the prevalence.
We can use the pair-wise equations to see how the steady state prevalence depends on various parameters, which is especially useful to see how interventions that aim to change a certain parameter may affect the prevalence. Figure 7 shows these results. For the parameter values for obesity, although is quite large, is still important. If changes from 0 to 0.005, the expected steady state changes from around 0.35 to 0.42. However, much larger changes can be realized by decreasing or increasing . For the obesity parameters, completely removing the contagious component is only expected to change the steady state prevalence by around 7. However, changing the spontaneous infection term can have much larger effects. While a 50 change in will result in only a 3 decrease in I, cutting in half will reduce the prevalence by 15. However, a similar absolute decrease of 0.005 would also lead to a 7 difference. The efficiency of changing one parameter versus the other can be looked directly at for various parameters, which will be shown in the next section.
In this section we will examine the more general properties of ‘infections’ following SISa model dynamics. While Figure 5 showed excellent agreement between the pair-wise equations and full simulations for the time dynamics, this is not true for all parameter regimes. When is larger and is smaller (as shown in Figure 8), and the network is strongly heterogeneous (as the Framingham network is), the pair-wise model deviates more. The reason is that heterogeneous network effects become more important for larger , and the pair-wise approximations are best for homogeneous networks. The extension of the pair-wise equations to heterogeneous networks is described in the supplement (Text S1).
We can use the pair-wise equations to see how the steady state prevalence depends on various parameters, which is especially useful to see how interventions that aim to change a certain parameter may affect the prevalence. Figure 9 shows how the steady state changes with the rate of transmission, . The blue line () shows what would happen in a classical epidemic, with no spontaneous infection. When is below a certain value (), the infection does not spread. The fraction infected increases rapidly with in this regime. As soon as we add , this thresholding behavior disappears. When the steady state is less sensitive to . The red line () shows the approximate parameter values for obesity. Here although is quite large, is still important. As with classical infectious disease models [29], disease spread on a network leads to decreased , the spatial correlation between infected and susceptible individuals, and increase and , the correlation between pairs of infected individuals and pairs of susceptible individuals, respectively. If we look at , we can see that we expect there to be some correlations of infected people at some values, but not all. So while seeing spatial correlation may hint there is a inductive process, it is definitely not necessary. You can have an infectious process without seeing correlations, just like you can see correlations without it being caused by the dynamics of influence. Spatial correlation is much higher when is small.
Figure 10 shows the dependence on the rate of spontaneous infection, . The more spontaneous infection, the more infected. When is larger (red line), increasing has less effect. The green line is for the parameters measured for obesity. We can use these graphs to compare the effects of various interventions which may reduce the rate of infection. In Figure 9 (vs ), we can see the expected decrease in the prevalence of the infection for a given decrease in . Changing has more effect when is small. The rate of recovery from an infection is , and in the obesity case, represents the rate at which obese people lose weight and transition to normal BMI values, in probability per year. Higher rates of recovery lead to lower fraction infected (Figure 11). One possible intervention is to increase the rate of recovery. For low recovery values, this has a large effect on , but for around 0.04 (the value for obesity), only small changes in result from changing .
In general, the spatial correlations () are negatively correlated with the fraction infected (I); more correlations are observed when a disease is not too infectious. If the spatial correlations were fixed to be a certain value (for example obese people cluster together due to selection bias in friendships or confounding factors), then this would actually serve to slow infection. Since we do not observe contagion of losing weight, it does not seem like it would be beneficial to have an intervention which broke up obese clusters.
The most direct way to compare various parameters for spread, and therefore interventions that reduce one of the parameters, is to look directly at for various parameters ( is the steady state fraction infected, is the parameter of interest. Figure 12 shows that for most parameter regimes, it is always best to increase the recovery rate, , as a method to reduce the fraction infected, . However, for low and low , it is best to decrease the spontaneous infection term , and for a window of intermediate , it is best to decrease the transmissive component . The third plot shows the results for the value measured for obesity, and because is low here we are in a regime where it decreasing has the most effect, so this is the best intervention.
Many analytic models of network phenomenon assume the transitivity, , is zero, meaning there are no triangles in the network. This is done to get the analytic expression presented here (Eq. 2), which is not necessary to numerically integrate the pair-wise equations, as presented in the results above. In the FHS network, we observed that is 0.64, suggesting human social networks are quite transitive. We want to examine the importance of in predicting the fraction infected. For the observed value, the effect of is negligible, as shown in Figure 13. The reason is that the dominant effect here is the spontaneous infection, which does not depend on the network structure. This justifies ignoring for infections that have only low infectivity terms. However, for large values (the equivalent of 2 is shown in Figure 14) has a more pronounced effect. While for a purely infectious process (blue line), at high , a disease can die out, even for , when , this doesn't occur, but still slightly reduces the spread. It also results in more observed spatial correlation of infected individuals. Overall, there is very little effect of in the SISa model.
We've already discussed how changes in parameters of infection affect the steady state prevalence, and we can consider this an analysis of different types of public health interventions that change rates of recovery, infection or network structure. In previous analysis of the obesity epidemic done by Bahr et al [40] they suggest a strategy of ‘pinning’ groups of people to stay in a non-obese state, similar to vaccinating against an microbial disease, as a method to remove the ‘infection’ from the population. However, in the Bahr model this intervention works (if enough people are ‘pinned’) because becoming non-obese is also contagious, which we don't find in this analysis. In the classical infectious disease setting, vaccinating can lower below the threshold for disease invasion, but in the SISa model there is no threshold, and so neither mechanism makes this an effective strategy in the SISa model. Two other possible intervention strategies come out of this model. Firstly, from Eq. 7 we can see that the fraction infected decreases with , the correlation of susceptible and infected people. If an intervention actively reduced this number, by isolating or clustering infected people, this could reduce the prevalence. Secondly, the fraction infected could be reduced if it were possible to make the ‘susceptible’ state also contagious through contacts.
The SISa model offers a framework for quantitatively analyzing and predicting the public health affects of socially contagious phenomenon. Using a longitudinally measured health outcome and social network data, the SISa model can be used to determine the dynamics of a health trend in terms of rates of acquisition, recovery and inter-personal transmission. From these rates, the relative importance of social contagion can be determined, and changes in prevalence over time can be predicted. The framework can also be used to examine how these rates themselves change over time, helping to understand the mechanisms behind drastic changes in disease prevalence, such as in the obesity epidemic current effecting the United States. Finally, understanding the dynamics of a health behavior using the SISa model allows us to evaluate the benefits of various interventions, especially those that may work within social networks.
The prevalence of obesity in the Framingham Heart Study cohort has increased from 14 in the 1970s to 30 in 2000, and continues to increase. We find that the most recent rate of becoming obese is 2 per year and increases by 0.5 for each obese social contact. The rate of recovering from obesity is 4 per year, and does not depend on the number of non-obese contacts. These results show that obesity has an infectious character: obesity can be acquired through social contagion as well as through non-social factors. Examining over 30 years of data, we find that these rates have changed throughout the course of the study, with the rate of becoming obese through mechanisms other than social contagion increasing approximately twofold since 1970, and the rate of transmission increasing approximately fourfold. The rate of recovery, however, has changed little. These results suggest that social norms are changing the propensity for becoming obese by non-social mechanisms, and also magnifying the affect that obese individuals have on their non-obese contacts. It is possible that while causing changes in prevalence, these rates may also be responding to changing prevalences (i.e. more obese people leads to increased social acceptability of obesity, which leads to higher rate of becoming obese), creating a positive feedback mechanism and a continuously increasing obese fraction of the population. It has been suggested that changing social norms that stigmatized smoking may have lead to its decline [48], and just the opposite may be true for obesity [49].
Using the SISa model with these parameter values estimated for obesity, we can make predictions about the future of the obesity epidemic and the important factors controlling it. Our models suggest that if the most recent rates stay constant, the population will stabilize at 42 obese. However, it is very likely that the rates of obesity infection may continue to increase if successful interventions are not conducted. Our results show that while the rate of automatic development of obesity appears to have leveled off in the past decade, the rate of transmission has been steadily increasing.
This model allows us to can predict how much spatial correlation is expected from a purely infectious process, and compare this to what is observed in the data, which could be influenced by confounding factors and selection bias in choosing friends. A coefficient of 1 indicates that arrangement of infected nodes is random, while higher values are indicative of spatial correlations. We observed a correlation coefficient for obese individuals of 1.30, which was quite close to what was predicted from epidemic simulations (1.33). This suggests that infection alone is sufficient for explaining the observed correlations, and there may not be much selection bias or confounding factors in effect. We also show that network transitivity is not predicted to have a strong affect on prevalences when there is an automatic component to infection. However, our model also shows that contrary to popular belief, a contagious process on a network does not always result in clustering of infected individuals. This is especially true if there is a large automatic infection term, which is likely with many trends and behaviors.
The SISa approach allows us to compare the effectiveness of different classes of intervention. For the parameter range observed, we find that decreasing the rate of transmission is the most effective intervention (largest decrease in prevalence per unit decrease in rate), although decreasing the automatic infection is almost as effective. More generally, while we find that gaining weight is contagious, we do not find that losing weight is contagious. Thus it does not seem to be beneficial to ‘break-up’ clusters of obese individuals or ‘pin’ the weight of certain people in these clusters. Our results actually suggest that clusters of obese people serve to slow the spread of obesity by reducing social contagion to non-obese others outside of the clusters. Another possible intervention would involve somehow facilitating the social spread of becoming non-obese (losing weight), creating a bi-directional transmissive process.
One possible limitation of this study is the incompleteness of the social network dataset used. Because the Framingham Heart Study was not designed as a study of social networks, no attempt was made to capture all of a person's important social contacts. Many close friends of a person could be missing (usually only one friend per person was recorded) and family and coworkers who play only a small part in ones actual social network may have been counted. However, even if under-sampling of real-world contacts did occur in the FHS Network, it does not change our results qualitatively: our data clearly show that rates of becoming obese increase with the number of ‘infected’ contacts (i.e. is contagious) while the rate of ‘recovery’ to a non-obese state does not depend on contacts. However, under-sampling could quantitatively effect our measurement of the rate constants. If a constant number of contacts for each person were missed, our estimate of the y intercept of the transition graphs would be shifted up from its true value, and the actual would be smaller than the we measured. If a constant fraction of contacts for each person were missed, then our estimate of the x axis would be compressed from its true value and the slope would be increased, so then the actual value of would be smaller than the we measured. While it is likely that the FHS network underestimates the total number of contacts, the relationship to the number of ‘influential’ contacts is unclear. In this sense, the observed value of the transmission rates, , are network dependent. Additionally, network connections may be weighted differently according to their ability to transmit behaviors. Longitudinal studies designed specifically with the intent of measuring social networks and health, which carefully define contacts, such as by amount of time spent together per day, influence, etc, are an important area for future research.
It has recently been suggested that certain, particular types of latent homophily, in which an unobservable trait influences both which friends one chooses and current and future behavior, may be impossible to distinguish from contagion in observational studies and hence may bias estimates of contagion and homophily [50]. The circumstances under which this is likely to be a serious source of bias (e.g., whether people, empirically, behave in these sorts of ways), and what (if anything) might be done about it (absent experimental data of the kind that some new networks studies are providing [22]) merits further study. Observational data invariably pose problems for causal inference, and require one set of assumptions or another to analyze; the plausibility of these assumptions (even of standard ones that are widely used) warrants constant review.
The SISa model as presented here assumes that all individuals have the same probability of changing state (though not everyone will actually change state within their lifetime). It is clearly possible, however, that there is heterogeneity between individuals in these rates. We do not have sufficient data on obesity in the Framingham dataset to explore this issue, which would require observing numerous transitions between states for each individual. Exploring individual differences in acquisition rate empirically is a very interesting topic for future research, as is extending the theoretical framework we introduce to take into account individual differences.
The results we have presented here reiterate an important general principle of network processes: networks tend to magnify whatever they are seeded with, but they must be seeded with something. The increase in obesity is not purely a network-diffusion phenomenon. Automatic infection serves to start and continuously seed the epidemic. Here we show that the dominant process in the increasing prevalence of obesity is contact-independent weight gain; however, the rate of interpersonal transmission contribute significantly to the overall prevalence and appears to be increasing steadily over time. Thus consideration of social transmission and network effects is an important issue for health and policy professionals.
|
10.1371/journal.pgen.1000957 | Integration of Light Signals by the Retinoblastoma Pathway in the Control of S Phase Entry in the Picophytoplanktonic Cell Ostreococcus | Although the decision to proceed through cell division depends largely on the metabolic status or the size of the cell, the timing of cell division is often set by internal clocks such as the circadian clock. Light is a major cue for circadian clock entrainment, and for photosynthetic organisms it is also the main source of energy supporting cell growth prior to cell division. Little is known about how light signals are integrated in the control of S phase entry. Here, we present an integrated study of light-dependent regulation of cell division in the marine green alga Ostreococcus. During early G1, the main genes of cell division were transcribed independently of the amount of light, and the timing of S phase did not occur prior to 6 hours after dawn. In contrast S phase commitment and the translation of a G1 A-type cyclin were dependent on the amount of light in a cAMP–dependent manner. CyclinA was shown to interact with the Retinoblastoma (Rb) protein during S phase. Down-regulating Rb bypassed the requirement for CyclinA and cAMP without altering the timing of S phase. Overexpression of CyclinA overrode the cAMP–dependent control of S phase entry and led to early cell division. Therefore, the Rb pathway appears to integrate light signals in the control of S phase entry in Ostreococcus, though differential transcriptional and posttranscriptional regulations of a G1 A-type cyclin. Furthermore, commitment to S phase depends on a cAMP pathway, which regulates the synthesis of CyclinA. We discuss the relative involvements of the metabolic and time/clock signals in the photoperiodic control of cell division.
| Microalgae from phytoplankton play an essential role in the biogeochemical cycles through carbon dioxide assimilation in the oceans where they account for more than half of organic carbon production. Photosynthetic cells use light energy for cell growth, but light can also reset the circadian clock, which is involved in the timing of cell division. How light signals are integrated in the control of cell division remains largely unknown in photosynthetic cells. We have used the marine picoeukaryotic alga Ostreococcus to dissect the molecular mechanisms of light-dependent control of cell division. We found that the Retinoblastoma pathway integrates light signals which regulate the synthesis of CyclinA in response to cAMP. Alteration of CyclinA or Rb levels triggers cell division in limiting light conditions and bypasses the need for cAMP. In addition, CyclinA overexpression affects the timing of S phase entry. This first integrated study of light-dependent regulation of cell division in photosynthetic cells provides insight into the underlying molecular mechanisms.
| The cell division cycle (CDC) is a highly conserved and regulated process, which controls the proliferation of unicellular organisms and development and tissue renewal in multicellular organisms. In eukaryotes the main steps of CDC progression are controlled by Cyclin Dependent Kinases (CDKs). From human to algae, the metabolic status regulates cell cycle progression. Cell growth can occur during different phases of CDC depending on the organism but the main decision to progress into the cell cycle is usually made in G1 and depends on environmental conditions. It is referred to as cell cycle commitment and known as START in yeast or restriction point in mammals. Commitment has been depicted as a point, beyond which the cell is irreversibly engaged in cell cycle progression and is no longer sensitive to nutrients and also in the case of photosynthetic organisms, light availability [1], [2]. The transcriptional regulation of cell cycle progression in S phase is controlled in mammals and plants by the E2F transcription factors and these are sequestrated by the Retinoblastoma protein (Rb). In budding yeast, it is controlled by the transcription factor Swi4/6-dependent cell cycle box-Binding Factor (SBF) which is sequestrated by Whi5. On phosphorylation of Rb by G1 Cyclin/CDC complexes, such as CyclinD-Cdk4, E2F transcription factors are released leading to S phase commitment. In yeast on phosphorylation of Whi5, by Cln3/Cdc28, SBF is released leading to S phase commitment. In plants, a CyclinD/CDKA complex has been shown to phosphorylate a Retinoblastoma related (RBR) protein and overexpression of CyclinD accelerates entry into S phase and mitosis of G0 cells [3]. G1 cyclin/CDK complexes are primary targets of environmental signals and cyclin levels can be regulated at the transcriptional or the post-transcriptional level by mitogenic factors such as hormones and nutrients availability [4]–[7]. In animals and yeast, Rb and Whi5 respectively are critical players in linking cell size or metabolic status to cell cycle progression [8], [9].
The gating of CDC, which restricts cell division to well defined windows of time during the day, has been described for organisms as diverse as microalgae [10], [11] and mammals [12]. Gating of CDC ensures that cell division occurs with a daily periodicity over a wide range of environmental conditions. The timing of cell division is relatively insensitive to changes in the environment, such as nutrients or temperature and persists under constant light with a period close to 24 hours, two features of circadian regulation. The Wee1 kinase, a key regulator of G2/M transition is transcriptionally regulated by the master clock complex CLOCK/BMAL1 in mouse regenerating liver cells, illustrating the direct control of cell cycle components by the circadian clock. In addition, striking experimental evidences showed that the circadian clock and the DNA damage pathway share common regulators from animals to fungi [13]–[15]. However, more experimental data is needed to unravel cross-talks between circadian, metabolic and cell cycle controls in the absence of injury or stress.
Unicellular algae such as Chlamydomonas or Euglena are very useful organisms to dissect the light-dependent regulation of cell division in photosynthetic organisms because cell division can be synchronized by light/dark cycles. In Chlamydomonas, commitment takes place in G1 whereas in Euglena the light-dependent control of CDC operates mainly in G2 but also at G1/S and S/G2 transitions [16]. In Chlamydomonas, cell division was shown to be under circadian control [17] but also to depend on the amount of light available for photosynthesis [1]. Until recently tools for gene function analysis were available only for Chlamydomonas, a microalga that exhibits multiple-fission division type. We have recently implemented molecular tools for gene function analysis in the picoeukaryotic alga Ostreococcus tauri, which divides by simple binary fission [18]. O.tauri has a very compact genome and displays very low gene redundancy [19]. A reduced set of cell cycle genes including Cyclins and Cyclin-Dependent Kinases (CDKs) were identified in the fully sequenced genome [20]. They encode functional CDKs and associated regulatory proteins [21]. Cell division and the transcription of the main cell cycle regulators were shown to be under circadian control and resetting by light demonstrated that the timing of cell division is mainly locked to the time of light on [22].
Here, we have performed an integrated study of light-dependent regulation of cell division in Ostreococcus, varying available light by modulating both light duration and intensity. In all conditions, the timing of cell cycle entry did not occur prior to 6 hours after dawn. No cell cycle arrest was observed outside the G1 phase. CDKA, CyclinA and Rb had patterns of expression and interactions compatible with a putative involvement in a functional Rb pathway. Cyclic AMP was necessary and sufficient for both S phase entry and CyclinA synthesis. Down-regulation of Rb or CyclinA overexpression triggered cell cycle entry under limiting light conditions demonstrating the antagonistic roles of cAMP and Rb in a “metabolic checkpoint”. Moreover, overexpression of CyclinA advanced the timing of S phase entry. Our work illustrates how combined light intensity-dependent and time-dependent signals regulate S phase entry and give insight into the role of a G1 cyclin in the light-dependent control of cell cycle progression.
Cells entrained under 12 hours light, 12 hours dark cycles (LD 12, 12) at 35 µmol.quanta. m−2.s−1 were in G1 phase at dawn (Time 0) (Figure 1). They were submitted to various light intensities and durations from Time 0 to modulate the amount of light provided (Figure 1A). Under these conditions, light is a source of energy for photosynthesis, that is required for cell growth prior to cell division (commitment) but it can also act as a signal (timer or clock) controlling the timing of cell cycle events. Estimation of DNA content by flow cytometry allowed monitoring of S phase as previously described [21], [22]. Cells in S phase were detected between 6 hours and 14 hours after light on (Time 6 and Time 14). Depending on the light intensity and duration, the cell population underwent from 0 to more than 1 division as determined by cell counting (Figure S1). G2 and M phases are very short in Ostreococcus as estimated from naturally or artificially synchronized cell populations [21]. Therefore the number of cells in G2/M is low and difficult to estimate [21]. Furthermore for low light intensities/durations, only a few cells divided (Figure 1) whereas for high light intensities/durations, two successive divisions could be observed (Figure S1) making it extremely difficult to estimate the rate of cells in S and G2/M phase and to discriminate between the first and second S phases. To determine the effect of light on cell division we chose to focus on the timing of entry into the first S phase (Figure 1). At the control fluence rate (35 µmol.quanta.m−2.s−1), S phase was detected from 6 hours after Light on (Time 6), that is, at the same time as under the entraining LD 12, 12 cycle. When increasing light intensity (from 35 to 100 or 150 µmol.quanta.m−2.s−1), only 3 to 4 hours of light were required for commitment to S phase. Exposure to light for 8 hours allowed S phase progression at all tested fluence rates with a maximum of cells entering S phase for highest intensities. In all conditions, cells entering S phase, completed their cell cycle and after 24 hours the cell population was back in G1. This suggests that the main light-dependent control of cell cycle progression occurs in G1 and that cell cycle progression is not impaired by darkness once cells are committed. Commitment to S phase was dependent both on light intensity and duration (blue area in Figure 1B). For example, at 150 µmol.quanta.m−2.s−1 the first committed cells were seen 2 to 3 hours before S phase was detected, whereas at 35 µmol.quanta.m−2.s−1 the first cells in S phase were detected at the same time as the first committed cells (Time 6) (Figure 1B). For the lowest light intensity, the timing of S phase was delayed (red area in Figure 1B), most likely because cells had not received enough light to commit at that time (intersection of the blue and red area on Figure 1B). Together these results indicate that the timing of entry into the first S phase is gated during several hours after dawn. For the highest light intensities some cells were able to divide twice in a row (Figure S1), suggesting that the timing mechanism which gates cell division until Time 5 to Time 6, has little effect on the timing of the second division. This is similar to the gating of division described in Chlamydomonas, which is restricted to a time window, the number of successive divisions being determined by the light conditions [1].
Limiting and non-limiting light conditions for cell division (referred to as limiting and non-limiting conditions) were chosen as three and eight hours respectively, of exposure to light at 100 µmol.quanta.m−2.s−1. At this level about 90% of an LD 12, 12 entrained cell population divided (see Figure S1). We investigated the transcription patterns of the main cell cycle actors of cell division in limiting and non-limiting conditions (Figure 2). Transcription of Cyclin-Dependent Kinases (CDKs), Cyclins and Retinoblastoma (Rb) were monitored by quantitative RT-PCR. CyclinB was the only transcript that was not detected in limiting conditions, whereas expression patterns of other cell cycle genes including CyclinA, CyclinD, CDKA, CDKB and Rb remained similar in both limiting and non-limiting conditions. CDKA and CyclinA transcripts were detected first, accumulating as early as two hours after light on, closely followed by Rb, CyclinD and CDKB mRNAs. Maximal transcripts levels were observed between 9 to 10 hours after light on when most of the cells were progressing through the cell cycle.
In the non-limiting light condition the transcription of CyclinB started after 6 hours, when S phase had begun, suggesting that its transcription might be dependent on cell cycle progression in G1. In contrast known and putative G1/S regulators, including CDKA, CyclinA and Rb, were not differentially expressed in limiting and non-limiting conditions, indicating that their transcriptional regulation is independent of commitment.
Together Figure 1 and Figure 2 suggested that commitment occurs upon light assimilation in G1 and that it does not primarily rely on transcriptional regulations of the putative G1/S regulators identified in silico. Because, the Retinoblastoma protein (Rb) is well known to play a central role in the restriction point of plant and animal cells, we chose to investigate the role of Rb in G1 progression in Ostreococcus. CyclinA is the only protein exhibiting a canonical (LXCXE) Rb-binding site [23] and CDKA is the only CDK expressed in G1. Thus, CDKA/CyclinA complex is the best candidate for regulating cell cycle progression in G1. To monitor Rb, CDKA, and CyclinA protein synthesis and quantify interacting partners, we generated stable translational luciferase reporter lines Rb-Luc, CDKA-Luc and CyclinA-Luc in the pOtLuc vector [18]. Estimation of the recombinant protein synthesis was achieved through luminescence measurement from either whole protein extracts or affinity-purified proteins. The human p9CKShs1 referred to as P9 was used to specifically purify CDKA [21]. An anti-CyclinA antibody was used for immunoprecipitation of CyclinA and associated proteins. Luminescence patterns measured in extracts from CDKA-Luc and CyclinA-Luc lines were similar to that of CDKA and CyclinA profiles as determined by western blot, demonstrating that in our experiments luciferase translational fusions (Figure 3A) reflected the expression patterns of these proteins (Figure 3B), which in the case of CyclinA resulted mainly from protein de novo synthesis since endogenous CyclinA was no detected at Time 0 (Figure 3A and 3B).
In non-limiting conditions, CyclinA-Luc accumulated from 4 hours after light on (Figure 4A). Similar profiles of CyclinA-Luc were obtained from raw extracts or P9-purified complexes (CDKA/CyclinA-Luc) (Figure 4A). Significantly, CyclinA-Luc protein was found to be bound to CDKA from Time 4 that is, as soon as CyclinA-Luc was detected in raw extracts. Conversely, CyclinA/CDKA-Luc complexes were immuno-precipitated with the anti-CyclinA antibody. While a steady state level of CDKA-Luc was detected in raw extract, the amount of CDKA-Luc copurified with CyclinA followed the profile of CyclinA-Luc in raw extract (Figure 4B). These results suggest that CyclinA may be a limiting factor in the formation of the CyclinA/CDKA complex, before commitment.
Rb-Luc level increased from 6 hours after light on, reaching a maximum at 7 to 8 hours after light on (Figure 4C). In contrast, Rb-Luc bound to CyclinA in immunoprecipitation experiments peaked 3 hours before Rb-Luc in raw extract (Figure 4C). Rb-Luc purified on P9 (bound to CDKA) had a similar profile as Rb-Luc bound to CyclinA (Figure 4D and 4C). Since CDKA-Luc was detected at a steady state level in raw extracts (Figure 4B), this suggests that CyclinA might be a limiting factor in CDKA/Rb interaction early after dawn. Significantly, the amount of Rb-Luc associated to CyclinA or CDKA was highest around Time 5 to Time 6, that is, 2 hours before CDKA/CyclinA maximal interaction. This suggests that at the light/dark transition (Time 8), Rb is released from the CyclinA/CDKA complex.
CyclinA transcript and CyclinA-Luc were monitored in limiting and non-limiting conditions (Figure 5). In non-limiting conditions CyclinA-Luc was detected from 4 hours after light on, i.e. two hours after CyclinA transcript (Figure 5A). No CyclinA-Luc could be detected in limiting conditions though CyclinA mRNA profile remained similar to that in non-limiting conditions (Figure 5B). When the light supply was modulated by changing the fluence rate instead of the duration of illumination (12 hours of light), all profiles of CyclinA mRNA were increasing from 1 hour after light on (Figure 5C). In contrast, translation products, as reported by CyclinA-Luc, appeared later for lower fluence rates (Figure 5D). These results indicate that the synthesis of CyclinA protein, unlike the CyclinA transcription, is regulated by the light conditions in a manner very similar to S phase commitment (e.g. Figure 1).
To gain insight into the signal transduction pathway leading to the light-dependent regulation of S phase entry, we chose to investigate the involvement of cAMP known to be an important signaling component for cell cycle progression [24]–[26]. Monitoring of cAMP level in cells exposed to various light intensities after LD 12, 12 entrainment revealed that the peak of cAMP occurred earlier and/or had higher amplitudes for high fluence rates (Figure S2). This suggests a possible correlation between cAMP level, CyclinA synthesis and S phase commitment since cells were committed sooner at high fluence rates (e.g. Figure 1).
Under non-limiting conditions, cAMP increased immediately after light on and returned to a basal level before S phase was detected (Figure 6A and 6B). We used a pharmacological approach to evaluate the role of cAMP in the synthesis of CyclinA and the control of S phase. Indomethacin and forskolin have been reported to inhibit and activate cAMP synthesis, respectively, including in photosynthetic organisms [27], [28]. Inhibiting cAMP synthesis with indomethacin (30 µM) impaired cell cycle entry and CyclinA-Luc protein synthesis without affecting CyclinA transcription (Figure 6A–6D). Conversely, the adenylate cyclase activator forskolin (20 µM) significantly increased cAMP levels under limiting light conditions (Figure 6E). Remarkably, S phase cells as well as CyclinA-Luc protein synthesis were detected, though at a low levels, in the presence of forskolin (Figure 6F and 6G). CyclinA-Luc accumulated as early as one hour after forskolin addition, i.e. two hours after light on but the first cells in were not detected before 6 hours after light on as in control cells. Our results indicate that cAMP is required for CyclinA synthesis and S phase entry. The delay between the peak of cAMP and commitment suggests that cAMP is an upstream signal in a signal transduction pathway leading ultimately to commitment rather than a direct regulator of commitment.
To better understand the respective involvement of CyclinA protein and cAMP in commitment, CyclinA was ectopically expressed under the strong and constitutive Ostreococcus High Affinity Phosphate Transporter (HAPT) promoter in the pOtoxLuc vector (Figure S3). Two lines, referred to as CyclinA-ox, were selected based on CyclinA expression levels. In limiting conditions CyclinA was detectable, though at low levels, as early as 1 hour after light on in CyclinA-ox lines (Figure 7A). In CyclinA-ox line S phase was detected under limiting light conditions as early as 3 hours after light on (Figure 7B). Under non-limiting conditions S phase entry was advanced by two hours in CyclinA-ox lines compared to WT cells (Figure 7C). Moreover, unlike control cells, CyclinA–ox cells treated with indomethacin were able to enter S phase (Figure 7D). Under non-limiting conditions, CyclinA displayed higher levels than control cells at all time points (Figure 7E). Noteworthy, the levels of CyclinA increased after dawn in CyclinA-ox lines as in control cells, suggesting that post- translational regulations operate also in overexpression lines. Higher levels of CyclinA were also detected in indomethacin-treated cells (Figure 7E).
Overexpression of CyclinA, therefore, bypasses the requirement for light and cAMP, allowing cells to commit and to enter into S phase with an earlier timing.
To test the role of Rb in the light-dependent regulation of cell division, we generated Retinoblastoma- knockdown (Rb-kd) by expressing antisense Rb sequence in the pOtoxLuc vector. Three lines, with reduced levels of Rb mRNA compared to WT cells were selected (Figure S4). Under limiting conditions, Rb-kd cells were still able to progress into S phase (Figure 8A). Inhibition of cAMP synthesis by indomethacin reduced the number of Rb-kd that entered S phase but a significant proportion progressed through S phase, while wild type cells remained in G1. Thus, down-regulation of Rb bypasses the need for light and cAMP in Rb-kd as in CyclinA-ox lines. Entry into S phase occurred at the same time in Rb-kd cells as in WT under non-limiting light conditions (Figure 8C). Therefore repression of Rb expression triggers commitment but has no obvious effect on the timing of S phase, unlike overexpression of CyclinA, which induces an earlier timing of cell division. Significantly, in limiting light conditions CyclinA was not detected in Rb-kd cells (Figure 8C) as in control cells (e.g. Figure 7A). As cells in S phase were observed in Rb-kd cells, this suggests that CyclinA may not be essential for S phase entry when Rb is repressed in limiting light conditions.
We have previously shown that in Ostreococcus, the CDC is synchronized by day/night cycles and that a timing mechanism, namely the circadian clock, regulates cell division and the transcription of the main cell cycle regulators [22]. The present study aims to decipher the molecular mechanisms involved in the photoperiodic control of cell division. Varying the fluence rate and/or the duration of exposure to light modulates the length of the G1 phase. Once committed, cells resume division indicating that the light-dependent regulation of cell division occurs mainly in G1 (Figure 1). G1 phase is lengthened by as much as 3 hours (from 6 to 9 hours) for the lowest light intensity, suggesting that metabolic status might control commitment. Conversely, for high fluence rates, 3 to 4 hours are sufficient for commitment but S phase is not observed prior to 6 hours after dawn indicating that engagement of committed cells into S phase is gated in time. Such a light-dependent control of cell cycle progression has been described in Chlamydomonas. In early G1, cell cycle progression is light-dependent but after commitment it becomes light-independent and cells wait for an additional 5 to 10 hours before entering S phase [1]. In contrast, in Euglena, the light-dependent regulation of cell division occurs in both G1 and G2 phases because on transfer to darkness cell cycle arrested at both stages [16].
The main cell cycle regulatory genes display similar patterns of transcription under limiting and non-limiting light conditions (Figure 2). From completion of mitosis (about 2 to 3 hours after light off) to the early morning, no transcription of the main cell cycle regulators is detected [22]. CyclinA is the first gene to be transcribed from one hour after light on closely followed by CDKA, CyclinD, CDKB and Rb. (Figure 9). Transcription of these genes occurs at fixed time intervals from light on independently of the commitment status of the cells, suggesting that transcription is mainly controlled by a dawn-dependent timing mechanism, which would be similar to the circadian control of cell division [22]. The main regulatory genes of cell division, including cyclins and CDKs have been shown to retain rhythmic expression under constant light [22] suggesting a circadian regulation of their transcription. We show here that CyclinB transcription is further dependent on commitment (Figure 2). The presence of an E2F-binding motif in the promoter of CyclinB, suggests that CyclinB transcription may depend on E2F transcription factor, once the cells have passed commitment. Besides being the first cyclin to be expressed soon after dawn, CyclinA is the only cyclin, which displays a fully conserved Rb-interaction motif in O. tauri. Furthermore CyclinA interacts with both CDKA and Rb protein in G1 (Figure 4). Since CDKA is the only canonical CDK present during G1 phase [21], CDKA/CyclinA, is potentially the only CDK/cyclin complex in the Rb pathway in Ostreococcus. As soon as CyclinA protein is detected, it is found in association with both CDKA and Rb protein. In contrast, CDKA does not associate with Rb in the absence of CyclinA suggesting that CyclinA may be essential for the formation of this complex. Finally CyclinA remains complexed with CDKA several hours after the maximal interaction between Rb/CyclinA and Rb/CDKA, consistent with the CyclinA/CDKA complex being involved in the control of S phase entry. Surprisingly, the level of Rb remains high during S phase, which may appear to be in contradiction with an exclusive role for Rb in S phase progression. This suggests that Rb might also be involved later during cell cycle progression as previously reported in several organisms [29]–[31].
CyclinA overexpression or Rb down-regulation induce cell division in limiting light conditions indicating that the Retinoblastoma pathway plays an essential role in cell cycle progression at commitment (Figure 9B and 9C). Such a commitment phenotype has been observed in animal knockout cells lacking the entire Retinoblastoma family, which cannot undergo growth arrest when starved of growth factors [32] and Chlamydomonas mat3 mutants cells were shown to divide at smaller size than wild type cells [33]. The best known function of Rb is to repress S phase transcription by sequestering the E2F transcription factor until Rb is phosphorylated by specific Cyclin/CDK complexes [34]. Our results would be consistent with CyclinA/CDKA controlling cell cycle progression in G1 by regulating Rb phosphorylation at commitment, Rb being a negative regulator of CyclinA/CDKA (Figure 9A). No CyclinA was detected in Rb-kd cells entering early into S phase under limiting light conditions. This would suggest that other cell cycle regulators downstream of CyclinA that are not under the control of Rb are sufficient to promote S phase entry in the absence of CyclinA (Figure 9B). Therefore a main function of CyclinA/CDKA may be to override the inhibitory effect of Rb in wild type cells (Figure 9A).
In non-limiting light conditions, the level of cAMP increases from light on and peaks before S phase entry (Figure S2), while cAMP level remains low under limiting light conditions (Figure 6). Mitogens such as the EGF are known to induce cell cycle re-entry in a cAMP dependent manner [35]. In S. cerevisiae, G1 progression is regulated by cAMP, which mediates the intracellular level of glucose [26]. In Ostrecoccus the inhibition of cAMP synthesis with indomethacin prevents cAMP accumulation and cell division under non-limiting conditions but has no effect on cell growth (Figure 6). Conversely forskolin, an activator of cAMP synthesis, triggers cell division, though at low rates, under limiting light conditions. The fact that cAMP is necessary for commitment suggests that our limiting light condition may correspond to a metabolic limitation by restricting the light energy available for photosynthesis (Figure 9A).
The transcription of CyclinA does not depend on light conditions but the synthesis of CyclinA protein occurs only under non-limiting light conditions. When CyclinA protein is detected earlier under high light, the cells commit sooner. While the rise in transcription of CyclinA is independent of the light conditions occurring at a fixed time after light on, the post-transcriptional regulation of CyclinA may ensure that CyclinA does not accumulate until optimal metabolic conditions allowing cell growth for commitment are met (Figure 9A). Inhibition of cAMP synthesis prevents CyclinA synthesis under non-limiting light conditions and activation of cAMP synthesis triggers CyclinA synthesis consistent with CyclinA accumulation being regulated by cAMP levels. In agreement with this hypothesis, down-regulation of Rb or overexpression of CyclinA bypasses the need for cAMP for cell division (Figure 9B and 9C). A similar mechanism of G1 progression by the metabolic status has been described in yeast. The differential translation rate of the G1 cyclin cln3 is dependent on a Ras-cAMP pathway, which reflects the metabolic status of the cell [26]. In a rich carbon source, cln3 appears earlier, interacts with cdc28 to phosphorylate Whi5 and induces an earlier progression through START, leading to a shortening of G1 phase. Our results suggest that in Ostreococcus cells the synthesis of CyclinA is regulated by a cAMP dependent mechanism when cells have accumulated enough light energy (Figure 9A). However, the time lapse between the peak of cAMP and S phase observed in most of the light conditions suggests that cAMP does not directly control CyclinA synthesis but that it activates a downstream pathway that controls CyclinA accumulation. An alternative explanation is that cAMP regulates cell cycle progression independently of commitment, consistent with the fact that the temporal changes in cAMP level do not correlate exactly with the timing of the commitment.
The timing of cell division in Ostreococcus has been shown previously to be regulated mainly by the dark-light transition and to occur in G1 because perturbations by light or dark pulses induced changes in timing of S phase entry, but not in the duration of the S, G2 and M phases [22]. In Figure 1, we show that the timing of S phase is delayed when entrained cells are exposed to low light from dawn. In all conditions S phase is not observed before 6 hours after light on, even for high fluence rates suggesting that the timing of S phase is gated (Figure 1). Previous light-resetting experiments have shown that the timing of cell division is mainly locked to light on at dawn defining a time window in which cells do not divide [22]. Timing mechanisms of cell division have been shown to rely on clocks such as the circadian clock in animal cells, which regulates cell cycle progression during both G1 and G2 phase. [36]–[38]. In mice liver cells re-entering the cell cycle upon liver partial ablation, the circadian clock gates entry into mitosis by regulating the transcription of the Wee1 kinase, which inhibits the activity of Cyclin B1/CDC2 kinase in G2 [36]. In Ostreococcus cells, rhythmic patterns of cell division and transcription of the main cell cycle regulators persist under constant light, supporting a circadian regulation of cell division [22]. CyclinA transcription is not affected under a wide range of light conditions further suggesting that it is regulated by a timing mechanism rather than by metabolic control. This would be consistent with a clock controlling the transcription of the main cell cycle actors after dawn. Alternatively we cannot rule out the possibility that the light on signal, on its own, is sufficient to trigger the transcription of the main cell cycle regulators in G1.
Remarkably, overexpression of CyclinA induces earlier entry into S phase in limiting light conditions, suggesting at first sight that the regulation of CDKA/CyclinA activity by a clock may account for the timing of S phase (Figure 7). However Rb-kd cells enter S phase without any detectable CyclinA under limiting light conditions and in these cells the timing of S phase is normal (Figure 8 and Figure 9B). As mentioned above it is possible that CyclinA/CDKA promotes S phase entry by counteracting the inhibitory effect of Rb (Figure 9A). In the absence of CyclinA and Rb, S phase would be controlled by another timing mechanism. An as yet unknown player, such as a Cyclin/CDK complex would control S phase entry independently of CyclinA/CDKA and it would be negatively regulated by Rb in normal conditions. Our results also suggest that this player would be controlled by an independent timing mechanism since the timing of S phase is normal in Rb-kd lines which display low levels of Rb transcript and no detectable CyclinA protein under limiting light conditions (Figure 9B). The earlier timing of S phase entry observed in CyclinA-ox lines under limiting light conditions could also be explained by a titration effect of an inhibitor such as a CDK inhibitor by CyclinA (Figure 9C). It is also possible that CyclinA overexpression non-specifically induces cell cycle progression by replacing another Cyclin/CDK complex since cyclins can have overlapping functions. Alternatively, the overexpression of CyclinA may induce cell cycle progression downstream of the G1/S progression if CyclinA is also involved later during cell cycle progression as previously hypothesized [21].
Finally, activation of cAMP synthesis by forskolin induces an early synthesis of CyclinA though at low levels but the timing of S phase is not advanced. It is therefore possible that CyclinA promotes cell cycle progression in a dose-dependent manner and that the forskolin-treatment prevents CyclinA from reaching sufficient levels to promote early S phase entry.
In summary, we propose a model of the light-dependent regulation of G1 phase progression, in which timing and metabolic signals are integrated in a sequential way (Figure 9A). First, the transcription of several genes needed for G1 phase progression is activated, independently of the amount of light provided. Among these genes, CyclinA is one of the earliest to be transcribed after dawn, likely depending on a timing mechanism such as the circadian clock. CyclinA synthesis appears to be controlled by a cAMP-dependent pathway, most probably under metabolic control. CyclinA protein binds both CDKA and Rb, which might result eventually in the release of E2F transcription factor upon phosphorylation of Rb. Finally another timing mechanism, independent of commitment prevents entry into S phase before 6 hours after dawn. Whether this timer is the circadian clock and which cell cycle regulators are involved is currently unknown. The limited number of cell cycle regulators in Ostreococcus as well as the recent identification of circadian clock players [18] should allow this question to be addressed in the future.
More generally, unraveling the molecular mechanisms of light-dependent regulation of growth and cell division in microalgae from the phytoplankton should lead to a better understanding of the physiology of these key organisms involved in carbon dioxide assimilation.
Ostreococcus tauri strain, 0TTH0595 isolated from the Thau lagoon [39], was cultivated in filtered sterile seawater supplemented with Keller enrichment medium (Sigma-Aldrich, Lyon, France). O. tauri strain was grown in aerated flasks (Sarstedt) at 20°C under 12 hours light/12 hours dark cycles as previously described [21]. Drugs were purchased from Sigma-Aldrich unless otherwise stated. For extraction, cells were harvested by centrifugation in conical bottles (10,000 g, 10 min, 4°C), after addition of pluronic (0.1%) to the medium and stored at −80°C until extraction. Frozen cell-pellets were ground by shaking (45 seconds, 30 Hz, twice) with 5 mm stainless steel beads using a TissueLyser (Retsch, Haan, Germany) after the appropriate buffer was added. Cell debris were removed by centrifugation (12 000 g, 10 min, 4°C).
A 1 ml cell sample was fixed with 0.25% glutaraldehyde (Sigma) for 15 min at room temperature and then stored at 4°C for 1 day or frozen in liquid nitrogen and stored at −80°C. Flow cytometry analysis was performed on a FACScan flow cytometer (FACScalibur; Becton-Dickinson, San Jose, CA). Cells were counted from the appropriate gate (FL3-H versus SSC-H) as described previously [39]. For analysis of the DNA content, whole fixed cells were stained with SYBR green I (3000X dilution of the commercial solution; Molecular Probes, Eugene, OR) for 30 min, and 20,000 cells per sample were analyzed using the CellQuest software. Cell cycle analysis was performed with the Modfit software (Verity Software House, Tophsam, ME) as previously described [21]. The Graphical trapezoidal model and the fixed ratio of G2/G1 of 1.85 gave the best fits and were kept for cell cycle analysis in all analysis.
Amplifications by PCR of CDKA, CyclinA, Retinoblastoma, full genes, including the promoter and coding region were achieved with the Triple Master polymerase mix (Eppendorf). A sub-cloning step in the pGEMT vector (Promega) was performed first. The pOtLuc vector was used to fuse the gene in frame with luciferase enabling protein quantification via in vitro luciferase assay [18]. The pOtoxLuc vector was designed to facilitate the selection of overexpression/antisense transformants on the basis of luminescence levels produced from luciferase fused to the CCA1 promoter (Figure S3). POtoxLuc allows the expression of the sequences of interest in sense or antisense orientation under control of the strong High Affinity Phosphate Transporter promoter (pHAPT). CyclinA coding sequence and antisense of 3′-end sequence of Retinoblastoma coding sequence (from position 2964 to 1824) were cloned in pOtoxLuc. Overexpression of CyclinA was confirmed by western blot, and Knock-down of Retinoblastoma by quantitative RT-PCR.
Transformation was performed as previously described [18]. Briefly, O. tauri was harvested by centrifugation (8000 g, 8 min, 10°C) after pluronic addition (0.1% final concentration) and cells were gently resuspended in 1 ml 1 M sorbitol. After one supplemental wash, cells were resuspended in 50 to 80 µL sorbitol (2 to 3×1010 cells per ml) and incubated with 5 µL of linearised DNA (1 µg/µL) before electroporation using a Bio-rad Gene Pulser apparatus (field strength 6 kV/cm, resistor 600 Ω, capacitor 25 µF). Cells were transferred into culture medium for 24 h. Stable transformant colonies were selected in semi-solid medium at 0.2% w/v agarose (low melting point agarose, Invitrogen) in Keller Medium supplemented with G418 (Calbiochem) at 1 mg/ml concentration. Individual clones were transferred in liquid medium in 96-well microplate until they reached stationary phase (4 to 6 107 cell/ml). Luciferase reporter lines were selected on the basis of reproducible patterns of luminescence under LD conditions. Overexpressing/Antisense lines were first selected from the lines displaying the highest luminescence level and then analyzed either by quantitative RT-PCR or by western blot.
RNA was extracted using RNeasy-Plus Mini kit (Qiagen, Hilden, Germany) following the manufacturer's instructions. Contaminating DNA was removed using Q1 RNAse-free DNAse (Promega). Absence of DNA contamination was checked by PCR. Reverse transcription was performed using the PowerScript Reverse Transcriptase synthesis kit (BD Bioscience, Palo Alto, CA). Real-time PCR was carried out on a LightCycler 1.5 (Roche Diagnostic) with LightCycler DNA Master SYBR Green I (Roche Molecular Biochemicals). Primers were designed with LightCycler Probe Design2 software (Roche Diagnostic, Mannhein, Germany). Primers are available in Table S1. Results were analyzed using the comparative critical threshold (ΔΔCT) method. The O. tauri elongation factor 1α (EF1α was used as internal reference. The analyses were performed in duplicate. Errors (SD) were usually below 1%.
Proteins were extracted in CCLR buffer (100 mM potassium phosphate pH 7.8, 1 mM EDTA, 1 mM DTT, 1% TritonX-100, 10% glycerol). For affinity purification, protein extracts were diluted 5 time in CCLR with antiprotease without glycerol and further incubated with either p9CKShs1 sepharose beads (Corellou, 2005) or specific anti-Ostreococcus CyclinA bound to protein A sepharose on a rotator at 4°C for one hour as previously described [40]. After western blotting, protein detection was achieved by enhanced chemiluminescence detection.
Luminescence of translational luciferase fusion proteins (50 µl of protein extract) was recorded on Centro LB 960 luminometer (Berthold Technologies, Germany), 1 minute after injection of 80 µl of luciferase assay reagent buffer (20 mM Tricine pH 7.8, 5 mM MgCl2, 0.1 mM EDTA, 3.3 mM DTT, 270 µM coenzyme A, 500 µM luciferin, 500 µM ATP). The luciferase background was determined by running controls lacking the primary antibody (e.g anti-CyclinA) and substracted for each time point. We also checked using various amount of recombinant luciferase expressed under control of the High affinity phosphate promoter that luciferase is not immunoprecipitated by antibodies or bound to P9. In these control experiments, the luciferase activity was below 0.1% of the initial luciferase activity in the extract before immunoprecipitation.
For each sample, 1 ml of cell culture (corresponding to five millions of cells), were extracted in HHBS buffer. cAMP measurements were performed with cAMP I HitHunter assay kit for cells in suspension (DiscoveRx Corp., CA) according to the manufacturer's instructions. Luminescence was recorded in 96 wells microplate using a Centro LB 960 luminometer (Berthold Technologies, Germany). A cAMP standard curve was established to quantify the cAMP levels which were normalized to the cell number as determined by flow cytometry.
|
10.1371/journal.ppat.1004895 | Edin Expression in the Fat Body Is Required in the Defense Against Parasitic Wasps in Drosophila melanogaster | The cellular immune response against parasitoid wasps in Drosophila involves the activation, mobilization, proliferation and differentiation of different blood cell types. Here, we have assessed the role of Edin (elevated during infection) in the immune response against the parasitoid wasp Leptopilina boulardi in Drosophila melanogaster larvae. The expression of edin was induced within hours after a wasp infection in larval fat bodies. Using tissue-specific RNAi, we show that Edin is an important determinant of the encapsulation response. Although edin expression in the fat body was required for the larvae to mount a normal encapsulation response, it was dispensable in hemocytes. Edin expression in the fat body was not required for lamellocyte differentiation, but it was needed for the increase in plasmatocyte numbers and for the release of sessile hemocytes into the hemolymph. We conclude that edin expression in the fat body affects the outcome of a wasp infection by regulating the increase of plasmatocyte numbers and the mobilization of sessile hemocytes in Drosophila larvae.
| The events leading to a successful encapsulation of parasitoid wasp eggs in the larvae of the fruit fly Drosophila melanogaster are insufficiently understood. The formation of a capsule seals off the wasp egg, and this process is often functionally compared to the formation of granulomas in vertebrates. Like granuloma formation in humans, the encapsulation process in fruit flies requires the activation, mobilization, proliferation and differentiation of different blood cell types. Here, we have studied the role of Edin (elevated during infection) in the immune defense against the parasitoid wasp Leptopilina boulardi in Drosophila larvae. We demonstrate that edin expression in the fat body (an immune-responsive organ in Drosophila functionally resembling the mammalian liver) is required for a normal defense against wasp eggs. Edin is required for the release of blood cells from larval tissues and for the subsequent increase in circulating blood cell numbers. Our results provide new knowledge of how the encapsulation process is regulated in Drosophila, and how blood cells are activated upon wasp parasitism. Understanding of the encapsulation process in invertebrates may eventually lead to a better knowledge of the pathophysiology of granuloma formation in human diseases, such as tuberculosis.
| Parasitoid wasps are natural enemies of insects such as the fruit fly Drosophila melanogaster. In the course of a successful wasp infection, a female wasp lays an egg in a fruit fly larva and the wasp larva hatches. Thereafter, the wasp larva develops inside the Drosophila larva using the host tissue as a source of nutrition to ultimately emerge as an adult wasp, unless the wasp larva is eliminated by the host’s immune response [1].
The initial oviposition of a wasp egg triggers changes in gene expression in the fruit fly and activates both humoral and cellular defense mechanisms [2–4]. The role of the humoral defense, i.e. the production of antimicrobial peptides by the fat body, via the Imd and Toll pathways in response to a microbial challenge, is well characterized in response to microbial challenge (reviewed in [5, 6]). However, in the context of wasp parasitism, cellular immunity is more striking than the humoral response. The cellular immune responses are mediated by three types of blood cells, or hemocytes: plasmatocytes, lamellocytes and crystal cells (reviewed for example in [7, 8]). The round and small plasmatocytes are the most abundant type tallying up to 95% of all of the larval hemocytes. Plasmatocytes are responsible for phagocytosing invading microorganisms and apoptotic particles and are also required for a normal resistance against bacteria [9–12]. Crystal cells comprise around 5% of all hemocytes and they contain phenoloxidase-containing crystals that are released in the melanization response [13]. Lamellocytes, on the other hand, are solely found in larvae and are rarely present in individuals that are not immune-challenged. The main task of lamellocytes is to participate in encapsulating objects that are too large to be phagocytosed, such as the eggs of parasitoids wasps. However, the encapsulation of wasp eggs requires the concerted action of all three types of hemocytes [7].
Upon a wasp infection, the presence of a wasp egg is first recognized. Plasmatocytes are the first cells that adhere to the wasp egg and they spread around the surface of the egg forming the first layer of the capsule [14]. A wasp infection also leads to the differentiation of a large number of lamellocytes [15–17], which migrate towards the wasp egg and attach onto the plasmatocyte-covered egg. During a successful immune response lamellocytes, together with plasmatocytes, form a multilayered capsule that surrounds the wasp egg. The capsule is melanized, phenol oxidases and reactive oxygen species are released within the capsule [18], and the wasp is ultimately killed.
Although many pathways, such as the Toll and JAK/STAT pathway, have been shown to have a role in the encapsulation response [3], the phenomenon is still insufficiently understood. In this current study, we investigate the role of Edin (elevated during infection) in a wasp infection. Edin is a small peptide that is secreted into the hemolymph upon infection [19, 20], and it is required for the immune response against Listeria monocytogenes [21]. Earlier, we have shown that the expression of edin is induced after a bacterial infection, and it has a minor role in the resistance against Enterococcus faecalis [20]. In this study, we investigated whether edin expression is induced by a wasp infection using the Leptopilina boulardi strain G486. We also examined the role of Edin in the encapsulation response and in the activation and formation of hemocytes upon a wasp infection. We report that edin expression is required in the fat body upon a wasp infection in order to mount an effective encapsulation response, and that knocking down edin in the fat body causes defects in hemocyte mobilization in Drosophila larvae.
We have previously shown that edin is induced both in vitro and in vivo upon a microbial infection, but were unable to find any essential role for Edin in this context [20]. To test whether a wasp infection induces the expression of edin, we infected Canton S larvae with the parasitoid wasp Leptopilina boulardi strain G486, and determined the expression levels of edin in whole larvae three hours after infection using qRT-PCR. As is seen in Fig 1A, the wasp infection led to a 7-fold induction in the expression levels of edin compared to uninfected larvae. Because the fat body is the main immune-responsive organ in the fruit fly, we next looked at edin mRNA levels in the fat bodies of wasp-infected larvae 24 hours post-infection. As is shown in Fig 1B, the expression of edin was more highly induced in the fat bodies of the wasp-infected larvae than in whole larvae (80-fold induction). Our results indicate that edin is upregulated after a wasp infection in larvae and that the fat body is a main source for its expression.
Fruit fly larvae can mount an effective immune response against invading parasitoids by encapsulating the wasp egg. To address the functional significance of edin expression for the encapsulation process upon an L. boulardi infection, we used the UAS-GAL4 system to knock down edin expression. The normal response against the wasp egg is the formation of a visible melanized capsule around the parasitoid egg, and in our hands, 45–66% of control larvae had a melanized capsule. First, we crossed edin14289 RNAi flies (#14289, hereafter referred to as edin14289) with flies carrying the C564-GAL4 driver, which is expressed in many organs, including the fat body, salivary glands and lymph glands [22], and looked for the presence of melanized capsules 27–29 hours after the wasp parasitization (Fig 2A). Parasitized w1118 controls showed an encapsulation rate of 47%. Similarly, w1118 crossed with C564-GAL4 or edin14289 showed encapsulation rates of 52% and 53%, respectively, while only 15% of edin14289 crossed with C564-GAL4 showed melanized capsules. To ensure that the observed phenotype was caused by reduced edin expression, we analyzed the encapsulation response of another edin RNAi line (#109528, hereafter referred to as edin109528). Similarly to the edin14289 line, edin109528 crossed with the driver line showed a clearly decreased encapsulation efficiency of 7% (Fig 2A), when compared to edin109528 crossed with w1118.
We next used a fat body-specific driver to examine specifically whether the lowered encapsulation response was due to the role of edin in the fat body. We crossed both the edin14289 and edin109528 RNAi lines with the Fb-GAL4 driver line and examined the encapsulation response of the offspring. Fb-GAL4 crossed with w1118 showed encapsulation levels of 45% (Fig 2A), whereas edin RNAi flies crossed with Fb-GAL4 showed an encapsulation activity of only 8% (edin14289) and 7% (edin109528). In addition, similar results were also obtained with another fat body-specific driver, Lsp2-GAL4 (edin109528, S1 Fig).
We also analyzed the encapsulation activity of edin RNAi larvae crossed with the pan-hemocyte driver HmlΔ;He-GAL4 and were not able to see any effect with either of the RNAi lines (60% and 70% encapsulation, Fig 2A). Together, these data suggest that Edin is required for a normal encapsulation response after parasitization, and that its expression is required in the larval fat body but not in the hemocytes.
Scoring for the ability of the fly larva to melanize the wasp egg does not indicate whether the fruit fly larva is actually able to overcome the parasitization. Therefore, we replicated the experimental setting in Fig 2A, but scored for the presence of living or dead wasp larvae 48–50 hours post infection. The parasite was scored as killed by the fruit fly larva if a melanized wasp egg was found in the hemocoel in the absence of a living wasp larva. As is seen in Fig 2B, the percentage of dead wasps in control larvae varied between 20–34%. When edin14289 RNAi was induced with either the C564-GAL4 or Fb-Gal4 driver, the percentage of dead wasps was significantly reduced (8% in both cases). A significant decrease was also observed with the combination of the edin109528 RNAi line and the C564-GAL4 driver (9% killing rate). These results, together with the encapsulation phenotype, indicate that edin is required for the resistance against wasp parasitism in Drosophila larvae.
Lamellocytes have a central role in the resistance against L. boulardi parasitism. They are not found in the hemocoel of healthy, unchallenged Drosophila larvae, but they are formed in response to a wasp infection [15–17]. To investigate whether the expression of edin in the fat body is required for lamellocyte formation, we bled hemocytes of wasp-challenged larvae 48–50 hours after infection. Plasmatocytes and lamellocytes were visualized using the eaterGFP (green) and msnCherry (red) reporters, respectively. As is shown in Fig 3A and 3B, all of the hemocytes in the unchallenged larvae express the eaterGFP reporter and are msnCherry-negative, indicating that only plasmatocytes are present. Lamellocytes are msnCherry-positive, large, and flat cells. They are present only in the infected larvae (Fig 3A’ and 3B’) and are found both in RNAi treated and control larvae, indicating that edin expression in the fat body is not required for lamellocyte formation upon a wasp infection (Fig 3B’). It is noteworthy that the infected larvae contain cells that express both eaterGFP and msnCherry reporters, showing that some of the cells are undergoing plasmatocyte to lamellocyte transition and are not yet fully differentiated lamellocytes (Fig 3A’ and S2 Fig).
In order to obtain additional information about the role of Edin after wasp infection, we used flow cytometry and the msnCherry,eaterGFP reporter to analyze hemocytes of larvae, where edin was knocked down in the fat body. Fig 3C–3D’ show representative scatter plots of hemocytes of uninfected and infected larvae with edin RNAi in the fat body as well as age-matched uninfected and infected control larvae at the 27–29 hour time point. Lamellocytes were induced in spite of edin depletion in the fat body. When comparing hemocyte numbers of uninfected and infected control larvae and edin RNAi larvae, we found that although lamellocyte numbers of infected animals did not differ (p = 0.061, Fig 3E), the plasmatocyte numbers generally increased approximately two to three fold after infection in controls but remained constant in edin knock-down larvae (Fig 3E). Taken together, Edin was dispensable for lamellocyte formation but seemed to be necessary to increase plasmatocyte numbers after a wasp infection.
In order to properly encapsulate wasp eggs, blood cells must adhere and spread on the egg surface until the egg is finally encapsulated. The Rac GTPase Rac2 regulates the actin cytoskeleton that mediates the spreading of plasmatocytes on the wasp egg [23]. To ensure that the defect in encapsulation is not caused by a defective plasmatocyte function, we tested whether plasmatocytes adhere and spread normally on glass slides and on wasp eggs. In our experimental setting, lamellocytes appear 20 hours after parasitization. To get only plasmatocytes, we bled larvae 14 hours after wasp infection and stained the microtubules and the actin cytoskeleton (Fig 4A and 4B”). We measured the tubulin to actin ratio from approximately 120 hemocytes of larvae with edin RNAi in fat body and control larvae, and found no significant difference in the spreading behavior (control: tubulin/actin = 0.46, standard deviation = 0.18; edin RNAi: tubulin/actin = 0.42, standard deviation = 0.21; p = n.s., S1 Table). Another way of looking at spreading behavior is assaying the distribution of the NimC1 protein that is specific for plasmatocytes. The NimC1 protein forms a cytoplasmic ring in control cells, whereas it accumulates in the center of the cell in Rac2 mutants [23]. NimC1 antibody staining of plasmatocytes on the wasp egg 14 hours after parasitization of edin RNAi larvae was indistinguishable from controls (Fig 4C and 4D) indicating normal adhesion and spreading of plasmatocytes in vivo.
The defining early events of capsule formation are the recognition of the wasp egg by plasmatocytes [14] and a significant increase of hemocytes in circulation. [24]. To study whether edin expression is required to increase plasmatocyte numbers in the early stages of an infection, we counted plasmatocytes 14 hours after wasp infection using flow cytometry. As is shown in Fig 4E, edin RNAi in the fat body resulted in more than three times fewer cells compared to controls (p<0.001). Taken together, Edin is dispensable for lamellocyte formation but it is necessary to increase plasmatocyte numbers in circulation in the early stages of a wasp infection.
Sessile plasmatocytes reside attached to the skin of Drosophila larvae and form a hematopoietic compartment that releases blood cells in response to a wasp infection [25, 26]. In order to see if the decreased numbers of plasmatocytes were due to a defect in releasing the sessile plasmatocytes into circulation, we imaged the Fb-GAL4-driven edin RNAi larvae and the respective control crosses 27–29 hours after the wasp parasitization, and again used the msnCherry,eaterGFP reporter line to allow the visualization of plasmatocytes (green) and lamellocytes (red). In the uninfected controls (Fig 5A–5D, top row), the banded pattern of plasmatocytes and the lymph gland could been seen. The bands represented plasmatocytes that resided in the sessile compartment in the absence of an immune stimulus. When the larvae were infected by wasps, the green banded pattern disappeared (Fig 5E–5G) and lamellocytes appeared in the hemolymph (Fig 5E’–5G’). This was due to the activation of the hemocytes in the sessile compartment in response to the wasp infection, which causes the cells to leave the compartment and enter the circulation, where many differentiate into lamellocytes [25, 26]. Consistent with our flow cytometry data (Fig 3), when edin was knocked down in the fat body, lamellocytes still appeared in the circulation showing that Edin did not affect the formation of lamellocytes (Fig 5H’). However, in the edin knockdown larvae the banded pattern of plasmatocytes was not disrupted as in the controls (Fig 5H and 5H’). Of note, overexpression of edin in the fat body did not disrupt the banded pattern indicating that the overexpression of edin alone was not sufficient for releasing the sessile hemocytes into the circulation (S3 Fig). In conclusion, our data suggest that edin expression in the fat body affects plasmatocyte activation and release from the sessile compartment. This suggests that the silencing of edin results in a compromised response to L. boulardi parasitism in the early stages of the infection, and that the altered resistance is due to insufficient plasmatocyte numbers in circulation.
Encapsulation is a complex response against a wasp attack in fruit fly larvae and it requires the concerted action of activated hemocytes. In the course of the encapsulation response, plasmatocytes and the encapsulation-specific lamellocytes form a multilayered capsule around the wasp egg and sequester the invading parasite from the hemocoel of the larva. In addition to inducing the encapsulation response, a wasp infection causes changes in the expression profile of the fruit fly genes [3, 4]. Our results show that edin was rapidly induced in response to an infection by the endoparasitoid wasp Leptopilina boulardi and that edin expression in the fat body, but not in hemocytes, was required to mount a normal encapsulation response against the wasp. Encapsulation was not blocked entirely, however, as approximately 10% of the larvae encapsulated the wasp egg, when edin was knocked down in the fat body. Nevertheless, lamellocyte numbers were unaffected and plasmatocyte spreading behavior was normal. Instead, in larvae where edin was knocked down in the fat body, fewer plasmatocytes were present in circulation, while more hemocytes were retained within the sessile compartment. These data indicate that the presence of lamellocytes alone is not enough for the fruit fly larva to kill the wasp egg. Sufficient numbers of plasmatocytes are also needed.
We discovered that knocking down edin in the fat body did not affect lamellocyte differentiation but compromised the increase of plasmatocyte numbers after a wasp infection. The impaired encapsulation response observed in our study could be therefore due to the misregulation of hemocyte proliferation and/or activation. Because plasmatocyte function was not impaired, as the cells were able to attach and spread normally onto glass slides and wasp eggs, the lowered plasmatocyte number could be the cause of the defects observed in the encapsulation response. Other studies have shown that high hemocyte numbers are associated with an increased resistance against parasitoid wasps in D. melanogaster as well as in other Drosophila species [27–30], although the molecular mechanisms behind this phenomenon are not understood. In our study, the lowered numbers of plasmatocytes are observed already early on during the wasp infection (14 h post infection), suggesting that the function of Edin is critical at the onset of an immune response. This might be the case also in the context of an antimicrobial response, where edin knock down seems to have a modest effect on the levels of some antimicrobial peptides during the early phases of a bacterial infection [20].
Studies have shown that, when hemocytes are activated after an immune stimulus, the banded pattern formed by plasmatocytes is disrupted and the cells are released into the circulation [25, 26, 31], where they can differentiate into lamellocytes [16, 17, 26]. The mobilization of sessile cells occurs prior to the release of hemocytes from the lymph gland [17, 26], and this disruption of the banded pattern is caused by changes in the adhesive properties of the cells. Several genes have been reported to be involved in the attachment of the sessile hemocytes to the sessile compartment [25, 32]. For example, the conserved Rho family of GTPases, namely Rac1 and Rho, regulate the release of sessile cells through the regulation of the adhesive properties of the cells [33, 34]. It has also been suggested that sessile hemocytes adhere to laminin under the larval integument in a syndecan-dependent manner [35]. Additionally, the EGF-repeat containing receptor Eater, which was originally identified for its role in the phagocytosis of bacteria [36], was recently reported to be required in plasmatocytes for the adhesion of hemocytes to the sessile compartment [37]. In our current study, we show that sessile plasmatocytes of edin RNAi larvae did not leave the sessile bands, and the numbers of circulating plasmatocytes did not change after a wasp infection, yet normal amounts of lamellocytes were formed. Despite comparatively normal amounts of lamellocytes, the encapsulation response was impaired when the sessile plasmatocytes could not be mobilized. Hence, besides forming the first layer of the capsule and giving rise to lamellocytes, plasmatocytes have other functions in the encapsulation response that are dependent on edin expression in the fat body.
Our results imply that the effect of Edin is non-cell autonomous and that it seems to act as a molecule that signals from fat body to hemocytes either directly or indirectly. Although the humoral and cellular aspects of Drosophila immunity are often depicted as separate, several studies have provided evidence of the interaction between hemocytes and the fat body. For example, the antimicrobial peptide response to an E. coli infection in domino mutants which lack hemocytes, is normal, but these mutants fail to induce Diptericin during a gut infection by Erwinia carotovora suggesting that hemocytes mediate a signal from the gut to the fat body [38, 39]. In line with these data, Brennan et al. have shown that Psidin acts in the hemocytes to activate the production of Defensin in the fat body [40]. Another example of crosstalk between hemocytes and the fat body is the requirement of Upd3 expression in hemocytes to activate the JAK-STAT pathway in the fat body of adult flies [41]. Furthermore, in larvae, the production of the cytokine Spätzle by hemocytes is needed for the activation of Toll-mediated AMP production in the fat body [42]. Hemocytes are also mediators of the transport of the nitric oxide from its site of production in the gut epithelia to the fat body, where AMP production via the Imd pathway is activated [43, 44]. However, contradicting data also exist for adult flies showing that the ablation of hemocytes by apoptosis does not affect AMP induction in the fat body [45, 46]. A more recent study has shown that the interaction between the fat body and hemocytes is crucial in controlling tumor cell death [47]. Recently, we also showed that Toll signaling in the fat body controlled hemocyte differentiation and activation, but that it did not play a major role in the immune response against L. boulardi as the wasps were able to suppress Toll signaling in the fat body [48]. These examples point to the existence of active tissue-to-tissue signaling that orchestrates appropriate immune responses against different immune challenges. According to our results, Edin functions as a cytokine-like molecule, but the receptor for Edin and its localization remain to be studied. Edin might signal directly from the fat body to the hemocytes, but it may also signal to other tissues or cells that then affect the function of the hemocytes in the sessile compartment (Fig 6). Although Edin is not structurally conserved outside brachyrecan flies [20], its cytokine-like function might be conserved, as in the case of the Spätzle-like function of the vertebrate nerve growth factor β [49], for example.
Based on our results Edin appears to be a key regulator in the cross-talk between fat body and hemocytes in the context of a wasp infection. As in the encapsulation response, the granuloma formation in vertebrates also requires the recruitment of many different cell types. For example, the adult zebrafish responds to a Mycobacterium marinum infection by enclosing the infectious foci in granulomas [50, 51], but also the intracellular bacterium Listeria monocytogenes is sequestered inside granulomas to constrain the infection [52]. Whether information obtained from genetically tractable model organisms such as Drosophila melanogaster, will lead to a better understanding of the pathophysiology of granuloma formation remains to be studied.
UAS-edin RNAi (CG32185) flies #109528 and #14289 (hereafter called edin109528 and edin14289) were obtained from the Vienna Drosophila Resource Center. The driver lines used in this study were the fat body-specific driver Fb-GAL4, the hemocyte-specific driver HmlΔ;He-GAL4 [48] and C564-GAL4, which was obtained from Prof. Bruno Lemaitre (Global Health Institute, EPFL, Switzerland). The C564-GAL4 driver is expressed in many tissues such as the fat body, lymph gland, salivary glands, gut and brain but not in hemocytes [22].
The hemocyte reporter lines eaterGFP (for plasmatocytes) [53] and MSNF9mo-mCherry (for lamellocytes, hereafter called msnCherry) [54] were obtained from Robert Schulz’s laboratory. The lines were crossed to create the msnCherry,eaterGFP reporter line. The mCherry,eaterGFP reporter was further crossed with Fb-GAL4 and edin RNAi109528 to obtain the mCherry,eaterGFP;Fb-GAL4 and mCherry,eaterGFP;edin109528 lines. Canton S flies were used for RNA extractions.
Ten GAL4-driver virgin females were crossed with five RNAi male flies and allowed to lay eggs at +25°C. w1118 flies and GAL4-driver virgin females crossed with w1118 males and w1118 virgin females crossed with RNAi males were used as controls. The flies were transferred daily into fresh vials and the vials containing eggs were transferred to +29°C. On the third day after egg-laying, the larvae were infected with 20 female and 10 male wasps of the Leptopilina boulardi strain G486. The larvae were infected for 2 hours at room temperature after which the wasps were removed and the larvae were transferred back to +29°C.
The encapsulation properties were assayed 27–29 hours after the infection and the killing ability of the larval immune system 48–50 hours after the wasp infection. The egg was scored as encapsulated when traces of melanin were found on it. To analyze the killing ability of the Drosophila larva, three types of phenotypes were scored. The wasp was scored as killed if a melanized wasp egg or melanized wasp larva without other living wasp larvae was found in the hemolymph, whereas the wasp was scored as living when a living wasp that had escaped a melanized capsule was present or when a living wasp larva without any melanized particles was found in the hemocoel.
Eight to ten Canton S larvae per sample were snap frozen on dry ice at 0 hours or 3 hours after the wasp infection. The fat bodies were dissected in 1x PBS 24 hours after the wasp infection and kept on ice. Both larvae and fat bodies were homogenized in TRIsure reagent (Bioline, London, UK) and total RNAs were extracted according to the manufacturer’s instructions.
Quantitative RT-PCR was carried out using the iScript One-Step RT-PCR kit with SYBR Green (Bio-Rad, Hercules, CA, USA) and the Bio-Rad CFX96 (Bio-Rad) instrument according to the manufacturer’s instructions. Results were analyzed with the Bio-Rad CFX Manager software version 1.6. Actin5C was used as a housekeeping gene. The following primers were used: Forward 5’-CTCGTGTCCTGCTGTCTG-3’ and reverse 5’-GCCTTCGTAGTTGTTCCG-3' for edin and forward 5’-CGAAGAAGTTGCTGCTCTGG-3’ and reverse 5’-AGAACGATACCGGTGGTACG-3’ for Actin5C.
Drosophila larvae were imaged using 3rd instar larvae 27–29 hours after the wasp infection. The larvae were washed three times in H2O and embedded on microscope slides in a drop of ice-cold glycerol. The larvae were immobilized at -20°C before imaging. The Zeiss ApoTome.2 was used for live imaging of larvae. For hemocyte imaging, the larvae were washed three times in H2O, and the hemocytes were bled into 1 x PBS 48–50 hours after the wasp infection. Uninfected controls of the same age were also used. The hemocytes were let to adhere to the glass surface of a microscope slide for 30 minutes, after which they were fixed with 3.7% paraformaldehyde for 5 minutes. The samples were washed with PBS and mounted with the Prolong Gold Anti-Fade reagent with DAPI (Molecular Probes). Hemocyte imaging was carried out with the Zeiss AxioImager.M2 microscope with Zeiss AxioCam and the Zen Blue 2011 software and with the Zeiss LSM780 in the case of the antibody-stained hemocytes. The hemocyte images were processed with ImageJ 1.49p (Rasband WS, ImageJ, U.S. National Institutes of Health, Bethesda, Maryland, USA, imagej.nih.gov/ij, 1997–2012).
Hemocytes from infected and control larvae were bled into 1 x PBS with 8% BSA to obtain the hemocytes. Flow cytometry was used to detect eaterGFP-positive and msnCherry-positive cells in these samples. The Accuri C6 flow cytometer (BD, Franklin Lakes, NJ, USA) was used to run the samples, and the data was analyzed using the BD Accuri C6 software. The gating strategy is explained in S2 Fig.
For F-actin and α-tubulin stainings, hemocytes were bled from 15 larvae per cross into 20 μl of 1 x PBS with 8% BSA in pools of three larvae per well and allowed to spread on a glass slide for 45 minutes. Cells were fixed with 3.7% paraformaldehyde/PBS solution for 10 minutes, washed three times with PBS and permeabilized for 5 minutes with 0.1% Triton X-100 before antibody staining. Cells were incubated for 2 hours with an unconjugated mouse α-tubulin monoclonal antibody (Life Technologies, 1μg/ml concentration) followed by one hour incubation with the Alexa Fluor 405 goat anti-mouse secondary antibody (Life Technologies, a 1:500 dilution in 1% BSA in PBS). F-actin was visualized by incubating the cells for 30 minutes with the Alexa Fluor 680 nm Phalloidin stain (Invitrogen) diluted to 1:50 in 1x PBS with 1% BSA. After this, the cells were washed 3 times with PBS and mounted using the ProLong Gold antifade mountant (Life Technologies). We measured the area of Phalloidin and α-tubulin staining with ImageJ 1.49p and calculated the ratio of α-tubulin to Phalloidin areas.
Wasp eggs with hemocytes attached onto them were collected from fly larvae 12–14 hours after infection in a drop of 8% BSA in 1 x PBS, fixed with 3.7% paraformaldehyde/PBS solution for 10 minutes, washed three times with PBS, and stained for 4 hours with an undiluted mixture of monoclonal P1a and P1b (NimC1) plasmatocyte-specific antibodies [55]. Thereafter, the samples were washed 3 times with PBS and incubated with the Alexa Fluor 405 goat anti-mouse secondary antibody (Life Technologies, 1:500 dilution). The eggs were mounted with 50% glycerol prior to imaging. Three eggs per cross were imaged.
Edin expression data was analyzed using an independent samples two-tailed T-test, with unequal variances assumed. The analysis was carried out using Microsoft Office Professional Plus Excel 2013. The threshold for statistical significance was established as p<0.05.
We applied a Generalized Linear Model (glm) in R 3.1.2 (2014-10-31)— “Pumpkin Helmet” (R Development Core, 2003) to analyze the encapsulation and parasite killing data (R Core Team 2014, R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria, http://www.R-project.org/). The categorical explanatory variable was “Cross” and the binary response variable was numbers of “successful encapsulation” or “killed parasites” and numbers of “failed encapsulation” or “failed parasite killing”. Differences between specific crosses were analyzed by Chi-square tests.
We analyzed the cell spreading data and cell numbers 14–16 hours post infection with Welch’s T-test implemented in R 3.1.2 (2014-10-31) (R Development Core, 2003). The data were log-transformed prior to the analyses to obtain normal distribution.
Full factorial analysis of variance (ANOVA) was applied to data on plasmatocyte and lamellocyte numbers 27–29 hours after infection with cross and infection status (infected or not infected) as explanatory variables. The data did not meet the requirement for normal distribution and was log transformed prior to the analyses. In the analysis of plasmatocyte numbers, a significant interaction term was found between cross and infection status and therefore plasmatocyte numbers were further analyzed conducting ANOVAs separately for each cross with infection status as explanatory variable. This data was analyzed using IBM SPSS Statistics version 22.
|
10.1371/journal.pcbi.1002292 | Estimated Comparative Integration Hotspots Identify Different Behaviors of Retroviral Gene Transfer Vectors | Integration of retroviral vectors in the human genome follows non random patterns that favor insertional deregulation of gene expression and may cause risks of insertional mutagenesis when used in clinical gene therapy. Understanding how viral vectors integrate into the human genome is a key issue in predicting these risks. We provide a new statistical method to compare retroviral integration patterns. We identified the positions where vectors derived from the Human Immunodeficiency Virus (HIV) and the Moloney Murine Leukemia Virus (MLV) show different integration behaviors in human hematopoietic progenitor cells. Non-parametric density estimation was used to identify candidate comparative hotspots, which were then tested and ranked. We found 100 significative comparative hotspots, distributed throughout the chromosomes. HIV hotspots were wider and contained more genes than MLV ones. A Gene Ontology analysis of HIV targets showed enrichment of genes involved in antigen processing and presentation, reflecting the high HIV integration frequency observed at the MHC locus on chromosome 6. Four histone modifications/variants had a different mean density in comparative hotspots (H2AZ, H3K4me1, H3K4me3, H3K9me1), while gene expression within the comparative hotspots did not differ from background. These findings suggest the existence of epigenetic or nuclear three-dimensional topology contexts guiding retroviral integration to specific chromosome areas.
| Understanding how retroviral vectors integrate in the human genome is a major safety issue in gene therapy, since a concrete risk of developing tumors associated with the integration process has been observed in several clinical trials. Statistical analyses confirmed the non randomness of the integration. Where and why do virus-specific integrations tend to accumulate in the genome? We compared integration preferences of two retroviral vectors derived from HIV and MLV, which are used in most gene therapy trials for hematological disorders, in their actual clinical targets, i.e., human hematopoietic stem/progenitor cells. We developed a new statistical method to find areas of the genome, called comparative hotspots, where integration preferences are significantly different. We modeled the integration process as a stochastic process, so that integration sites are seen as samples from an unknown virus-specific probability density function. Thus, the problem became to identify areas where two empirical density functions differ significantly. The comparison of nonparametric variability bands around the estimated integration densities allowed identifying and ranking candidate comparative hotspots. Results indicated clear differential patterns of integration between HIV and MLV, leading to new hypotheses on the mechanisms governing retroviral integration.
| Seminal clinical studies have recently shown that transplantation of stem cells, genetically modified by retroviral vectors, may cure severe genetic diseases such as immunodeficiencies [1], [2], skin adhesion defects [3] and lysosomal storage disorders [4]. Unfortunately, some of these studies also showed the limitations of retroviral gene transfer technology, which may cause severe and sometimes fatal adverse effects. In particular, insertional activation of proto-oncogenes by vectors derived from the Moloney murine leukemia virus (MLV) caused T-cell lymphoproliferative disorders in five patients undergoing gene therapy for X-linked severe combined immunodeficiency [5], [6], and pre-malignant expansion of myeloid progenitors in two patients treated for chronic granulomatous disease [7]. Pre-clinical experiments showed that HIV-derived lentiviral vectors are less likely to cause insertional gene activation than MLV vectors. Most of the studies on retroviral integration preferences, however, have been carried out on cell lines that poorly represent the genomic characteristics of somatic stem cells, or on limited numbers of patient-derived cells. A better understanding of the interactions between retroviral vectors and the genome of clinically relevant target cells may provide a more rational basis for predicting genotoxic risks in clinical gene therapy.
A large number of studies have focused on the molecular mechanisms by which mammalian retroviruses choose their integration sites in the target cell genome. After entering a cell, the retroviral RNA genome is reverse transcribed into double-stranded DNA, and assembled in pre-integration complexes (PICs) containing viral as well as cellular proteins. PICs associate with the host cell chromatin, where the virally encoded integrase mediates proviral insertion in the genomic DNA. Retroviral integration is a non-random process, whereby PICs of different viruses recognize components or features of the host cell chromatin in a specific fashion [8]. The LEDGF/p75 protein has been identified as the main factor tethering HIV PICs to active chromatin [9], while mechanisms underlying integration site selection of other retroviruses remain largely unknown. We recently showed that MLV-derived vectors integrate preferentially in hotspots near genes involved in the control of growth, differentiation and development of hematopoietic cells and flanked by defined subsets of transcription factor binding sites; this suggested that MLV PICs are tethered to transcriptionally active regulatory regions engaged by basal components of the RNA Pol II transcriptional machinery [10], [11]. On the contrary, HIV-derived vectors target expressed genes in their transcribed portions away from regulatory elements, suggesting a different evolutionary strategy for these two viruses.
The molecular basis of retroviral target site selection is still poorly understood. The concept of integration “hotspot” was introduced to describe areas of the genome where integrations accumulate more than expected by chance in the absence of any selection process [10]. Hotspots therefore differ from “common integration sites” (CIS), which were defined as sites recurrently associated with virus-induced malignant expansion [12]. The final goal in finding hotspots is to investigate genomic properties that lead certain areas to “attract” or “refuse” integration. We suggested in previous work that integration preferences are dependent on the intrinsic gene density distribution and on the type of vector [13], [14].
In this paper we develop a statistical methodology to detect “comparative” hotspots, i.e. areas of the genome where integration intensities of MLV and HIV appear to differ. We do not find regions where the viruses prefer to integrate, but where the integration patterns are different. Our approach followed two steps: first candidate comparative hotspots were identified by comparing variability bands around estimated integration intensities along the genome, and then each candidate comparative hotspot was tested in turn. After multiplicity correction we produced a list of 100 comparative hotspots, ready for further biological validation. Our analysis discriminated regions which were targeted by both viruses, most likely on the basis of their accessibility (high content of active genes), from regions specifically that are preferred by either MLV or HIV. We show that HIV and MLV integrate differently in regions spanning 0.2 to >6 Mb in the human genome, with specific patterns. In particular, HIV-specific hotspots are wider and contain a larger number of genes. The preference of HIV or MLV for these regions cannot be explained by the known viral target site selection preferences, or by the expression characteristics of the targeted genes, suggesting the existence of epigenetic or nuclear topology contexts that drive retroviral integration to specific chromosome territories.
We developed a new statistical method to compare the integration preferences of distinct retroviral vectors in the human genome and we used it to analyze a collection of ∼30,000 MLV and HIV-vector insertion sites in human CD34+ hematopoietic stem-progenitor cells [15]. Figure 1 illustrates how the methodology performs on chromosome 6. We compared the two integration propensities for each arm and strand separately. The blue 99% variability band corresponds to the integration density of HIV, estimated from our data; in red the band for MLV. When the two bands stay apart, one above the other, a candidate comparative hotspot is identified as the segment of such empty intersection. These are depicted as blue and red thick segments in the center of the plot. In chromosome 6 we identified 12 candidate hotspots where MLV shows more integrations than HIV, and 5 candidate hotspots where HIV is dominating. In most of chromosome 6, we found no differences in integration patterns. Note the high peak of integration on the p-arm for HIV, on both strands, corresponding to the MHC locus. Similar plots for all chromosomes are available in supplementary material Figure S1 and Figure S2.
The panels of Figure 2 show two typical situations in detail. The panel A (left side) from the HIV HLA locus in chromosome 6, arm p; upper panel refers to strand+, lower panel refers to strand -. We see how the estimated variability bands, around the non-parametrically estimated integration densities, are clearly apart from each other. The bands overlap at both ends of the comparative hotspot, which is therefore well defined. The width of the bands describes the statistical uncertainty attached to the estimated densities: in both cases the MLV bands are quite thin, as there are a total large number of integrations. The bands for HIV are larger; the exact density function is difficult to estimate with limited sample size. Despite the uncertainty, the candidate hotspot in panel A is clearly identified. The panels B of Figure 2 show a candidate comparative hotspot in the plus strand of chromosome 6, arm q, which has no corresponding in the minus strand. In other locations of the genome, the two bands often overlap simply due to lack of data, rather than because the two vectors are equally distributed. This indicates that our method will leave undiscovered comparative hotspots (false negatives). Not all the candidate comparative hotspots that we identified were clearly distinguishable.
Our analysis led to 256 candidate comparative hotspots on all chromosomes (see, supplementary material, Table S1). Each candidate comparative hotspot was then tested individually. We computed odds ratios, between HIV and MLV odds of integrations in each hotspot, and tested the null hypothesis that the odds ratio is one. P-values were then corrected for multiple testing. This reduced the number of significative comparative hotspot to 100, reported in Table 1.
The length of the hotspots varies between ca. 200,000 bp and 7,000,000 bp, but most are longer than 106 bp. They include between 1 and 179 genes. Of the 100 significative comparative hotspots, 49 have a higher density of HIV integrations (lengths ranging between 378,200 and 6,857,000 bp; median: 2,651,000 bp, 2,027 unique target genes) while 51 contain a higher MLV density (lengths ranging between 211,300 and 6,021,000 bp; median: 1,319,000 bp, 475 unique target genes). The median length of MLV hotspots is about the half of the median length of HIV hotspots with a significative difference (p-value: 2.108•10−06; Mann-Whitney test, p-values computed by permutations). The wideness of HIV hotspots only partially accounts for the higher number of target genes compared to MLV (2,027 vs. 475), as shown by plotting the number of targets per hotspot normalized by the hotspot length (see Figure 3: p-value = 1.953*10−8, Mann-Whitney test, p-values computed by permutations). This hints at gene density as a critical parameter for HIV integration site selection and is in accordance with the recent finding that MLV integration is associated to transcription regulatory regions rather than to genes 11,15,16.
To investigate the categories of genes preferentially targeted by the comparative hotspots we performed a Gene Ontology (GO) classification of HIV and MLV target genes (supplementary material, Table S2). Among the 2,027 genes in the comparative hotspots with HIV preference, the analysis showed a significant enrichment over the background (0.005<p-values<0.05, Fischer's exact test with Bonferroni correction for multiple testing) for genes involved in antigen processing and presentation, and in hormone nuclear receptor activity. Remarkably, both GO terms exclusively contained genes located in the MHC locus on chromosome 6 (highest peak in Figure 1). Among the 475 MLV targets, genes participating in adaptive immune response, signal transduction, and regulation of biological processes were over-represented (0.005<p-values<0.05). Differently from HIV targets, these genes did not belong to the same chromosomal region.
The annotation of oncogenes 17 (Sanger Cancer Gene Census, http://www.sanger.ac.uk/genetics/CGP/Census) incorporated into comparative hotspots (see the full gene Table S1 in supplementary material) did not reveal a significant difference in the targeting frequency between the two vectors, both when considering all genes (0.045 for HIV vs. 0.050 for MLV, p-value: 0.4109) or the sole genes in the 100 significative comparative hotspots (0.047 for HIV vs. 0.054 for MLV, p-value: 0.06707). We next investigated the relation between gene expressions and comparative integration hotspots. We compared the frequency of expressed genes belonging to comparative integration hotspots with the frequency of transcribed genes located elsewhere in the genome. After multiple testing corrections, we found just one hotspot with increased presence of expressed genes with respect to the genomic baseline (hiv_55, adjusted p-value 0.02800; see Table S1, supplementary material for full results). Three comparative hotspots, all with higher MLV density (mlv_50, mlv_51, mlv_124, p-values: 0.00075, 00.00197, 0.00778 respectively), showed instead a reduced presence of expressed genes.
Since there is strong evidence of association between integration sites and specific histone modifications 18,19,20, we also investigated the histone methylation 21 density in comparative hotspots, defined as the methylation intensities (i.e., the number of ChIP-seq reads) in each comparative hotspot, divided by the hotspot length; the same was done for the histone variant H2A.Z. We compared the mean density of histone modifications associated to transcription or heterochromatin in HIV vs. MLV hotspots using the Welch statistic test, which does not assume the same variance for the two groups, with p-values computed by permutations. After adjustment for multiplicity (Bonferroni-Holm) three methylations and the one histone variant analyzed were found to have different mean density in HIV vs. MLV hotspots (H3K4me1, H3K4me3, H3K9me1, H2AZ; adjusted p-values: 0.000096, 0.000010, 0.020986, 0.000018 respectively). Results are summarized in table 2.
The construction of the variability bands of an integration density depends on a design smoothing parameter, as described in the Materials and Methods section. The choice of such smoothing parameters controls the regularity of the variability bands and therefore had an effect on the comparison. We estimated the smoothing parameters in an optimal fashion (in correspondence to which results were reported), but also studied robustness of the hotspots by varying them systematically. We systematically checked if a comparative hotspot would have persisted for larger and smaller smoothing parameters. Figure 4 shows the results of such a sensitivity study for two strands of chromosome 6. The middle line, corresponding to 1, shows the hotspot identified by the two optimal smoothing parameters, while above and below that we see how hotspots would grow and shrink by increasing and reducing the smoothing level. It is important that the chosen segments at level 1 continue to appear for values just above and under, as happened systematically. This visual inspection strengthens the validity of the way we chose the smoothing parameters. See Methods for more details and supplementary material Figures S3, S4, S5, S6, S7 and S8 for robustness plots for all chromosomes.
Integration of MLV-derived retroviral vectors may have significant consequences on gene expression and homeostasis of transduced and transplanted target cells, particularly in the hematopoietic system. The enhancer activity of the MLV LTRs may de-regulate proto-oncogenes, and cause pre-neoplastic clonal expansion 7,22, leukemic transformation without clonal expansion 5,6,23, or no apparent adverse effect 1 depending on the disease context and a number of still ill-defined factors. Integration sites can be used as markers of clonality to study the clonal dynamics of transduced cells in vivo, and provide important clues to predict the potential genotoxicity of MLV integration in a specific cell or disease context 23,24,25,26,27. We used LM-PCR and pyrosequencing to derive high-definition maps of MLV and HIV integration sites in the genome of human CD34+ hematopoietic progenitors. As previously reported 14, MLV integrations were clustered around gene regulatory elements (promoters, enhancers, evolutionarily conserved non-coding regions) bearing epigenetic marks of active transcription (H3K4me1, H3K4me2, H3K4me3, H3K9Ac) and specialized chromatin configurations (H2A.Z). On the contrary, HIV integrations occurred away from regulatory elements, and are associated with histone modification enriched in the body of transcribed genes (H3K36me3 and H2BK5me1). In both cases, statistical analysis identified hotspots of clustered integrations with strong correlation with transcriptional activity, using random integration datasets as controls.
In this study, we identify broad areas of the genome where HIV and MLV integrate differently; therefore it was not expected to find comparative hotspots in areas of high gene expression. This is in accordance with the fact that a single hotspot showed an increased expression level with respect to the rest of the genome. We used non-parametric density estimation and variability bands to identify regions of the genome as candidate comparative, i.e., virus-specific, hotspots. Thereafter, these were tested for significance. The first step delivers a series of bins, of variable length, were the two integration frequencies appear to be different. This strategy is more effective than binning the chromosome with equal size bins, since some of them might not be large enough to contain enough integrations. An optimal bin size algorithm, producing a constant bun size, would easily divide a chromosome in a dozen bins, which would be too large to be of practical interest as candidate hotspots. Our approach generates a list of bins of variable and adaptive length, only in areas of interest. Interestingly, this analysis identified large genomic regions (0.2 to >6 Mb in length) rather than local (<100 kb) hotspots. Most genomic regions are targeted by both virus types, most likely because they contain a high proportion of active genes and regulatory elements. Some regions, however, are targeted by either virus in a specific fashion, where HIV-specific hotspots tend to be larger in size and to contain more genes. The expression and gene ontology characteristics of the genes contained in MLV and HIV-specific regions, however, were comparable, and there are no obvious characteristics that would predict such a striking virus-specific preference. While MLV-specific regions are enriched for histone modifications/variants correlated with active regulatory regions (H3K4me1, H2A.Z), HIV-specific regions have a higher density of H3K4me3, associated to active transcription start sites. Although counterintuitive, given the well-known MLV preference for transcription start sites, this might be simply explained by the higher gene content of HIV-specific hotspots. Unfortunately, genomic distribution of HIV tethering factors, such as LEDGF/p75, is not known, particularly for hematopoietic progenitors, and it is therefore impossible to test whether high protein concentration in specific chromosomal region may explain the HIV-specific preferences.
Interestingly, we found a significant comparative hotspot spanning the entire MHC locus on chromosome 6 (from the MHC class I to the extended MHC class II subregions 28) with increased HIV, but not MLV, integration propensity. Importantly, a gene-centric hotspot definition would have failed to detect this locus, since in this particular case intergenic regions rather than single genes are highly targeted by HIV.
Large, virus-specific hotspots may suggest that tethering of PICs to chromatin favors relatively wide chromosomal territories independently from their content or local concentration of “attractive” features, such as GC content of DNA, binding of factors or transcriptional complexes, nucleosome density or epigenetic marks. This type of preference may instead reflect larger scale, nuclear topology factors that make these regions more accessible to one or another virus type. The modalities by which HIV and MLV access target cell chromatin, may be a critical factor underlying these preferences. MLV is incapable of entering intact nuclei and requires cell division in order to integrate, while HIV is actively imported in interphase nuclei through the nuclear pores. MLV and HIV PICs therefore “see” chromatin in different phases of the cell cycle, and may have access to different regions simply because they are differently exposed. Recent studies showed that alterations in the nuclear pore architecture impairs HIV nuclear import and impacts on integration efficiency, suggesting that access to chromatin is mediated by the nuclear pore and may be a critical component of target site selection 29,30. The HIV-specific hot regions identified in this study may therefore reflect the chromatin organization in the vicinity of the nuclear pore. Studies are in progress to test this hypothesis in clinically relevant target cells.
We worked with a previously published collection of 28,382 HIV and 32,631 MLV retroviral integration sites isolated by linker-mediated PCR (LM-PCR) and pyrosequenced by GS-FLX Genome Sequencer (Roche/454 Life Sciences, Branford, CT) from cord blood-derived human CD34+ hematopoietic stem-progenitor cells 15. The bioinformatics pipeline used to process crude MLV and HIV sequence reads was previously described 15. Briefly, valid reads 20-bp or longer were used to generate a non-redundant dataset using the nrdb tool (available at http://www.advbiocomp.com/blast.html in the AB-BLAST software package). Non-perfectly redundant reads were than mapped onto the human genome, requiring the alignment to start within the first three nucleotides and to possess a minimum of 90% identity. Sequences were discarded when mapping to multiple sites if they had more than one match on the human genome differing in identity less than 2%. Overall valid sequence recovery was similar between MLV and HIV (13.3% and 17.3%, respectively). The expression profile of CD34+ cells was determined by microarray analysis of cytokine-activated cells from three independent umbilical cords. RNA was extracted from 1–2×106 cells, transcribed into biotinylated cRNA and hybridized to Affymetrix HG-U133A plus 2.0 Gene Chip arrays.
Functional clustering of target genes was performed by the DAVID 2.0 Functional Annotation Tool and EASE score, as previously described 10. GO categories were considered over-represented when yielding an EASE score <0.05, after Bonferroni-Holm correction for multiple testing.
Certain areas of the genome cannot be scanned in order to investigate the presence of integrations. This is mainly due to two reasons: genome mappability and the presence of what we call “blind regions”. Although extremely critical in determining the randomness of single integration patterns, genome mappability was not a concern in our comparative study, since only unequivocally mapping reads were considered, for the comparison of MLV and HIV integration patterns (i.e., the mappability bias, if any, was the same for the two vectors). Blind regions instead derive from the use of restriction enzymes and size-selection during the integration library preparation, and represent portions of the genome that are scarcely accessible to detection due to their distance to the closest 3′ restriction site (Figure 5). Specifically, if this distance is shorter than the sensibility of alignment programs, in terms of minimum length of the processable sequence, integration is not identifiable. For example, if a viral vector integrated 10 bps far from the closest 3′ cut sequence, then from the sequencing platform we obtained a 10 bps sequence, that for most of alignment program is not processable. We used Blat 31 which has minimum sequence length of 20 nt. On the other hand, the size-fractionation step only includes fragments <500 nt, this being the maximum estimated length for efficient 454 bead loading (see supplemental methods in 10). Therefore, integrations with a distance to the closest 3′ restriction enzyme site of, for example, 600 bps, would not be detected. These blind regions need to be excluded from further analysis, as it was impossible to determine accurately integration frequencies occurring therein. We first identified these blind regions by looking for the position of restriction sequences over the whole genome. We then cut off the blind regions. When performing density estimation, we skipped blind regions and connected together successive non-blind parts. We assumed smoothness of the density at mending points, as the blind regions were comparably short. Once hotspots were found, the blind areas were placed back in the original topology. Integration analysis was performed separately for each chromosomal arm, so that it was not affected by the centromere, which is a giant blind region. Furthermore, we studied separately each strands, since blind regions are strand specific. The presence of blind regions due to the restriction enzyme digestion is known. It has been shown that the “invisible” portion of the genome is substantially affected by the use of different and/or multiple restriction enzymes 32. We also found the percentage of blind regions to be very significant, ranging from 10% up to 40% of the length of the chromosome. For example, 30.5% of chromosome 1 was blind (total length 247.249.719 bp); see supplementary material Table S1 for percentages for all chromosomes. Blind regions were identified by means of a custom R-script (R ver 2.10 33 and Bioconductor 34) which searched for the TTAA sequences (MseI) on the Hg18 UCSC genome. Once occurrences were identified, blind regions were estimated as follows: from TTAA to 20 bp downstream (due to algorithm limitation) and from 500 bp downstream to the consecutive restriction site (due to deep sequencing platform limitation).
The integration dynamics was modelled as a stochastic process, where integration points were considered as samples from an unknown density function on the region of study D. We assumed that each integration was independent of any other. Each virus was considered as a random variable v with its own unknown probability density function . Comparing integration preferences of two viruses v1 and v2 was then turned into the statistical problem of comparing two unknown densities and , defined on the same genomic range D, based on an independent and identically distributed sample from each of the two densities. The samples were allowed to have different sample size. Our approach was fully nonparametric and led to candidate comparative hotspots, which were then individually tested. Specifically, non-parametric kernel density estimation with Gaussian kernels was used 35. In a basepair , the estimated density , based on the sample in D, is given by the kernel density estimatorwhere K(·) is the kernel and h>0 is the smoothing parameter (bandwidth). We used the Gaussian kernelNotice the scaling of the kernel with h, which controls how much weight each integration xi has in the estimate at a basepair x. We performed a small approximation, as the kernel should integrate to 1 over D, and the domain D is discrete. However, the resolution at basepair level of the chromosome arms is extremely high, so that the effect of this was negligible.
We wished to construct simultaneous confidence bands (at 0.99 level, say) for the two densities to be compared, in order to identify areas (if any) where the confidence bands did not overlap: in such segments of D, one density must clearly be below the other. However, such confidence bands depend on the second derivative of the unknown density, controlling both bias and variance; approximations are available only in some special cases under very strong conditions. We instead calculated pointwise variability bands around the estimated densities, where the variation in the density estimates were taken into account, but the bias was ignored. The segments of the chromosome D where the two variability bands had empty intersection were considered as candidate comparative hotspots. The 0.99 variability band for the estimated density was computed 36 starting with the Taylor expansionwhereis the integral of the squared kernel function and n is the sample size. The root transform allowed obtaining an approximation of the variance which was independent from the unknown density. Therefore, on the square root scale, a level error band could be computed, using the half widtharound the squared root of the estimate, where is the quantile of the normal standard distribution. Then, as in 36, the edges of this band were transformed back to the original scale aswhere the lower bound is set to zero if it took a negative value. We used .
This is not a confidence band and there is no nominal coverage probability. The effect of the bias is to diminish modes and fill valleys, as it depends on the curvature of (and on the bandwidth), see 36. Variability bands of this type were computed for both densities. Typically, a detected candidate comparative hotspot (where the two variability bands had empty intersection) resulted from a pronounced peak in one density and a valley or flat area in the other. In these situations, adjusting for the bias would have strengthened further the indication of a hotspot. On the other hand, the absence of bias adjustment could in some special situations hide a difference. This indicates that in most cases we have identified candidate comparative hotspots conservatively.
We compared the two pointwise variability bands at level , one for each virus, to detect where the bands did not overlap. These segments in D were considered as candidate comparative hotspots. This approach is different from 37, where bins are decided in advance, instead than being data-driven.
Though the band was computed pointwise, it inherited smoothness from the smooth density estimate around which it was built. For computational efficiency, the density was estimated on a grid of points, which were then interpolated with a spline function 38. We did not implement any particular boundary control at the border of the chromosome arm not flanking the centromere.
The choice of the smoothing parameters h1 and h2, one for each viral integration density, is important: too much smoothing would flatten the kernel estimates, hiding every difference; too little smoothing would lead to a too rich and fragmented identification of comparative hotspots, with very high false positive findings. Our choice was to perform an automatic and optimal choice of the smoothing parameter for each density and then study how results would change when this value was perturbed in both directions, towards more and towards less smoothing. We chose the optimal smoothing parameters, hopt, one for each density, using unbiased cross-validation 39. Briefly, hopt is chosen to minimize the measure of closeness of to given by the Integrated Squared Errorthrough a least square, leave-one-out crossvalidation criterion. For this purpose we minimized the estimate of the first two terms of the ISE (the last term does not depend on h) given bywhere denotes the kernel estimator constructed from the data without the observation . See 39,40 for more details. In order to test sensitivity of results with respect to the choice of h, we reparameterized the smoothing parameter as h = hopt s, where the sensitivity factor s was left to vary in [0.05, 20]. We then repeated the comparison of the variability bands for the two viral integration densities, using the crossvalidated optimal smoothing parameter for each virus, adjusted with the same s. We plotted the comparative hotspots while varying s, to see the robustness of each hotspot, as in Figure 4.
Candidate comparative hotspots were then tested individually, to confirm (or not) that the integration propensities of the two viruses were significantly different. As many comparisons were performed, multiple testing correction was done. We computed the odds ratio of the two integration intensities, one for each virus, for each candidate hotspot aswhen HIV had a higher density and the inverse of it when the MLV density was higher instead. Here is the number of integration of HIV falling inside the candidate hotspot H, is the number of integration outside hotspot H, and similarly for MLV. We computed 0.95 confidence intervals for this odds ratio and tested the null hypothesis that the odd ratio is 1. We used the Fisher exact test. Raw p-values were then corrected for multiple testing by Bonferroni-Holm 41. All computations and analyses were performed in R and Bioconductor environment 33,34.
|
10.1371/journal.pgen.1003337 | CELF Family RNA–Binding Protein UNC-75 Regulates Two Sets of Mutually Exclusive Exons of the unc-32 Gene in Neuron-Specific Manners in Caenorhabditis elegans | An enormous number of alternative pre–mRNA splicing patterns in multicellular organisms are coordinately defined by a limited number of regulatory proteins and cis elements. Mutually exclusive alternative splicing should be strictly regulated and is a challenging model for elucidating regulation mechanisms. Here we provide models of the regulation of two sets of mutually exclusive exons, 4a–4c and 7a–7b, of the Caenorhabditis elegans uncoordinated (unc)-32 gene, encoding the a subunit of V0 complex of vacuolar-type H+-ATPases. We visualize selection patterns of exon 4 and exon 7 in vivo by utilizing a trio and a pair of symmetric fluorescence splicing reporter minigenes, respectively, to demonstrate that they are regulated in tissue-specific manners. Genetic analyses reveal that RBFOX family RNA–binding proteins ASD-1 and FOX-1 and a UGCAUG stretch in intron 7b are involved in the neuron-specific selection of exon 7a. Through further forward genetic screening, we identify UNC-75, a neuron-specific CELF family RNA–binding protein of unknown function, as an essential regulator for the exon 7a selection. Electrophoretic mobility shift assays specify a short fragment in intron 7a as the recognition site for UNC-75 and demonstrate that UNC-75 specifically binds via its three RNA recognition motifs to the element including a UUGUUGUGUUGU stretch. The UUGUUGUGUUGU stretch in the reporter minigenes is actually required for the selection of exon 7a in the nervous system. We compare the amounts of partially spliced RNAs in the wild-type and unc-75 mutant backgrounds and raise a model for the mutually exclusive selection of unc-32 exon 7 by the RBFOX family and UNC-75. The neuron-specific selection of unc-32 exon 4b is also regulated by UNC-75 and the unc-75 mutation suppresses the Unc phenotype of the exon-4b-specific allele of unc-32 mutants. Taken together, UNC-75 is the neuron-specific splicing factor and regulates both sets of the mutually exclusive exons of the unc-32 gene.
| Tissue-specific and mutually exclusive alternative pre–mRNA splicing is a challenging model for elucidating regulation mechanisms. We previously demonstrated that evolutionarily conserved RBFOX family RNA–binding proteins ASD-1 and FOX-1 and a muscle-specific RNA–binding protein SUP-12 cooperatively direct muscle-specific selection of exon 5B of the C. elegans egl-15 gene. Here we demonstrate that two sets of mutually exclusive exons, 4a–4c and 7a–7b, of the unc-32 gene are regulated in tissue-specific manners and that ASD-1 and FOX-1, expressed in a variety of tissues, can regulate the neuron-specific selection of unc-32 exon 7a in combination with the neuron-specific CELF family RNA–binding protein UNC-75. We determine the cis-elements for the RBFOX family and UNC-75, which separately reside in intron 7b and intron 7a, respectively. By analyzing the partially spliced RNA species, we propose the orders of intron removal and the sites of action for the RBFOX family and UNC-75 in the mutually exclusive selection of exon 7a and exon 7b. We also demonstrate that UNC-75 regulates the neuron-specific selection of exon 4b and propose the models of the mutually exclusive selection of exons 4a, 4b, and 4c. These studies thus provide novel modes of regulation for tissue-specific and mutually exclusive alternative splicing in vivo.
| Alternative splicing of pre-mRNAs is a major source of proteomic complexity in metazoans. More than 90% of human multi-exon genes undergo alternative pre-mRNA processing and many alternative splicing events are controlled in tissue- and cell-type dependent manners [1]. Mis-splicing of pre-mRNAs underlie many inherited diseases [2]. A variety of auxiliary trans-acting factors and cis-acting elements regulating alternative splicing have been identified [3], [4], [5], [6]. Recent genome-wide studies of protein-RNA interactions for trans-acting splicing factors led to creation of RNA splicing maps [7]. Combinations of hundreds of RNA features were used to assemble ‘splicing codes’ to predict splicing patterns in four major tissues to a significant extent [8]. However, much of our knowledge of splicing regulation relies on experiments utilizing cultured cells, and therefore complex mechanisms of the tissue-specific regulation of pre-mRNA splicing by coordination of multiple trans-factors and cis-elements in living organisms remain less understood.
Mutually exclusive splicing should consist of multiple steps of strictly regulated splicing events and offers good models for elucidating regulation mechanisms for alternative pre-mRNA splicing [9], [10]. Among them, fibroblast growth factor receptor (FGFR) genes have been well studied because tissue-specific and mutually exclusive selection of exons encoding a part of the extracellular domain determines the ligand specificity of the receptors [11], [12], [13], [14]. The most extraordinary examples of the mutually exclusive exons are in the Drosophila Dscam gene [9], [10], which has four clusters of mutually exclusive exons. Selection of only one exon out of 48 candidate exons at a time for the exon 6 cluster is considered to be regulated by a complex system of competing RNA structures and a globally-acting cluster-specific splicing repressor [15], [16]. However, the molecular mechanisms governing the selection patterns for the entire Dscam mRNA remain poorly understood [10].
A nematode Caenorhabditis elegans is intron-rich like vertebrates and is an excellent model organism for studying the regulation mechanisms of pre-mRNA processing in vivo [17]. Up to 25% of its protein-coding genes are estimated to undergo alternative pre-mRNA processing and hundreds of the events are developmentally regulated [18]. We developed a fluorescence alternative splicing reporter system and visualized spatio-temporal selection patterns of mutually exclusive exons in living worms [19], [20], [21]. Through genetic and biochemical analyses, we successfully identified evolutionarily-conserved and broadly-expressed RBFOX (named after RNA binding protein, fox-1 homolog (C. elegans)) family splicing regulators ASD-1 and FOX-1 and a muscle-specific RNA-binding protein SUP-12 as the co-regulators of the muscle-specific selection of exon 5B of the egl-15 gene encoding the sole homolog of the FGFRs in C. elegans [19], [22].
The unc-32 gene of C. elegans, analyzed in this study, encodes the a subunit of V0 complex of vacuolar-type H+-ATPases considered to be proton pumps that acidify intracellular organelles [23], [24]. The unique property of the unc-32 gene as a model for studying alternative splicing regulation is that it has two sets of mutually exclusive exons (Figure 1A). Only one exon at a time is selected from three exons 4a, 4b and 4c; only one exon is selected at a time from two exons 7a and 7b. Of the six possible combinations of exons 4 and 7, the three isoforms UNC-32A (4a/7b), UNC-32B (4b/7a) and UNC-32C (4c/7b) were predominantly detected [25] and appear to be developmentally regulated [18], raising questions about the exact selection patterns and the regulation mechanisms in vivo. In the present study, we demonstrate that unc-32 exon 4 and exon 7 are selected in tissue-specific manners and that a neuron-specific RNA-binding protein UNC-75 regulates the neuron-specific selection of exons 4b and 7a.
We first confirmed mutually exclusive selection of endogenous unc-32 exon 4 and exon 7 by RT-PCR (Figure 1B). Consistent with the previous report based on microarray profiling and high-throughput sequencing of mRNAs from synchronized worms [18], the splicing patterns of both exon 4 and exon 7 appeared to be developmentally regulated; the relative amounts of the exon 4b isoform and the exon 7a isoform dramatically decreased at the L4 stage (Figure 1B).
Next we visualized the selection patterns of unc-32 exon 4 and exon 7 in vivo by applying our fluorescence alternative splicing reporter system [20]. A trio of symmetric reporter minigenes for exon 4 was constructed by cloning the genomic fragment spanning from exon 3 through exon 5 upstream of one of three fluorescent protein cDNA cassettes and by introducing artificial termination codons into two of the three mutually exclusive exons in each construct (Figure 1C). From these minigenes, we expect expression of Venus-fusion protein (E4a-Venus), monomeric red fluorescent protein (mRFP)-fusion protein (E4b-mRFP) and enhanced cyan fluorescent protein (ECFP)-fusion protein (E4c-ECFP) only when exon 4a alone, 4b alone and 4c alone are selected, respectively (Figure 1C). In the same way, a pair of symmetric exon 7 reporter minigenes was constructed by cloning the genomic fragment spanning from exon 6 through exon 8 upstream of either of two fluorescent protein cDNA cassettes and by introducing an artificial termination codon into one of the two mutually exclusive exons in each construct (Figure 1D). From these minigenes, we expect expression of enhanced green fluorescent protein (EGFP)-fusion protein (E7a-EGFP) and mRFP-fusion protein (E7b-mRFP) when exon 7a and exon 7b are selected, respectively (Figure 1D).
We utilized a ubiquitous promoter to drive expression of the minigenes and generated transgenic reporter worms (Figure 1E–1G). Expression of the three fluorescent proteins in the exon 4 reporter worms varied among tissues; intestine, the nervous system and pharynx predominantly or exclusively expressed E4a-Venus, E4b-mRFP and E4c-ECFP, respectively (Figure 1E). Expression of the two fluorescent proteins in the exon 7 reporter worms also showed tissue-specificity. Most tissues predominantly expressed E7b-mRFP and therefore the worms appear almost Red (Figure 1F). Confocal microscopy revealed that neurons in head ganglia predominantly expressed E7a-EGFP (Figure 1G). The expression patterns of the exon 4 and exon 7 reporters were consistent throughout development. We suspected that lack of the developmental change in the reporter expression was due to ectopic expression of the reporters in tissues that do not express the endogenous unc-32 gene. A transcriptional fusion reporter, however, revealed that the unc-32 promoter drives expression in intestine, neurons and pharynx (Figure 1H), the major tissues where the exon 4 and exon 7 reporters were expressed. We therefore concluded that the mutually exclusive exons of the unc-32 exon 4 and exon 7 reporter minigenes are selected in tissue-specific and not developmentally regulated manners.
To focus on the neuron-specific selection of exon 7a, we utilized the rgef-1 (also known as F25B3.3) promoter to drive pan-neuronal expression of the exon 7 reporter. As expected, transgenic worms with an integrated reporter allele ybIs1622 [rgef-1::unc-32E7a-EGFP rgef-1::unc-32E7b-mRFP] predominantly expressed E7a-EGFP in the nervous system and appeared Green with a dual-bandpass filter (Figure 2A). We therefore used the rgef-1 promoter for further analyses described below.
As cis-elements regulating alternative splicing are often evolutionarily conserved in the genus Caenorhabditis [19], [21], [22], [26], we first focused on the five stretches in flanking introns of exons 7a and 7b conserved among C. elegans, C. briggsae and C. remanei (Figure 2B, Figure S1A). We constructed five pairs of mutagenized exon 7 reporter minigenes M1 to M5 (Figure 2C, Figure S1A) and found that disruption of the UGCAUG stretch in intron 7b (M1) changed the color of the exon 7 reporter from Green to Orange (Figure 2D), while disruption of the other stretches had no apparent effect (Figure S1B). We therefore concluded that the UGCAUG stretch in intron 7b is required for the neuron-specific selection of exon 7a.
The UGCAUG stretches are known to be specifically recognized by the RBFOX family splicing regulators in metazoans including C. elegans [27]. We have previously reported that the RBFOX family proteins in C. elegans, ASD-1 and FOX-1, redundantly repress egl-15 exon 5B by specifically binding to the UGCAUG stretch in the upstream intron [19]. The asd-1; fox-1 double mutant is defective in expression of a muscle-specific fibroblast growth factor receptor (FGFR) isoform EGL-15(5A) and shows the egg-laying-defective (Egl-d) phenotype [19]. To test whether ASD-1 and FOX-1 also regulate the neuron-specific selection of unc-32 exon 7a, we crossed the reporter allele ybIs1622 with the asd-1 and fox-1 mutants. As expected, the reporter worms turned the color from Green to Yellow in the single mutant backgrounds (Figure 2E, top and middle) and to Orange in the double (Figure 2E, bottom), confirming that ASD-1 and FOX-1 are redundantly involved in the neuron-specific selection of exon 7a from the exon 7 reporter.
To confirm direct and specific binding of ASD-1 and FOX-1 to the UGCAUG stretch in intron 7b in vitro, we performed an electrophoretic mobility shift assay (EMSA) using the radiolabelled RNA probes with an intact (WT) and a mutagenized (M1) sequence as in the reporters (Figure 2F, top). Recombinant full-length ASD-1 and FOX-1 proteins (Figure 2F, bottom left) efficiently shifted the mobility of the WT probe (Figure 2F, bottom right, lanes 1–4, 9–12) and less efficiently of the M1 probe (lanes 5–8, 13–16) in a dose-dependent manner, demonstrating direct and specific binding of ASD-1 and FOX-1 to the UGCAUG stretch. These results led to the conclusion that ASD-1 and FOX-1 regulate the selection of exon 7a from the unc-32 exon 7 reporter via the UGCAUG stretch in intron 7b in the nervous system.
To identify other regulator(s) that confer the neuron-specificity to the exon 7 reporter, we mutagenized the ybIs1622 strain to screen for mutants exhibiting altered colors. We successfully isolated many homozygous viable strains with Yellow, Orange or Red phenotype (Figure 3A, Figure S2). In some other strains, most neurons turned red while some remained green (Red/Green) (Figure 3A, Figure S2). The color phenotypes were completely penetrated within the strains. Notably, all the Red and Red/Green strains also showed an uncoordinated (Unc) phenotype while the Orange or Yellow strains did not.
By single-nucleotide polymorphism (SNP)-based mapping and sequencing candidate genes, we identified mutations in the unc-75 gene in the color mutants. The unc-75 gene was originally identified as the gene responsible for the Unc phenotype caused by defects in synaptic transmission [28]. The exon 7 reporter allele ybIs1622 crossed with an existing null allele unc-75 (e950), which lacks exon 1 through exon 5 and exhibits the Unc phenotype [28], showed the RedUnc phenotype (data not shown), confirming that the color phenotype is caused by loss of function of the unc-75 gene.
UNC-75 belongs to the CUG-BP and ETR-3-like factor (CELF) family of RNA-binding proteins, which have two N-terminal RNA recognition motifs (RRMs) followed by a so-called divergent domain and the third RRM at the C-terminus. The CELF family can be divided into two subfamilies CELF1–2 and CELF3–6 according to sequence similarities [29] and UNC-75 is the sole member of the CELF3–6 subfamily in C. elegans [29]. Although UNC-75 has been shown to be expressed exclusively in the nervous system and localized to subnuclear speckles [28], it is still unknown what process UNC-75 is involved in.
The mutations identified in the unc-75 gene are summarized in Figure 3B. All of the five alleles with the RedUnc phenotype have nonsense mutations in exon 6 or exon 7 (Figure 3B). Figure 3C shows amino acid sequence alignments of the three RRMs from the CELF family members in C. elegans and human. A missense mutation (yb1714) in the conserved glycine residue in the α1β2 loop of RRM1 and four other mutations (yb1697, yb1705, yb1709 and yb1718) in the region between exon 1 and exon 3 were associated with the Red/GreenUnc phenotype (Figure 3B and 3C, top). A missense mutation (yb1700) in the conserved glycine residue in the RNP1 motif of RRM2 (Figure 3B and 3C, middle) and a missense mutation (yb1698) in the conserved arginine residue in the divergent domain (Figure S3) were associated with the Yellow phenotype. A missense mutation (yb1723) in the RNP2 motif and a 4-aa deletion (yb1725) in the RNP1 motif in RRM3 were associated with the Orange phenotype (Figure 3B and 3C, bottom). These results suggested that all the three RRMs and the divergent domain are required for UNC-75 to properly regulate the selection of exon 7a in the nervous system.
During the course of cDNA cloning, we found another UNC-75 mRNA isoform lacking exon 8 corresponding to the anterior half of RRM3 (Figure 3B and 3D, lane 1). Although the skipping of exon 8 does not cause a frame-shift or nonsense-mediated mRNA decay (NMD), the deletion of the half of RRM3 would more significantly affect the function of UNC-75 than the yb1723 and yb1725 mutations (Figure 3C, bottom). As many splicing factors are known to regulate their own expression at the pre-mRNA splicing level, we analyzed the effect of the nonsense mutation in the unc-75 gene on its own mRNAs. The splicing patterns of the UNC-75 mRNAs were not affected in the asd-1; fox-1 mutant (Figure 3D, lane 2), while the Δexon 8 isoform was undetected in the unc-75 (yb1701) mutant (lane 3), consistent with the idea that UNC-75 negatively regulates its own expression by repressing exon 8.
We noticed that the C-termini of the CELF family proteins as well as the RBFOX family proteins are evolutionarily conserved and match the consensus of the hydrophobic PY nuclear localization signal (PY-NLS) [30] (Figure 4A). To test this idea, we analyzed the effect of substitution or deletion of the C-terminal motifs upon subcellular localization of the proteins (Figure 4B–4G). The substitution of the three residues in the PY element of UNC-75 (Figure 4A) disrupted the nuclear localization of UNC-75 (Figure 4B, 4C), confirming that the C-terminal motif of UNC-75 functions as the PY-NLS. In the same way, the deletion of the 7 and 16 residues from the C-termini of ASD-1 and FOX-1, respectively (Figure 4A), disrupted the nuclear localization of the proteins (Figure 4D–4G), indicating that the C-terminal portions of ASD-1 and FOX-1 are the sole NLSs.
To determine the element(s) in the exon 7 cluster region that UNC-75 directly and specifically recognizes in vitro, we performed EMSAs with the radiolabelled RNA probes schematically illustrated in Figure 5A (top panel). Recombinant full-length UNC-75 protein shifted the mobility of Probe 2 (Figure 5B, lanes 3,4) and Probe 2-1 (lanes 9–12) and not of the other probes (Figure 5B). As more than half of Probe 2-1 overlapped with Probe 1 or Probe 2-2, we prepared a shorter probe 2-1-1 (Figure 5A) containing most of the sequence unique to Probe 2-1. UNC-75 shifted the mobility of Probe 2-2-1 (Figure 5C, lanes 1–4, 25–28), demonstrating that UNC-75 directly and specifically binds to the 2-1-1 fragment in this region.
To further specify the element(s) necessary for the UNC-75-binding, we prepared the five mutant probes 2-1-1a to -1e, in each of which G and C residues in a short stretch were replaced with A (Figure 5A, bottom panel). UNC-75 shifted the mobility of the probes 2-1-1a to -1d (lanes 5–20) similarly to Probe 2-1-1, while the mobility of Probe 2-1-1e was unaffected by UNC-75 (lane 21–24), indicating that the UUGUUGUGUUGU stretch disrupted in Probe 2-1-1e is essential for UNC-75 to specifically recognize the 2-1-1 fragment.
To test whether all the three RRMs of UNC-75 are involved in the recognition of the 2-2-1 fragment, we performed EMSAs using three mutant recombinant proteins UNC-75 (G53S), UNC-75 (G165E) and UNC-75 (L431F) (Figure 6A, left), each of which had a single missense mutation in one of the three RRMs as found in the mutant alleles. UNC-75 (G53S) and UNC-75 (G165E) less efficiently shifted the mobility of Probe 2-1 and Probe 2-1-1 than wild-type UNC-75 (Figure S4, lanes 1–10; Figure 6A, right, lanes 1–13). UNC-75 (L431F) failed to shift the mobility of these probes (Figure S4, lanes 11–13; Figure 6A, right, lanes 14–17). These results indicated that the missense mutations affected the RNA-binding properties of UNC-75 in vitro and that all the three RRMs of UNC-75 are required for the specific recognition of unc-32 intron 7a.
To specify which of the three RRMs of UNC-75 mediates the specific recognition of the elements in Probe 2-1-1, we prepared recombinant proteins for each of the three RRMs and performed an EMSA (Figure 6B). The RRM3 protein (Figure 6B, right, lanes 12–16) as well as full-length UNC-75 (lanes 17,18) shifted the mobility of Probe 2-1-1, while the RRM1 or RRM2 protein did not (lanes 1-11), indicating that only RRM3 can bind to Probe 2-1-1 by itself. So we used only RRM3 protein for a further EMSA with the mutant 2-1-1 probes. The RRM3 protein shifted the mobility of Probe 2-2-1 (Figure 6C, lanes 1–3) and the mutant probes 2-2-1a to -1d (lanes 4–15) and not of Probe 2-2-1e (lanes 16–18), indicating that RRM3 specifically recognizes the UUGUUGUGUUGU stretch.
To test the requirement of the UUGUUGUGUUGU stretch for the splicing regulation in vivo, we constructed another mutant pair of the exon 7 reporter minigenes M6 that has the same substitutions as in Probe 2-2-1e and generated transgenic worms. The disruption of the UUGUUGUGUUGU stretch turned the color into Orange (Figure 6D), confirming that the stretch is essential for the selection of exon 7a in the nervous system in vivo.
Next we analyzed the effects of the RBFOX family and UNC-75 on the endogenous unc-32 gene. In the wild-type L1 larvae, the exon 7a and exon 7b mRNA isoforms were almost equally detected (Figure 7A, left, lane 1). The relative amount of the exon 7a isoform was reduced in the asd-1; fox-1 double mutant (lane 2) and unc-75 mutant (lane 3) backgrounds. A double inclusion isoform or a double skipping isoform was not detected in either of the mutants. These results are consistent with their color phenotypes and the splicing patterns of the exon 7 reporter expressed in the nervous system (Figure 7A, right) and confirm that the RBFOX family and UNC-75 regulate the mutually exclusive splicing of exons 7a and 7b of the endogenous unc-32 gene.
For mutually exclusive alternative splicing, upstream and downstream flanking introns should be sequentially excised (Figure S5A). To obtain insight into the orders of intron removal for the production of the exon 7a and 7b mRNA isoforms, we analyzed the relative amounts of the four partially spliced RNA species to the unspliced RNA from the exon 7 reporter expressed in the nervous system by RT-PCR using two pairs of an intronic primer and a reporter-specific exonic primer. With one primer set, the partially spliced RNA in which exon 6 was spliced to exon 7b (E6/E7b–E8) was detected but the other partially spliced RNA in which intron 6 was removed (E6/E7a-E7b-E8) was almost undetectable in the wild-type, asd-1; fox-1 double mutant and unc-75 mutant worms (Figure 7B, left). With the other primer set, the partially spliced RNA in which exon 7a was spliced to exon 8 (E6–E7a/E8) was detected but the other partially spliced RNA in which intron 7b was removed (E6-E7a-E7b/E8) was almost undetectable in these worms (Figure 7B, right). The relative amounts of the four partially spliced RNAs to the unspliced RNA are summarized in Figure 7C. Of the two partially spliced RNAs that are the putative intermediates for the exon 7a isoform, E6–E7a/E8 was predominantly detected and its relative amount was decreased in the mutants. Of the two partially spliced RNAs that are the putative intermediates for the exon 7b isoform, E6/E7b–E8 was predominantly detected and its relative amount was increased in the mutants. Although these partially spliced RNAs may not necessarily be the processing intermediates but instead dead-end products, the changes in the relative amounts of the partially spliced RNAs are in good correlation with the changes in the amounts of the mature mRNA isoforms in the mutants. These results suggest that E6–E7a/E8 and E6/E7b–E8 are the major processing intermediates for the exon 7a and exon 7b isoforms, respectively. Notably, the mutations in the RBFOX family genes and unc-75 differentially affected the relative amounts of these partially spliced RNAs, suggesting their differential roles in the alternative splicing regulation of unc-32 exon 7.
We also analyzed the partially spliced RNAs from the endogenous unc-32 gene with endogenous RNA-specific pairs of primers. The result revealed consistent but weaker effects of the mutations in the RBFOX family genes and unc-75 on the partially spliced RNAs (Figure S5B–S5C). Considering that the endogenous unc-32 gene is expressed not only in the nervous system but also in pharynx and intestine that select exon 7b, this result is consistent with the idea that the RBFOX family and UNC-75 regulate the selection exon 7a from the endogenous unc-32 gene in the same way as from the reporter in the nervous system. Taking the relative strength of the splice sites in this region (Figure S5D) into account, Figure 7D summarizes the schematic models for the mutually exclusive selection of unc-32 exon 7, which will be discussed later (see Discussion).
As unc-32 exon 4b is also selected in a neuron-specific manner (Figure 1F), we tested whether the RBFOX family and UNC-75 are also involved in the regulation of the exon 4 cluster. Consistent with the absence of a (U)GCAUG stretch in the exon 4 cluster region, the asd-1; fox-1 double mutation did not affect the splicing patterns of exon 4 of the endogenous unc-32 gene (Figure 8A, lanes 1, 2). On the other hand, the unc-75 mutation caused marked reduction of the exon 4b isoform (lane 3). Furthermore, the neuron-specific expression of E4b-mRFP from the exon 4 reporter ybIs1891 was also abolished in the unc-75 mutant (Figure 8B, compare with Figure 1F). These results indicated that UNC-75 is required for the selection of exon 4b in the nervous system. We performed an EMSA to localize the UNC-75-binding site(s) with four overlapping probes in the exon 4 cluster region, but none of the probes were shifted as effectively as Probe 2 in Figure 5B by full-length UNC-75 (data not shown). We speculate that other cooperative factors may be required for the specific recognition of the exon 4 cluster region by UNC-75.
We next analyzed the amounts of the six theoretical partially spliced RNAs or putative processing intermediates (Figure S6) from the endogenous unc-32 gene in the wild type and unc-75 mutant. Both of the two putative processing intermediate RNAs for the exon 4b isoform were detected in the wild type (Figure 8C, left and right panels, lanes 1, 2) but almost undetectable in the unc-75 mutant (lanes 3, 4) consistently with the amount of the mature exon 4b isoform. Only one (E3–E4a/E5) of the two partially spliced RNAs that are the putative intermediate RNAs for the exon 4a isoform was detected and its relative amount was increased in the unc-75 mutant (Figure 8C–8D). Only one (E3/E4c–E5) of the two partially spliced RNAs that are the putative intermediate RNAs for the exon 4c isoform was detected and its relative amount was increased in the unc-75 mutant (Figure 8C–8D). These results propose a model schematically illustrated in Figure 8E; UNC-75 represses splicing of exon 3 to exon 4c and exon 4a to exon 5 and promotes splicing of exon 4b to exons 3 and 5.
The exon 4b-specific mutation in the unc-32 (e189) allele causes the uncoordinated (Unc) phenotype (Figure 1A) [25] and our results demonstrated that exon 4b is specifically selected in the nervous system in an UNC-75-dependent manner. So we speculated that the mutations in unc-75 should bypass the requirement of exon 4b in the nervous system. Consistent with this idea, the OrangeNon-Unc allele unc-75 (yb1725) suppressed the Unc phenotype of the unc-32 (e189) mutant (Figure 8F). As neuron-specific ectopic expression of any of the three major isoforms can rescue unc-32 (e189) (Figure S7), we reasoned that unc-75 (yb1725) suppressed unc-32 (e189) via the ectopic expression of the exon 4a or exon 4c isoform in the nervous system. Thus, UNC-75 is the critical splicing factor for the nervous system to specifically select unc-32 exon 4b in vivo.
In this study, we demonstrated that the two sets of the mutually exclusive exons of the unc-32 gene are independently regulated in tissue-specific manners by utilizing the fluorescence alternative splicing reporters. Our study revealed that intestine, neurons and pharynx express the UNC-32A (4a/7b), UNC-32B (4b/7a) and UNC-32C (4c/7b) isoforms, respectively. The expression patterns are consistent with the previous report that these three are the major isoforms and that the translational fusion reporter consisting of the unc-32 promoter through exon 4b is expressed in the nervous system [25]. The neuron-specific isoforms become relatively less abundant in elder stages in the RT-PCR experiments (Figure 1B) probably due to decrease in the relative population and/or mass of the nervous system. Our study thus demonstrated the importance of carefully analyzing alternative splicing patterns at a single cell resolution in vivo.
Figure 7D illustrates the proposed models of the neuron-specific selection of exon 7a. In the non-neuronal tissues, exon 7a is skipped presumably due to its weak splice sites and exon 6 is readily spliced to exon 7b (right panel). In neurons, UNC-75 specifically binds to its cis-elements in intron 7a to repress exon 7b and the RBFOX family and UNC-75 activate splicing between exon 7a and exon 8 (left panel). The models may explain why the mutations in unc-75 exerted more sever effects on the selection of exon 7a in the nervous system than the disruption of the RBFOX family genes; in the absence of UNC-75, exon 7b would be readily spliced to exon 6, where the target exon of the RBFOX family is no longer left (right panel).
Figure 8E illustrates the proposed model of the mutually exclusive selection of unc-32 exon 4. In neurons, UNC-75 activates splicing both between exon 3 and exon 4b and between exon 4b and exon 5 so that exon 4b alone is selected. In intestine and pharynx, splicing between exon 4a and exon 5 and between exon 3 and exon 4c, respectively, occurs first to determine the fate of the pre-mRNA presumably depending on other tissue-specific factor(s). The proposed order of intron excision for each isoform in this model explains the fidelity of the mutually exclusive selection from the three exons of the unc-32 exon 4 cluster.
The number of the mutually exclusive exons in a cluster is at most two in mammals. The fidelity of the mutually exclusive splicing relies on steric hindrance due to close proximity of the mutually exclusive exons [31], incompatibility between U2-type and U12-type splice sites [32], splicing regulators that repress one exon and activate the other [12], [33] and/or mRNA surveillance system [34]. We have previously raised regulation models for two genes with mutually exclusive exons in C. elegans. In the case of egl-15, the RBFOX family and SUP-12 cooperatively repress the splice acceptor of the upstream exon [22]. In the case of let-2, ASD-2 activates the splice donor of the downstream exon [21]. In the present study, we demonstrate novel types of regulation; for unc-32 exons 7a and 7b, UNC-75 and the RBFOX family switch the first splicing from E6/E7b to E7a/E8; for unc-32 exons 4a, 4b and 4c, UNC-75 activates both the splice acceptor and the donor of exon 4b. It has been recently suggested that the mutually exclusive exons in the slo-1 gene are regulated in intragenic coordination with downstream alternative splicing events although the splicing patters are not analyzed at a single cell resolution [35]. Thus, the order of intron excision and the modes of regulation for the mutually exclusive exons vary from case to case even in the simple model organism.
In this study, we identified the first endogenous alternative splicing events regulated by the CELF3–6 subfamily. A recent splicing-sensitive microarray analysis of the unc-75 mutant suggested only one affected gene, lec-3 [36], but the selection patterns of the putative target exons in each tissue in vivo are not known yet and the function of UNC-75 in the splicing regulation of the lec-3 gene are to be experimentally defined. In vertebrates, the CELF1–2 subfamily proteins CELF1 (also known as CUG-BP1) and CELF2 (also known as ETR-3 and CUG-BP2) are broadly expressed, highest in heart, skeletal muscle and brain, and their biological functions and biochemical properties are well characterized [29], [37]. On the other hand, CELF3 to CELF6 are predominantly expressed in the nervous system [38], [39], [40], [41], [42] and have been shown to regulate alternative splicing in heterologous minigene systems [33], [38], [43], [44], [45], [46]. However, the in vivo functions and biochemical properties of the CELF3–6 subfamily are less characterized [29] presumably due to their functional redundancy.
We identified the short fragment specifically recognized by UNC-75 in unc-32 intron 7a and provided the genetic and biochemical evidence that all the three RRMs are required for the recognition and regulation of the unc-32 pre-mRNA (Figure 3C, Figure 6A). Among them, RRM3 recognizes the UUGUUGUGUUGU stretch in the target element by itself (Figure 6C). On the other hand, the stretches that RRM1 and RRM2 recognize could not be determined, although our data shown in Figure 5 and Figure 6 do not preclude the possibility that RRM1 and/or RRM2 also recognize the UUGUUGUGUUGU stretch. These results suggest that recognition of target RNAs by RRM1 and RRM2 is context-dependent or cooperative, which may explain why it is difficult to determine the precise binding sites or consensus sequences for RRM1 and RRM2. The CELF1–2 subfamily has been shown to bind to a variety of UG-rich and related sequences via the three RRMs in a context-dependent manner [47], [48], [49], [50], [51], [52]. Considering the amino acid sequence similarities between the two subfamilies (Figure 3C), it is reasonable that UNC-75 also recognizes the UG-rich sequences. Collection of the unc-75 mutant alleles revealed that the conserved stretch in the N-terminal portion of the divergent domain is also involved in the recognition and/or splicing regulation of unc-32 (Figure S3). This is consistent with the previous reports that the N-terminal portion of the divergent domain of CELF4 is involved in the RNA recognition and/or splicing regulation in minigene contexts [44], [46].
The RedUnc mutant alleles have nonsense mutations in unc-75 exon 6 or 7 (Figure 3A, 3B), while some other mutants show the Red/Green phenotype (Figure 3A, Figure S2), suggesting cell-type-dependent remaining activity of UNC-75 in such mutants. Paradoxically, most of the Red/Green alleles have nonsense mutations or splice site mutations in exon 1, 2 or 3 (Figure 3B), indicative of fatal effects on the UNC-75 expression. The remaining activity of UNC-75 in certain neurons might derive from the use of alternative promoters in the upstream region or in intron 3 to bypass exons 1–3, although we have not experimentally identified such mRNA isoforms from the unc-75 gene.
We demonstrated that the C-termini of all the CELF family and the RBFOX family proteins match the consensus of the PY-NLS and that the C-termini are indeed required for the proper nuclear localization of UNC-75, ASD-1 and FOX-1 (Figure 4). As RRM3 of the CELF family resides at the C-terminus, the PY-NLS is overlapping with RRM3 and is highly conserved. It has been reported that deletion of a C-terminal KRP stretch affected the nuclear localization of UNC-75 in neurons [28], consistent with our finding. In contrast to the PY-NLSs in the RBFOX and CELF families, the PY-NLS was originally identified in the internal portion of hnRNP A1 and other RNA-binding proteins including hnRNP D, TAP, HuR, hnRNP F and hnRNP M [30]. Most of the PY-NLSs predicted in many other proteins are structurally divergent and reside in the internal portion [30]. Evolutionary conservation of the sequences and positions of the PY-NLSs in the RBFOX and CELF families may suggest importance of their positions for the functions of these proteins.
In this and previous studies, we demonstrated that the broadly-expressed RBFOX family proteins ASD-1 and FOX-1 regulate the neuron- and muscle-specific alternative splicing events in a target-specific manner in combination with the neuron-specific RNA-binding protein UNC-75 and the muscle-specific RNA-binding protein SUP-12 [22], respectively. Similarly, an RBFOX family protein RBFOX2 is expressed in a variety of cell types in mammals, yet it can regulate the epithelium-specific alternative splicing of the FGFR2 gene in coordination with epithelium-specific splicing factors ESRP1 and ESRP2 [11], [12]. The RBFOX family splicing regulators have only one RNA-binding domain that can specifically recognize the (U)GCAUG stretch in the target pre-mRNAs [27], [53]. Therefore, the presence of the (U)GCAUG stretch in the pre-mRNAs is not the sole determinant of the tissue-specificity but can be considered to offer an opportunity for the combinatorial and context-dependent regulation of alternative splicing. Considering their broad expression, the RBFOX family may regulate alternative splicing with a variety of tissue-specificity in cooperation with other tissue-specific factors in both mammals and C. elegans.
To construct the unc-32 exon 7 reporter cassettes, the unc-32 genomic fragment was cloned upstream of either mRFP1 [54] or EGFP (Clontech) cDNA in the Entry vectors by using In-Fusion system (BD Biosciences) and the artificial termination codons were introduced with QuickChange (Stratagene). The reporter minigenes for unc-32 exon 4 and the unc-32 transcriptional fusion were constructed as described previously [20]. The sequences of the primers used in the plasmid construction are available in Table S1.
The worms were cultured following standard methods. Generation of transgenic worms, mutant screening and mapping were performed as described previously [20]. The images of the fluorescence reporter and mutant worms were captured using fluorescence stereoscopes (MZ16FA and M205FA, Leica) equipped with color, cooled CCD cameras (DP71, Olympus and DFC310FX, Leica) or a confocal microscope (FV500, Olympus) and processed with Metamorph (Molecular Devices) or Photoshop (Adobe).
The RT-PCRs were performed essentially as described previously for amplifying the mature mRNAs [20] and the partially spliced RNAs [55]. The RT-PCR products were analyzed by standard agarose gel electrophoresis or by using BioAnalyzer (Agilent) and the sequences of the RT-PCR products were confirmed by direct sequencing or cloning and sequencing. The sequences of the primers used in the RT-PCR assays are available in Table S2.
The amino acid sequences of the proteins used in the alignments were retrieved from the protein sequence databases derived from GenBank and RefSeq. The accession numbers are as follows: human CELF1, NP_006551; CELF2, NP_001020247; CELF3, AAK07474; CELF4, NP_064565; CELF5, NP_068757; CELF6, NP_443072; hnRNP A1, AAH02355; hnRNP D, BAA09525; TAP, AAD20016; hnRNP F, NP_004957; hnRNP H1, NP_005511; hnRNP H2, NP_062543; RBFOX1, Q9NWB1; RBFOX2, NP_001026865; mouse RBFOX3, NP_001034256; Drosophila A2bp1, AAQ22527; C. elegans UNC-75, AAQ19851; ETR-1, NP_493673; EXC-7, CAA85327; HRPF-1, AAK21490; ASD-1, NP_497841; FOX-1, NP_508446. The amino acid sequences were aligned by Clustal W using Lasergene (DNASTAR).
The rabbit polyclonal anti-UNC-75 antiserum (9493R2R) was generated by using denatured His-tagged full-length UNC-75 protein as described previously [55]. The rabbit polyclonal anti-ASD-1 (RbD8211) and -FOX-1 (RbD8209) antisera were generated with the mixtures of synthetic peptides TVEKLNDFDYKVAL+C and C+RGVPQPGRIPTSTA for anti-ASD-1 and C+GKVKDDPNSDYDLQ and C+LPSYQMNPALRTLN for anti-FOX-1 by Operon Biotechnologies (Tokyo, Japan). The expression vectors for untagged UNC-75 and HA-tagged ASD-1 and FOX-1 were constructed by using Destination vectors pDEST-cDNA3 and pDEST-ME18S-3HA (H.K.), respectively. HeLa cells were transfected with the expression vectors by utilizing GeneJuice (Novagen). For UNC-75, the cells were stained with anti-UNC-75 (9493R2R), Alexa488-conjugated goat anti-rabbit IgG (Molecular Probes) and DAPI (Vector Laboratories) and fluorescence images were captured with a compound microscope (Eclipse E600, Nikon) and a CCD camera (DP71, Olympus). For ASD-1 and FOX-1, the cells were stained with anti-ASD-1 (RbD8211) or -FOX-1 (RbD8209), Cy3-conjugated goat anti-rabbit IgG (Jackson) and TO-PRO3 (Molecular Probes) and the confocal images were acquired with FV1000 (Olympus).
The expression vectors for FLAG-tagged ASD-1, FOX-1 and UNC-75 proteins were constructed using the primers listed in Table S3 and the recombinant proteins were prepared as previously described [55]. The 32P-labelled RNA probes were prepared as described previously [55] using the template oligo DNAs listed in Table S4 and the PCR products amplified with the primes in Table S5. The EMSAs were performed as described previously [55].
|
10.1371/journal.ppat.1006108 | Probability of Transmission of Malaria from Mosquito to Human Is Regulated by Mosquito Parasite Density in Naïve and Vaccinated Hosts | Over a century since Ronald Ross discovered that malaria is caused by the bite of an infectious mosquito it is still unclear how the number of parasites injected influences disease transmission. Currently it is assumed that all mosquitoes with salivary gland sporozoites are equally infectious irrespective of the number of parasites they harbour, though this has never been rigorously tested. Here we analyse >1000 experimental infections of humans and mice and demonstrate a dose-dependency for probability of infection and the length of the host pre-patent period. Mosquitoes with a higher numbers of sporozoites in their salivary glands following blood-feeding are more likely to have caused infection (and have done so quicker) than mosquitoes with fewer parasites. A similar dose response for the probability of infection was seen for humans given a pre-erythrocytic vaccine candidate targeting circumsporozoite protein (CSP), and in mice with and without transfusion of anti-CSP antibodies. These interventions prevented infection more efficiently from bites made by mosquitoes with fewer parasites. The importance of parasite number has widespread implications across malariology, ranging from our basic understanding of the parasite, how vaccines are evaluated and the way in which transmission should be measured in the field. It also provides direct evidence for why the only registered malaria vaccine RTS,S was partially effective in recent clinical trials.
| Malaria is transmitted to humans by the bite of an infectious mosquito though it is unclear whether a mosquito with a high number of parasites is more infectious than one with only a few. Here we show that the greater the number of parasites within the salivary gland of the mosquito following blood-feeding the more likely it is to have transmitted the disease. A clear dose-response is seen with highly infected mosquitoes being more likely to have caused infection (and to have done so quicker) than lightly infected mosquitoes. This suggesting that mosquito-based methods for measuring transmission in the field need to be refined as they currently only consider whether a mosquito is infected or not (and not how heavily infected the mosquito is). Novel transmission reducing drugs and vaccines are tested by experimentally infecting people using infectious mosquitoes. This work indicates that it is important to further standardise infectious dose in malaria experimental infections to enable the efficacy of new interventions to be accurately compared. The work also provides direct evidence to suggest that the world’s first licenced malaria vaccine may be partially effective because it fails to provide protection against highly infected mosquitoes.
| Mosquito-to-human malaria transmission occurs when sporozoites from the salivary gland of the mosquito are injected into the skin during blood-feeding. Parasites then pass to the liver where they replicate, each sporozoite yielding many thousands of merozoites which go on to cause patent infection. Relatively little is known about the population dynamics of malaria between the bite of the infected mosquito and patency. It is currently assumed that the probability of mosquito-to-human transmission is determined simply by the presence of salivary gland sporozoites and not the total number of parasites. For example malaria transmission intensity and the human force of infection are measured using the entomological inoculation rate (EIR) the average number of infectious mosquito bites per person per year [1]. EIR is calculated by multiplying the human biting rate by the proportion of mosquitoes with salivary gland sporozoites and therefore does not explicitly consider how heavily infected the mosquitoes are.
Controlled human malaria infection (CHMI) trials are increasingly being used to evaluate new vaccines [2] and offer a new way to examine the basic biology of the parasite. Volunteers are deliberately infected through the bite of a blood-feeding mosquito or from the direct syringe injection of titrated cryopreserved sporozoites harvested from laboratory reared mosquitoes [3]. The number of syringe injected sporozoites has been shown to correlate with the probability of infection, though the number of viable/infectious sporozoites inoculated is unknown [3, 4]. Therefore it is unclear whether the dose response is caused by an increased probability of the person being injected by a rare viable sporozoite, or whether it is the higher number of viable sporozoites which increases the chance of infection.
Mosquito-delivered sporozoites more reliably recreate the natural infection process than syringe inoculation [4–6], here the probability of infection increases with the number of infectious bites [7, 8]. However a review of the CHMI trial literature and model-system data indicates that no study has thoroughly examined the impact of the number of sporozoites within those bites. The limited number of studies on the subject have given inconsistent results [5, 9–11], potentially resulting from small sample sizes and the use of insensitive statistical methods [12]. Many highly infected mosquitoes fail to inject sporozoites during blood feeding [13–15] so it is again unclear whether the dose response is caused by more bites having a higher probability of injecting a viable sporozoite, or whether the increased quantity of sporozoites injected increases transmission. Currently it is not possible to accurately determine the salivary gland sporozoite burden prior to blood feeding nor the quantity inoculated into the host, which current estimates suggest may vary by multiple orders of magnitude [14–16]. Consequently CHMI studies estimate parasite challenge by dissecting mosquitoes after blood-feeding, counting the number of sporozoites remaining in the salivary glands using microscopy (here referred to as the number of residual-sporozoites, which is scored on a logarithmic scale) [9]. Residual-sporozoite score has been shown to correlate with the number of sporozoites injected into the skin [15].
Time to detectable blood-stage infection is used in CHMI studies to differentiate the efficacy of vaccines which did not elicit sterilising immunity, with longer pre-patent periods indicating a smaller number of parasites passing from the liver to the blood and therefore higher intervention efficacy. Historically, shorter pre-patent periods have been correlated with volunteers receiving more infectious bites [8, 17, 18] but recent analyses of the Plasmodium vivax malaria datasets using more appropriate statistical techniques suggest evidence for such an association is weak [12], though once again the impact of sporozoite intensity was not thoroughly investigated.
The relationship between pathogen dose and disease transmission may have important implications for the epidemiology of the disease, how transmission is measured in the field, and the impact of vaccination [19]. It appears intuitive that transmission will increase with the size of the inoculum though there is little empirical evidence to support this assumption and no direct evidence from human malaria. The dose required to infect a host varies dramatically between diseases [20] so the importance of sporozoite load and its relationship with the probability of transmission needs to be clarified.
This study combines results from novel and previously published experimental infections and analyses them with new statistical methods to rigorously determine the impact of the number of residual-sporozoites on the probability of infection. In the CHMI trials analysed here volunteers are bitten by infected mosquitoes until they have received 5 bites from mosquitoes with >10 residual-sporozoites [21]. This ensures that all control volunteers develop malaria [22] though precludes these volunteers being used to investigate factors influencing the probability of naïve human infection. We therefore analyse the results of a recent CHMI study where volunteers were given a partially effective pre-erythrocytic vaccine (PEV) which targets the circumsporozoite protein (CSP) [23]. The analysis is extended to the Plasmodium berghei–A. stephensi mouse model system [24] where the importance of the number of sporozoites in a mosquito bite can be assessed more thoroughly, with or without the presence of an anti-CSP antibody. To investigate whether sporozoite load influences the size of the liver-to-blood inocula in hosts which become infected (and not just whether or not blood-stage infection develops) the relationship between residual-sporozoite number and the time to patency is assessed. This allows the impact of parasite load to be investigated in a larger human dataset with and without vaccine where most volunteers became infected (S1 Table). For the sake of brevity both humans pre-administered a pre-erythrocytic vaccine and mice pre-administered the anti-CSP 3D11 antibody are referred to as being given a PEV, whilst those that did not are called naïve.
The association between the number of residual-sporozoites and the probability of infection was assessed in 47 vaccinated volunteers (challenged with Plasmodium falciparum using Anopheles stephensi mosquitoes), 9 of which became infected (Fig 1A). Though there was no difference in the mean (or total) residual-sporozoite number in mosquitoes biting persons who became infected or remained uninfected (S2 Table) a binomial model [25] shows that infection was significantly more likely from mosquitoes with >1000 residual-sporozoites (Fig 1B). The best fit model indicates that mosquitoes harbouring >1000 sporozoites had a per bite transmission probability of 9.2% (4.5%-16.0%) whilst those with a lower number of parasites did not measurably contribute to transmission (Fig 1B, see S3 Table for the results of the full model comparison). This model is used to estimate the probability that a volunteer became infected according to the number of residual-sporozoites in the bites they received (Fig 1C). Though there is still considerable variation, the predicted probability of infection is significantly higher in infected volunteers (p = 0.040) supporting the original analysis that mosquito parasite load influences the (per bite) probability of mosquito-to-human transmission. None of the 13 volunteers who received 0 or 1 bites from mosquitoes with >1000 residual-sporozoites became infected.
The analysis is extended to the mouse model system [24] where the importance of the number of sporozoites in a mosquito bite can be assessed more thoroughly, with or without the presence of PEV. A total of 844 mice were included in the analysis, 99 of whom were given a PEV (Fig 1D). Results show a clear dose-dependency, with mosquitoes with a higher number of residual-sporozoites being more likely to transmit malaria both in the naïve mice, and those given a PEV. The probability of infection in a naïve host from a single bite is 32% (19%-46%) from mosquitoes with 1–10 sporozoites and 78% (53%-93%) from those with >1000 sporozoites (Fig 1E). The approximately linear increase in transmission probability with the log-sporozoite number suggests that the per-sporozoite transmission probability is a negative density-dependent process. Model predictions of the probability of infection for each mouse (according to the number of residual-sporozoites of the bites they received) are compared to whether or not they became infected (Fig 1F). There is a significant association between infection status and the model-derived probability of infection (p<0.0001). The model is able to predict infection relatively accurately (for example 95.4% of mice with a >95% probability of infection became infected). Greater variability is seen in the probability of infection for mice than for humans. This is because in the rodent dataset the number of bites a mouse receives varies (in addition to the number of residual-sporozoites within those bites) whereas in the human dataset all volunteers receive 5 bites from mosquitoes with >10 residual-sporozoites.
Mosquitoes with no detected residual-sporozoites by microscopy can, very rarely, infect mice. In total, 291 naïve mice were bitten by mosquitoes with no detectable salivary gland sporozoites following blood-feeding. Of these 13 (4.5%) went on to develop a blood stage infection. The lack of observable sporozoites could be caused by mosquitoes injecting all sporozoites during blood-feeding, or due to measurement error in the counting process. If all mice bitten exclusively by mosquitoes with no detectable sporozoites were removed from the analysis then each additional bite from mosquitoes with no detected residual-sporozoites reduces the probability of infection by 4.6% (0.5%-7.4%). Comparing this model to a reduced version where zero scores have no contribution to transmission shows that including the negative impact of these bites does significantly improve model fit (Akaike information criterion, AIC, with 0 = 445.9, AIC without 0 = 452.4). The cause of this negative impact on transmission is unknown and requires further investigation though it could result from a non-specific immune response initiated by an uninfected mosquito bite[26]. As expected the PEV significantly reduces the probability of infection, reducing the per bite transmission probability by 29.1%. The most parsimonious model indicates that the PEV is more effective against lower sporozoite challenges. It reduced the probability of infection by 70.5% for bites with <101 residual-sporozoites and 16.3% in mice bitten by mosquitoes with >100 sporozoites (AIC constant efficacy model = 623.5, variable efficacy model = 622.5, Fig 1E).
The number of residual-sporozoites significantly influences the time to patency in both naïve volunteers and those given different PEVs. Data from a total of 267 CHMI volunteers who developed malaria were analysed, 192 of these had received a PEV candidate (Fig 2A and 2B). The best fit survival analysis model was unable to differentiate between mosquitoes with no detectable salivary gland sporozoites and those with 1–10 post-feeding perhaps because the numbers of bites with 1–10 sporozoites was relatively low. A full description of the time invariant components of the survival analysis are given in S4 Table. Importantly, the model predicts that naïve volunteers who were infected by 5 mosquitoes with >1000 sporozoites would have detectable parasitemia on average >2 days earlier than those infected by 5 mosquitoes with 11–101 sporozoites. There is a significant association between observed time to patency and survival analysis model predictions based on the number of residual-sporozoites (Fig 2C, p<0.0001). Time to patency analysis was repeated with those P. berghei challenges of mice which developed infection (Fig 2D–2F). Data from 429 mouse experimental challenges that generated blood-stage infection were used in the survival analysis, 31 of which had received a PEV. These results are consistent with the human data, with naïve mice and those given PEV reaching patent parasitemia earlier if they were bitten by mosquitoes with higher number of residual-sporozoites. Again there is a significant association between observed time to patency in mice and model predictions (Fig 2F, p<0.0001). In both the human and rodent system the number of bites and the associated residual-sporozoite loads explains some of the variability in time to patency but not all. This is why the variation seen in observed time to patency is substantially greater than that predicted by the model. The causes of this additional variation are unknown though in the human dataset it could be due in part to the different efficacies of the PEV candidates administered. There is greater variability in the predicted time to patency in the rodent system as the number of times the mice were bitten had a bigger range (between 1 to 10 bites) than the human dataset where all volunteers were bitten by 5 mosquitoes with >10 residual-sporozoites.
Human volunteers bitten by ‘additional’ mosquitoes that had ≤10 residual-sporozoites (in addition to the other 5 bites of >10 residual-sporozoites) had a shorter time to patency (Fig 3). These lightly infected mosquitoes are currently not considered infectious in CHMI trials though they appear to have a significant impact on transmission (S4 Table). Similar results were seen in the mouse data, with each additional bite with ≤10 residual-sporozoites reduced the time to patency in infected mice. In time-to-patency data of both humans and mice the best fit model did not differentiate between bites from mosquitoes with no detectable and 1–10 residual-sporozoites. The occurrence of mosquito bites with ≤10 sporozoites were relatively common in CHMI trials. On average each infected volunteer received 1.6 bites from mosquitoes with ≤10 residual-sporozoites, with 2/3 of the volunteers receiving one or more (Fig 3). The impact of these additional bites depends on the other bites that the volunteer received. For example, if a volunteer was bitten by 5 mosquitoes with 101–1000 salivary gland sporozoites then the additional bites reduced the time-to-patency by on average 1.5 days (Fig 3).
Together the human and rodent data provide strong evidence that mosquitoes with more salivary gland sporozoites post-feeding are more infectious. Though this cannot currently be tested directly in vivo, the most plausible explanation is that that mosquitoes with a higher number of residual-sporozoites injected more sporozoites into the skin during blood-feeding which increased the probability and the speed of the ensuing blood-stage infection. Whilst the probability of infection analysis could only be tested in humans who had been given a PEV, the consistent association between residual-sporozoites and the time to patency in naïve and vaccinated volunteers and the mouse experiments suggest that parasite intensity, and not just the number of infectious bites, is key to understanding the underlying biology of malaria transmission. The mouse experiments indicate a continuum of infection: with every sporozoite injected into the vertebrate increasing transmission.
The relationship between the number of parasites in the salivary gland and the size of the inoculum is poorly understood. Early laboratory work proffer apparently contradictory results [9, 13, 16, 27]. The thin diameter of the mosquito proximal duct means that only one or two sporozoites can pass down it at the same time [13] which supported the hypothesis that inoculum size, if confined to the void volume of the salivary duct, would be independent of sporozoite gland burden. Nevertheless more recent experiments which better resemble natural transmission show a clear positive association between residual-sporozoite load and the number of sporozoites injected into mice [15]. On blood feeding the majority of sporozoites appear to be injected during the first few seconds [13, 27] suggesting their presence in the duct prior to the bite being taken. The number in the ducts will be dependent upon the density of sporozoites in the glands, the time over which they have accrued in the duct since last feeding (on blood or sugar) and the available volume in the duct. There could be additional reasons other than the number of parasites injected which could explain why heavily infected mosquitoes are more infectious. These could include increased (per sporozoite) infectivity [28] or more frequent mosquito probing [29] and further work will be required to determine the causes of the observed infectivity.
Previous studies that failed to find a relationship between residual-sporozoite number and infection used relatively insensitive statistical methods [5, 11] which may have caused them to be underpowered to detect a difference due to the wide variability in the size of the inoculum from similarly infected mosquitoes. The linear increase in the probability of infection on the logarithmic scale (Fig 1E) shows that for every additional residual-sporozoite the increase in transmission declines (i.e. mosquito-mouse transmission is a negative density-dependent process). This has been observed by others [13] and could be modulated by the limited diameter of the duct restricting passage of sporozoites in highly infected mosquitoes. The linear increase in the probability of infection on the logarithmic scale (Fig 1E) shows that for every additional residual-sporozoite the increase in transmission declines (i.e. mosquito-mouse transmission is a negative density-dependent process). This has been observed by others [13] and could be due to the limited diameter of the duct restricting passage of sporozoites in highly infected mosquitoes.
This analysis allows vaccine efficacy estimates to be standardized from mosquito delivered CHMI trials. It is very technically difficult to homogenize the number of salivary gland sporozoites within a mosquito before blood-feeding, making it hard to fully control the infection challenge from CHMI mosquitoes. Including sporozoite intensity information in the analysis of CHMI studies could help reduce between volunteer variability (lowering sample sizes) [21] and improving the characterisation of immune responses [30]. The number of volunteers used in CHMI trials is understandably small. This study shows that failing to account for the (random) variability in parasite challenge will make PEV candidates harder to compare, both within and between studies. The results from the human and murine systems are consistent. In the rodent model a naïve mouse would require 2 bites from mosquitoes with >1000 residual-sporozoites or 4 bites from mosquitoes with 11–100 residual-sporozoites to ensure a 95% probability of infection. This is in line with the human data where all naïve volunteers became infected having received 5 bites from mosquitoes with >10 sporozoites. Naïve humans could not be included in the analysis as the parasite challenge was so high that all became infected (so there was no variability in that dependent variable). Nevertheless, the consistent pattern seen in mice with and without PEV suggests that residual-sporozoite number would have been associated with the probability of infection in naïve volunteers had parasite challenge been lower. By comparison with published estimates on the infectivity of P. berghei sporozoites [31], the observed probability of infection is high suggesting current arrangements for laboratory transmission have increased overall transmission probability.
The epidemiological importance of parasite intensity will depend on the distribution of sporozoites in wild mosquitoes and whether similar trends are observed in people with a prior history of malaria infection. All the volunteers analysed here were malaria naïve and received all infections within minutes of each other. It will therefore be important to repeat the experiments on people from malaria endemic regions who may have acquired a degree of immunity to infection. It will also be necessary to confirm the dose effect in singly bitten humans as it is unlikely that many people in malaria endemic Africa will be bitten by 5 infectious mosquitoes in such a short time period. There are relatively little data on the number of sporozoites in naturally infected mosquitoes [32–35] though laboratory reared and infected A. stephensi have been suggested to inject a similar number of sporozoites to natural infections [9]. In a very low transmission site on the Thai–Myanmar border a recent study reported a geometric mean of 57 sporozoites per mosquito (range 9–11,428) [33]. This compares to earlier studies in Africa and PNG where endemicity was higher and geometric means were >4000 (range 150–10,000) [32, 34]. In these datasets the distribution of sporozoites between mosquitoes appears highly over dispersed (aggregated), with a large number of lightly infected mosquitoes and few with very heavy infections. Notwithstanding the above, more than 45% of infectious mosquitoes caught in a low transmission site in Kenya had >1000 salivary gland sporozoites [35], the same number of residual-sporozoites that was associated with successful infection in the anti-CSP CHMI.
In this study mice were given anti-CSP antibody by passive transfer (i.v), whilst humans were vaccinated with the anti-CSP RTS,S/AS01B. RTS,S is thought to induce T cell responses as well as antibody. Though we see a similar relationship between both anti-CSP interventions disentangling the impact of the different immune responses is beyond the scope of this study. In both the human and mouse data analysed here the anti-CSP interventions were more effective against bites from mosquitoes with a lower residual-sporozoite number. This may explain why recent Phase II and III trials of RTS,S/AS01 (which targets CSP) were partially effective as the vaccine may only be providing sterilizing immunity against lightly infected mosquitoes [36]. Further work is needed to establish whether the anti-CSP antibodies/immunity provided sterilizing immunity against all bites from mosquitoes with a low number of residual-sporozoites (a “threshold” type of protection) or whether it prevents infection from a certain percentage of inoculated sporozoites (a “leaky” type of vaccine (see S1 Fig for a graphical explanation of the two hypotheses). The RTS,S/AS01 Phase III trial was not powered to detect a difference between sites but evidence suggests that it had a higher efficacy against clinical malaria in low transmission settings [37]. Though this is likely to be caused in part by different levels of immunity in the human population variable parasite challenge could also have contributed. It is unknown whether mosquito parasite challenge will diminish as the disease is successfully controlled (and therefore vaccine efficacy might be expected to rise) or whether the sporozoite dose may remain broadly the same [38].
The work shows that lightly infected mosquitoes (with ≤10 residual-sporozoites) contribute to transmission though they have a lower chance of causing onwards infection than more heavily infected mosquitoes. In areas approaching local elimination a large proportion of mosquitoes are thought to be infected by people with low density infections [39]. The contribution of different infection classes to this reservoir of infection has been assessed by their ability to infect mosquitoes [39] (human-mosquito-transmission) but has failed to consider how onwardly-infectious these mosquitoes might be (mosquito-to-human transmission). Work presented here strongly suggests that not all mosquitoes are equally infectious and therefore low density human infections might have a smaller epidemiological impact if they result in less infectious lightly infected mosquitoes [33]. Further work is needed to investigate this hypothesis as it will have consequences for the required sensitivity of malaria diagnostics and the effectiveness of different drug-based control strategies [40].
Malaria transmission intensity in the field is currently measured using estimates of the number of infectious mosquito bites per person per year (the entomological inoculation rate, EIR) [1]. This metric does not explicitly distinguish between light and heavy infections and therefore fails to describe accurately the onwards infectivity of mosquitoes. This will reduce its sensitivity and potentially generate additional bias in different settings [1]. The proportion of infectious mosquitoes in EIR estimates is typically measured by ELISA which is relatively poor at detecting infection in mosquitoes with fewer than 150 sporozoites [41]. The work presented here shows that these infections contribute to transmission thus causing the human force of infection to be underestimated (particularly in areas approaching local elimination [33]). The EIR is central to our understanding of malaria and how control interventions are compared. The influence of the number of parasites on mosquito-to-human transmission could therefore have wide reaching implications across malariology.
All human data analysed here has previously been published elsewhere. See individual publications for specific ethical statements. All animal procedures were performed in accordance with the terms of the UK Animals (Scientific Procedures) Act (PPL 70/7877) and were approved by the Imperial College Animal Welfare and Ethical Review Body (AWERB) LASA guidelines were adhered to at all points. The Office of Laboratory Animal Welfare Assurance for Imperial College covers all Public Health Service supported activities involving live vertebrates in the US (no. A5634-01). Recovery anaesthesia and terminal anaesthesia was performed during the course of this study as per PPL 70/7877. Anaesthesia used was Ketamine/Rompun, and Schedule 1 was performed by overdose, exsanguination or cervical dislocation as part of Home Office approved regulated procedures.”
A full outline of the volunteer selection, procedure and ethical considerations of the humans challenge studies is given by Sheehy et al. [42] whilst the mouse data was generated using methods outlined in Blagborough et al. [24]. Briefly laboratory reared Anopheles stephensi mosquitoes were fed on Plasmodium infected blood and maintained for ~21 days to enable them to develop salivary gland sporozoites. P. falciparum NF54/3D7 strain was used to infect humans, P. berghei clone 2.34 was used in the mouse studies. Mosquitoes were considered to have delivered sporozoites only if they ingested red blood cells in the midgut. Fed mosquitoes were immediately dissected and the number of sporozoites remaining in the salivary glands categorized on a logarithmic scale: 0 (no sporozoites), 1 (1–10), 2 (11–100), 3 (101–1000), 4 (>1000) [9]. In the human trials the feeding procedure was repeated until each volunteer had received 5 bites with >10 residual-sporozoites (in studies carried out before 2004 this value was of score 3 or above, S1 Table). Mice were bitten 1, 2, 3, 4, 5 or 10 times irrespective of the residual-sporozoite score. All bites were received in a short period of time so it is assumed that all mosquitoes contribute equally to the probability of infection and the time to patency (i.e. bite order is not important). Historical CHMI challenge data were collated by returning to the original paper records of each volunteer. Infections where information on the full parasite challenge was unavailable were excluded from the analysis.
The CHMI trials were originally carried out to test different potential malaria PEV candidates (S1 Table). All human control volunteers developed patent infection so could not be used in the probability of infection analysis. Instead data from a single study where 33 volunteers received a three dose regime of RTS,S/AS01B (18) that provided sterilizing immunity to 14 patients (who were subsequently re-challenged 6 months later). There was no significant difference in the impact of the two intervention arms or between initial challenge and re-challenge infections so all data were pooled making a total of 47 experimental infections.
In the mouse study data was collated from population studies to evaluate the impact of compounds inhibiting malarial transmission from mouse-to-mosquitoes which are not present during mosquito-to-mouse transmission (atovaquone [24], primaquine [43], NITD609 [43], artemether + lumefantrine [43] and OZ439 [43]). Other populations of mice were administered the monoclonal antibody 3D11 which targets the same homologous circumsporozoite protein (PbCSP) as the human vaccine RTS,S (PfCSP) which is the first malaria vaccine to be licensed [44]. A number of the human vaccines presented in S1 Table also target PfCSP. Asexual parasite density was regularly measured to determine infection status and time to patency (every day from day 4 to 10 for the mouse model and twice daily from day 6 to 15 and daily up to day 21 in the human volunteers). Volunteers were given a curative anti-malarial treatment on the detection of blood-stage parasites by microscopy. Up to this time-point samples were analyzed by using quantitative real-time PCR (qPCR, analyzing 150μl blood) [42] whilst parasiteamia in mouse samples were quantified by microscopy (by reading 4 microscopy fields, each with approximately 300 red blood cells) [24].
The mean residual-sporozoite score (the average score across all mosquito bites) and the total residual sporozoite score (the sum of the scores across all bites) were calculated for all hosts. To disentangle the effect of the number of bites and the residual-sporozoite scores within those bites a linear-mixed effect model is used to test whether hosts that got infected were bitten by mosquitoes with a higher residual-sporozoite score. The mean and total residual-sporozoite score is taken as the dependent variable whilst a binary value denoting whether or not the host developed infection is included as a fixed effect. The number of bites received is incorporated as a random effect allowing the difference in mean and total residual-sporozoite scores to vary between biting groups. Models with and without the fixed effect were compared using a likelihood ratio test to see whether there was a significantly different residual-sporozoite scores between infected and uninfected hosts.
It is likely that the relationship between the probability of infection and residual-sporozoite score is non-linear so a more advanced binomial statistical model is tested [25]. It is assumed that all bites are independent, i.e. the transmission probability of a bite from a mosquito with a certain residual-sporozoite score is the same irrespective of the number of other bites received. The overall probability of infection is described by a binomial distribution allowing data from multiply bitten hosts to be accurately included (i.e. the probability of infection from two bites each of which have an infection probability of 50% is 75%, not 100%). Sporozoite score is categorised into bins on the logarithmic scale so the probability of infection is estimated for each group independently (i.e. not using a continuous function) as the distribution of parasite numbers within these bins is unknown. Let ϕ denote the probability of a susceptible host becoming infected and bj be the per bite transmission probability (the proportion of bites by an infected mosquito with a residual-sporozoite score j which go onto develop patent infection). The probability of transmission can then be estimated using a binomial distribution,
ϕ=1−∏g=1..q(1−[bS(g)(1−ν)+bS(g)ν (1−εS(g))]),
(1)
where q is the number of bites received and S(g) signifies the residual-sporozoite score in the gth bite (i.e. S(g) = 0,1,2,3 or 4). Parameter v is a binary variable denoting whether the vertebrate host received a PEV or not whilst εj indicates the efficacy of that PEV against a bite with sporozoite score j (the proportional reduction in the probability of infection of a single bite caused by the PEV). Let i denote a group of hosts which receive the same parasite challenge (i.e. the same number of bites with the same residual-sporozoite scores and either a PEV vaccine or not), ni the number of hosts which receive exposure i and xi the number of these that become infected. Using the notation x¯ to denote the corresponding list of observations, under a binomial model the likelihood is proportional to,
L(x¯,n¯|b0,b1,b2,b3,b4,ν)∝∏i=1..zϕxi(1−ϕ)ni−xi,
(2)
where z is the number of combinations of different parasite exposures. This likelihood can be maximised to obtain estimates of the transmission probabilities of mosquitoes with different residual-sporozoite scores from the infection data.
A suite of nested models were fit to the human and mouse data and the most parsimonious were selected using the Akaike information criterion (AIC, lowest value giving the most parsimonious model). These nested models ranged from the full model outlined above (model 6 in S3 Table where mosquitoes with each of the different residual-sporozoite scores has a different contribution to transmission) down to models where all sporozoites positive scores were pooled together (i.e. transmission was independent of sporozoite load but dependent on the number of mosquito bites, model 2) or infection is independent of the number of bites (model 1). Residual-sporozoite scores of zero were also included to determine whether they had an impact on transmission. To test whether a bite with a certain score (or a range of scores) had a significant impact on the transmission metrics each model was systematically reduced to determine whether setting the contribution of a bite to zero improved the parsimony of the model. The full range of different models investigated is outlined in S3 Table. The impact of the PEV was subsequently assessed by initially assuming that vaccination had a different impact on the probability of infection of each sporozoite score before the model was systematically collapsed, grouping the impact of the vaccine on adjacent sporozoite scores together, until the most parsimonious combination was found. Multivariate 95% confidence intervals for each of the parameters were generated from the likelihood profile using a likelihood ratio test. Residual-sporozoite scores (or groups of) whose 95% confidence intervals spanned 0 were further collapsed to determine whether setting them to zero improved the parsimony of the model. To investigate the full impact of (apparently) uninfected bites the datasets were reduced, removing all hosts that only received bites with a residual-sporozoite score of zero. The best fit model was used to predict the likelihood of infection for each experimental infection according to the number of bites they received, the number of residual-sporozoites within those bites and whether or not they received a vaccine candidate. The association between these predicted values and the observed infectious status was evaluated using simple logistic regression (with null and full models compared using a likelihood ratio test).
The pre-patent period is typically defined as the first time-point at which infected red blood cells are detected. In this study it is estimated in humans using qPCR [45, 46] and microscopy in mice [24]. The appearance of the first detectable parasite is relatively variable as some PCR runs generate false positive (low) values [47] and the number of parasites in the sample will vary by chance at low densities [48]. This is shown in the dataset by high heteroscedacity when linear random effects models are fit to the individual parasite growth rate curves [47]. Instead studies measure a time to a low parasite density threshold as this reduces sampling variability and the influence of false positive results. Here a density of 1000 parasites per μl for humans and 2% infected red blood cells for mice are used though the same qualitative results are seen with different thresholds. Therefore throughout the manuscript the time to patency specifically refers to the time to the threshold parasite density.
The influence of residual-sporozoite score on the time to patency is analysed using a semi-parametric additive hazard model [49] implemented in the “timereg” R package [50]. This type of survival analysis enables the covariates to act additively on the baseline hazard instead of multiplicatively (as done in standard proportional hazard model) allowing the number of bites to act on an absolute scale rather than a relative one [50]. As there is a minimum time between parasite challenge and patency the baseline hazard is allowed to vary over time whilst the covariates (the number of bites with a score of 0,1,2,3 or 4 and whether the host was given a vaccine) are assumed to be time invariant. In the human dataset a variety of different vaccine candidates were tested. For simplicity it is assumed that all vaccines have an equal impact on the time to patency as the number of volunteers for each vaccine type is very small (typically 3–8 volunteers per candidate, S1 Table). Parasite challenges which did not result in infection were removed from the analysis to ensure that the influence of sporozoite number on the time to patency is independent of whether the host became infected. Infected hosts who had not reached the parasite density threshold (42 out of 429 mice) were classified as being censored. The mean time to patency was estimated by integrating over the survivorship function and the association between model derived prediction and the observed time to patency was tested using simple linear regression.
Likelihood estimates are not available for the additive hazard model so the significance of the different covariates (from zero) is estimated by resampling [50]. Without more formal model selection procedure the full model is collapsed (using the model combinations presented in S3 Table) until all covariates are significantly different from zero (p values of <0.05 were significant). For example, if the model collapses to model 2 then this suggests that the time to patency is independent of residual-sporozoite score but is significantly associated with the number of mosquito bites. For some mice only a binary outcome for infection was available instead of a parasite density estimate. To enable the whole dataset to be utilized a dummy variable was added denoting whether parasite density estimates were measured or not, providing an estimate of the time between the first detection and 2% red blood cell infection. Point estimates of the 95% confidence intervals were generated by resampling [50]. Raw data used in the mouse work is given in S1 Dataset.
The size of the Liver-to-Blood parasite Inoculum (LBI) can be estimated using linear regression from the parasite growth rate in the vertebrate host [47]. LBI estimates has been used to disentangle the impact of vaccine in CHMI trials and may be associated with residual-sporozoites number. Unfortunately its estimation is beyond the scope of this work as it requires accurate measures of asexual parasite density over multiple timepoints. This information was also unavailable for the mouse dataset as parasitemia was estimated by microscopy (which has high measurement error, particularly at low densities) and because relatively few data points are available (on average <3 per host above the density threshold). In the human data information from different vaccines is pooled as each study is relatively small. This would make LBI estimates highly uncertain due to differences in the parasite multiplication and variability in PCR heteroscedacity between studies [47].
|
10.1371/journal.pbio.0050133 | Gamma Oscillations in Human Primary Somatosensory Cortex Reflect Pain Perception | Successful behavior requires selection and preferred processing of relevant sensory information. The cortical representation of relevant sensory information has been related to neuronal oscillations in the gamma frequency band. Pain is of invariably high behavioral relevance and, thus, nociceptive stimuli receive preferred processing. Here, by using magnetoencephalography, we show that selective nociceptive stimuli induce gamma oscillations between 60 and 95 Hz in primary somatosensory cortex. Amplitudes of pain-induced gamma oscillations vary with objective stimulus intensity and subjective pain intensity. However, around pain threshold, perceived stimuli yielded stronger gamma oscillations than unperceived stimuli of equal stimulus intensity. These results show that pain induces gamma oscillations in primary somatosensory cortex that are particularly related to the subjective perception of pain. Our findings support the hypothesis that gamma oscillations are related to the internal representation of behaviorally relevant stimuli that should receive preferred processing.
| Pain is a highly subjective sensation of inherent behavioral importance and is therefore expected to receive enhanced processing in relevant brain regions. We show that painful stimuli induce high-frequency oscillations in the electrical activity of the human primary somatosensory cortex. Amplitudes of these pain-induced gamma oscillations were more closely related to the subjective perception of pain than to the objective stimulus attributes. They correlated with participants' ratings of pain and were stronger for laser stimuli that caused pain, compared with the same stimuli when no pain was perceived. These findings indicate that gamma oscillations may represent an important mechanism for processing behaviorally relevant sensory information.
| Within the continuous flow of sensory information, a huge number of events compete for neural representation and perception. This sensory overflow requires the selection and preferential processing of relevant information in order to optimize the utilization of cerebral processing resources. Recently, induced neuronal oscillations in the gamma frequency range (about 40–100 Hz) have been suggested to represent one mechanism of the selection and preferred processing of sensory information [1–7]. These induced gamma oscillations represent event-related modulations of neuronal oscillations, are often observed in early sensory cortices and differ from evoked neuronal responses in a lack of phase locking to the sensory stimulus. Functionally, the association between induced gamma oscillations, and selection and preferred processing of sensory stimuli suggests that these responses may not only be related to the physical stimulus attributes, but also related particularly to the subjectively weighted percept of a sensory event.
Painful stimuli signal threats and are therefore of utmost behavioral relevance [8,9]. Thus, we hypothesized that painful stimuli induce gamma oscillations in somatosensory cortices. Moreover, we speculated that these pain-induced gamma oscillations may not only relate to the objective attributes of painful stimuli, but may also particularly reflect the subjective experience of pain. To address this issue, we used magnetoencephalography to record neural responses to noxious stimuli in healthy human subjects. We investigated the effects of noxious stimuli on neuronal activity in the gamma band and related these effects to objective stimulus intensity and subjectively perceived pain intensity. Our results show that pain induces gamma oscillations in the contralateral primary somatosensory cortex. Amplitudes of pain-induced gamma oscillations increase with objective stimulus intensity and subjective pain intensity. However, around pain threshold, perceived stimuli induce significantly stronger gamma oscillations than unperceived stimuli of equal stimulus intensity. These observations provide direct evidence for a close association between induced gamma oscillations and the conscious and subjective perception of behaviorally relevant sensory events.
First, we aimed to identify and to characterize spatially and temporally pain-induced gamma oscillations in human somatosensory cortices. In 12 healthy male participants, 40 moderately painful cutaneous laser stimuli (intensity 600 mJ) were applied to the dorsum of the right hand. Participants were instructed to passively perceive the stimuli without any further task. The contralateral primary (S1) and bilateral secondary (S2) somatosensory cortices were localized by analyzing the well-known [10] pain-evoked (phase-locked) responses from these areas (Figure 1A). Next, we investigated possible pain-induced gamma oscillations in these areas. To this end, time-frequency representations (TFRs) were calculated for each trial and area. The analysis revealed that pain induces strong and significant increases in gamma power at frequencies between 60 and 95 Hz in the contralateral S1 cortex (Figure 1B). These pain-induced gamma oscillations were observed between 100 ms and 300 ms after stimulus application coinciding with the pain-evoked response from S1 (Figure 1C). Please note that these pain-induced gamma oscillations were observed without any particular task and, thus, do not depend on the task relevance of painful stimuli, but rather on their sensory quality and their inherent behavioral relevance. No pain-induced changes in gamma power were observed in the bilateral S2 cortices. Analysis of amplitude and phase dynamics confirmed that gamma oscillations were not phase locked to stimuli, and therefore represent induced, but not evoked oscillations (Figure 2).
Second, we investigated the relationship between amplitudes of induced gamma oscillations in S1, stimulus intensity, and perceived pain intensity. To this end, randomly varied intensities of noxious laser stimuli were applied to the right hand of 13 healthy human participants. Possible stimulus intensities were 150, 300, 450, and 600 mJ, which yield sensations ranging from barely detectable to moderately painful. Forty stimuli were presented for each stimulus intensity, and subjects were asked to rate the perception of each stimulus on a scale from 0 to 100 anchored at “no pain” and “worst imaginable pain.” Pain ratings were cued by an auditory signal presented 3 s after each stimulus. The contralateral S1 cortex was localized, and evoked responses and induced gamma oscillations to stimuli of different intensities were analyzed. Mean amplitudes of evoked responses and induced gamma oscillations from S1 were calculated during the time window from 100 ms to 300 ms as compared to baseline amplitudes. Figure 3 shows amplitudes of evoked (phase-locked) responses and induced (non–phase-locked) gamma oscillations as a function of stimulus intensity. The results reveal that amplitudes of induced gamma oscillations and amplitudes of evoked responses from S1 increase with stimulus intensity. This increase in response amplitudes was paralleled by an increase in perceived pain intensity. These observations show that amplitudes of pain-induced gamma oscillations in S1 vary with objective stimulus intensity and subjective pain intensity.
Third, we aimed at further defining the relationship between pain-induced gamma oscillations in S1 and the subjective experience of pain. Low intensity (150 and 300 mJ) trials around pain threshold were chosen for the analysis. Per definition, some of these trials are perceived as painful (“percept” trials, rating > 0) and some are not (“no percept” trials, rating = 0). We matched both trials for stimulus intensity and number of stimuli, i.e., for each individual, the “percept” and “no percept” sets of trials did not differ with respect to objective stimulus intensity or number of stimuli, but only with respect to subjective perception. Mean amplitudes of gamma oscillations at 100 ms to 300 ms as compared to baseline were computed for “percept” and “no percept” trials. The analysis reveals that gamma oscillations were significantly stronger for “percept” than for “no percept” trials (p = 0.013, Wilcoxon signed-rank test, Figure 4). Gamma power during baseline did not differ between conditions (p = 0.15) and only “percept” trials (p = 0.003), not “no percept” trials (p = 0.735), yielded significant increases of gamma power as compared to baseline. In contrast, amplitudes of phase-locked, evoked S1 responses did not differ between “percept” and “no percept” trials (p = 0.38). These observations show that pain-induced gamma oscillations are particularly related to the subjective perception of pain.
Our study demonstrates that painful stimuli induce gamma oscillations in the contralateral S1 cortex. These non–phase-locked gamma oscillations differ from evoked responses in a trial-by-trial jitter in latency, and occur at latencies around 200 ms and at frequencies between 60 and 95 Hz. Amplitudes of pain-induced gamma oscillations increase with both objective stimulus intensity and subjective pain intensity. However, around pain threshold, differences in the subjective perception of objectively similar stimuli were related to differences in amplitudes of induced gamma oscillations. These results show that pain-induced gamma oscillations in S1 are particularly related to the subjective perception of pain.
Here, latencies of pain-induced gamma oscillations between 100 ms and 300 ms indicate that these responses are mediated by A-delta-fibers relating to first pain sensation. Later sensations of warmth or second pain mediated by slowly conducting C-fibers occur at latencies of about 1,000 ms [11], and are unlikely to relate to cortical responses at latencies around 200 ms. Correspondingly, in the present study, “percept” and “no percept” refer to the presence and absence of A-delta-fiber–mediated first pain sensation. Moreover, our findings provide evidence for first pain–related gamma oscillations in S1, but do not preclude gamma oscillations from outside the somatosensory cortices, which were beyond the scope of our analysis.
Induced gamma oscillations have been demonstrated in tasks that require activation and further processing of object representations [12,13]. As common features across sensory modalities, these induced gamma oscillations share a similar timing (around 200 ms to 400 ms) and a focal localization [12,14–16] as compared to low-frequency components. A few previous studies investigated pain-related changes in gamma oscillations [17–19]. However, these studies did not apply selective nociceptive stimuli and could not provide consistent results on location, timing, and functional characteristics of possible pain-induced gamma responses. Other studies investigating the relationship between S1 and S2 responses and pain intensity [20–25] revealed a positive correlation between response amplitudes and pain intensity. However, these studies did not investigate induced but evoked responses or blood oxygen level-dependent (BOLD) effects and, thus, cannot be directly compared to the present results.
Painful stimuli possess utmost behavioral relevance that invariably interrupts ongoing processes, demand full access to system resources, and thereby lead to preferred processing [8,9]. Our study demonstrates that induced gamma oscillations are particularly related to the subjectively weighted percept of noxious stimuli. Thus, our results provide an important link between gamma oscillations and a neural filtering mechanism selecting behaviorally relevant information for action. Indeed, neural oscillations as regular excitability changes of neuronal populations have the capability of gating information flow and adding relevance to spike trains [4,26–28]. Thus, the gamma oscillations observed in the present study may modulate the cerebral processing of painful stimuli and the perceptual quality of the stimulus rather than representing the neural substrate of perception per se. Pain-induced gamma oscillations may thereby participate in activating the sensory representation of a painful stimulus and its selection for further processing. Preferred processing of these stimuli may facilitate behavioral responses aimed at preserving the integrity of the individual.
Beyond, our findings suggest that pain-induced gamma oscillations in S1 are related to a complex cerebral network subserving conscious perception of sensory events. This network includes sensory areas as well as higher order frontal and parietal association cortices [1,29–34]. Within this network, perception appears to depend on the complex relationship between ongoing neuronal activity [30,35], phase-locked evoked responses [30,34], and non-phase-locked–induced [1,32] responses.
Taken together, the present findings show that noxious stimuli induce gamma oscillations in S1 that are particularly related to the subjective experience of pain. These observations are compatible with the hypothesis that induced gamma oscillations are related to the internal representation of behaviorally relevant stimuli which should receive preferred processing. Our findings may, thus, contribute to our understanding of the neural mechanisms of perception and, in particular, to the understanding of the highly subjective experience of pain in health and disease.
Twelve healthy male participants (mean age: 33 y, range 22–41 y) participated in the experiment. All participants gave informed consent, and the study was performed according to the Declaration of Helsinki with the local ethics committee's approval.
Subjects were comfortably seated in a reclining chair. Forty noxious cutaneous laser stimuli were delivered to the dorsum of the right hand, and subjects were instructed to passively perceive the stimuli with closed eyes [36]. Stimuli were cutaneous laser stimuli that selectively activate nociceptive afferents without concomitant activation of tactile afferents. The laser device was a Tm:YAG-laser (Carl Baasel Lasertechnik, Starnberg, Germany) with a wavelength of 2,000 nm, a pulse duration of 1 ms, and a spot diameter of 6 mm. An optical fiber transmitted the laser beam into the magnetically shielded recording room. Stimulation site was slightly changed after each stimulus. Interstimulus intervals were randomly varied between 10 and 14 s. Applied stimulus intensity was 600 mJ, which evoked moderately painful sensations.
Neural activity was recorded with a Neuromag-122 whole-head neuromagnetometer [37] with passbands of 0.03–170 Hz and digitized with 514 Hz. The exact position of the head with respect to the sensor array was determined by measuring magnetic signals from four coils placed on the scalp. High-resolution T1-weighted magnetic resonance images (MRI) were obtained for each subject. Anatomical landmarks (nasion and preauricular points) were localized in each individual and used for the alignment of the MRI- and magnetoencephalography (MEG) coordinate systems.
Thirteen healthy, right-handed men participated (mean age: 28 y, range 25–33 y) in experiment 2. In this experiment, stimulus strength varied randomly with possible values of 150, 300, 450, or 600 mJ. Forty stimuli for each intensity were applied. Three seconds after each laser stimulus, an auditory signal prompted the participants to rate the intensity of the initial “pinprick”-like first pain on a rating scale from 0–100. Zero was defined as “no pain,” and 100 was defined as the “worst imaginable pain.” Because the rating was explicitly focused on first pain, a rating of zero did not preclude later sensations of warmth or second pain. The sample rate was 483 Hz, and signals were band-pass filtered between 0.03 and 160 Hz [24]. The other parameters were the same as in experiment 1.
Recorded signals were high-pass filtered (1 Hz) and visually inspected for artifacts. Contaminated epochs were excluded, leaving a minimum of 36 trials per participants and stimulus intensity.
First, somatosensory cortices were localized by analyzing the well-known pain-evoked responses from these areas [10]. Somatosensory cortices were selected for the analysis because previous studies showed that induced gamma oscillations mainly occur in early sensory cortices [2,7,12,14,16]. To this end, covariance matrices across all sensors were calculated for a prestimulus baseline interval (−400 to 0 ms) and a poststimulus interval (0 to 400 ms) including strongest pain-evoked activity as evident in global field power. From these covariance matrices, neural activity during both intervals was localized by using a spatial filtering algorithm [38–40]. The spatial filter was used with a realistic head model to estimate power in the whole brain, resulting in individual tomographic power maps with voxel sizes of 6 × 6 × 6 mm. For each voxel, the ratio of poststimulus power to prestimulus power was computed resulting in individual functional tomographic power maps that show cortical areas with a strong increase of neural activity following noxious stimuli. These functional maps were individually normalized to one, spatially normalized using SPM2 (Wellcome Department of Cognitive Neurology, Institute of Neurology, London, United Kingdom: http://www.fil.ion.ucl.ac.uk/spm), thresholded at 0.8 of the individual maximum, and averaged across participants. This procedure yields group-mean tomographic maps of pain-evoked power increases with a dimensionless maximum value of approximately 0.4. An arbitrary threshold was used for visualization because only the local maxima, but not the extent of activations, were used for subsequent analysis. The Analysis of Functional Neuroimages/AFNI Surface Mapper (AFNI/SUMA) programs were used for surface rendering (National Institute of Mental Health, Bethesda, Maryland, United States: http://afni.nimh.nih.gov/afni). As in previous MEG studies [10], strongest pain-evoked activity was seen in contralateral S1 and bilateral S2. Note that this analysis aimed at localizing somatosensory cortices and does not preclude activations of additional areas, e.g., insular or cingulate cortex, which may yield lower signal-to-noise ratios due to their deep location and/or radial orientation barely detected by MEG.
Second, individual locations of evoked S1 and S2 responses were optimized beyond the 6-mm grid of the first analysis step. To this end, a multi-dimensional, constrained nonlinear minimization (Nelder-Mead, modified fminsearch function in Matlab, Mathworks: http://www.mathworks.com) was employed. The S1 and S2 maxima from the individual tomographic power maps were used as starting points for the optimization. The position was allowed to change by 1 cm in each direction, whereas orientation was constrained to be tangential to the center of the head. For each position, the ratio of poststimulus activity (0 to 400 ms) to prestimulus activity (−400 to 0 ms) was calculated, and the position with the maximum stimulus-evoked power increase was chosen for further analysis. The same optimization procedure was applied to induced gamma power in S1, revealing that optimized locations of evoked S1 responses and induced gamma oscillations in S1 did not differ (x-coordinates, p = 0.15; y-coordinates, p = 0.10; z-coordinates, p = 0.84; two-tailed Wilcoxon signed-rank tests).
Third, for optimized locations of S1 and S2, time courses of activity were computed for all single trials, using the adaptive spatial filter [38,39]. Note that time courses of all activations were analyzed in source space. These time courses were subjected to a time-frequency analysis based on multi-tapers [41] using the fieldtrip toolbox (F. C. Donders Centre for Cognitive Neuroimaging: http://www.ru.nl/fcdonders/fieldtrip). A multi-taper–based analysis was chosen since this approach provides a robust and optimal way to smooth spectra in the frequency domain, and thereby enhances higher frequency oscillatory components with large-frequency jitter-like induced gamma oscillations. The analysis yields TFRs showing power as a function of time and frequency. TFRs were computed from 30 to 100 Hz in 400-ms–long windows with a spacing of 20 ms between windows. A 400-ms time window was chosen to allow a multi-taper frequency smoothing of ±5 Hz. For each frequency, relative change to a 1,000-ms baseline was computed. Power was coded as z-scores calculated from the 1,000-ms baseline. Significance of differences between poststimulus and prestimulus activity in TFRs was determined by applying permutation statistics. To this end, the 12 prestimulus baseline (−1,000 to 0 ms) and the 12 poststimulus (0 to 1,000 ms) parts of the TFRs of the 12 different subjects were randomly permuted 5,000 times. Each time, the maximum difference was computed across time and frequency. The 95th percentile of all 5,000 maximum differences was taken as threshold for the TFRs. This maximum statistics takes multiple comparisons into account [42].
Fourth, in order to distinguish between phase-locked (evoked) and non–phase-locked (induced) neural responses, phase locking of stimulus-related neural activity was determined. For each cortical area, single trial time courses were bandpass filtered (forward and reverse with a fourth-order Butterworth filter) in the gamma frequency band (60–95 Hz) defined from TFRs. The Hilbert transformation yielded instantaneous phase and amplitude estimates for each single trial with an optimum temporal resolution. Phase-locking value (plv, bounded between zero and one) was computed as the absolute value of the mean of complex phase Φ across N trials.
Evoked components show a consistent phase relationship to stimulus that is evident in a high plv. Amplitude and phase dynamics were calculated with reference to a 1,000-ms prestimulus baseline. To establish a confidence level for stimulus-induced phase locking, the time course for each region of interest was randomly permuted 5,000 times and then subjected to the same phase-locking analysis (i.e., filtering, Hilbert transformation, averaging, and baseline correction). The maximum value was used as the confidence level.
|
10.1371/journal.pntd.0003706 | Community-Centered Responses to Ebola in Urban Liberia: The View from Below | The West African Ebola epidemic has demonstrated that the existing range of medical and epidemiological responses to emerging disease outbreaks is insufficient, especially in post-conflict contexts with exceedingly poor healthcare infrastructures. In this context, community-based responses have proven vital for containing Ebola virus disease (EVD) and shifting the epidemic curve. Despite a surge in interest in local innovations that effectively contained the epidemic, the mechanisms for community-based response remain unclear. This study provides baseline information on community-based epidemic control priorities and identifies innovative local strategies for containing EVD in Liberia.
This study was conducted in September 2014 in 15 communities in Monrovia and Montserrado County, Liberia – one of the epicenters of the Ebola outbreak. Findings from 15 focus group discussions with 386 community leaders identified strategies being undertaken and recommendations for what a community-based response to Ebola should look like under then-existing conditions. Data were collected on the following topics: prevention, surveillance, care-giving, community-based treatment and support, networks and hotlines, response teams, Ebola treatment units (ETUs) and hospitals, the management of corpses, quarantine and isolation, orphans, memorialization, and the need for community-based training and education. Findings have been presented as community-based strategies and recommendations for (1) prevention, (2) treatment and response, and (3) community sequelae and recovery. Several models for community-based management of the current Ebola outbreak were proposed. Additional findings indicate positive attitudes towards early Ebola survivors, and the need for community-based psychosocial support.
Local communities’ strategies and recommendations give insight into how urban Liberian communities contained the EVD outbreak while navigating the systemic failures of the initial state and international response. Communities in urban Liberia adapted to the epidemic using multiple coping strategies. In the absence of health, infrastructural and material supports, local people engaged in self-reliance in order to contain the epidemic at the micro-social level. These innovations were regarded as necessary, but as less desirable than a well-supported health-systems based response; and were seen as involving considerable individual, social, and public health costs, including heightened vulnerability to infection.
| In this study the authors analyzed data from the 2014 Ebola outbreak in Monrovia and Montserrado County, Liberia. The data were collected for the purposes of program design and evaluation by the World Health Organization (WHO) and the Government of Liberia (GOL), in order to identify: (1) local knowledge about EVD, (2) local responses to the outbreak, and (3) community-based innovations to contain the virus. At the time of data collection, the international Ebola response had little insight into how much local Liberian communities knew about Ebola, and how communities managed the epidemic when they could not get access to care due to widespread hospital and clinic closures. Methods included 15 focus group discussions with community leaders from areas with active Ebola cases. Participants were asked about best practices and what they were currently doing to manage EVD in their respective communities, with the goal of developing conceptual models of local responses informed by local narratives. Findings reveal that communities responded to the outbreak in numerous ways that both supported and discouraged formal efforts to contain the spread of the disease. This research will inform global health policy for both this, and future, epidemic and pandemic responses.
| The West African Ebola epidemic emerged in the forest region of Guinea in late December 2013 and appeared to be contained until May 2014, when it rapidly accelerated its rate of incidence and crossed into urban areas in Sierra Leone and Liberia [1]. Upon entering Sierra Leone and Liberia, the rate of Ebola virus disease (EVD) transmission increased rapidly resulting in 1,711 cases by August 8, 2014, when the WHO declared that the conditions for a Public Health Emergency of International Concern (PHEIC) had been met—the third announcement of its kind in history. Although the tide has since turned, as of January 2015 there were over 21,000 confirmed, probable, or suspected cases of EVD, and more than 8,600 Ebola-related deaths, 3605 of which occurred in Liberia alone [2].
The West African Ebola epidemic has demonstrated that the existing range of medical and epidemiological responses to emergency epidemics is insufficient in some particularly vulnerable post-conflict low-income countries like Liberia and Sierra Leone, while other low-income countries like Senegal and Nigeria are able to rapidly contain outbreaks with international support. This fact has provoked key discussions regarding the need to strengthen African health systems, redress exceedingly poor public health and healthcare infrastructures [3], and examine the capability of local communities to respond to global health crises.
A key lesson from the West African Ebola epidemic is that local community engagement is crucial for response, and may have played a role in the decline in transmission rates [4–7]. In a context in which health surveillance systems had failed, healthcare workers were experiencing disproportionately high mortality rates due to Ebola infection, clinics and hospitals across Liberia were shut down, and the construction of hospitals and Ebola Treatments Units (ETU’s) could not keep pace with demand [8], communities were compelled to generate solutions of their own. The social context for local innovations and response did not appear to be auspicious, as local communities exhibited resistance towards hospitals, ETUs, and mandatory cremation policies [9], and global health experts expressed concern regarding local communities’ abilities to adequately contain the highly infectious disease and treat the sick.
Declining transmission rates are likely to be attributed to multiple factors. Mathematical modeling, having featured prominently in the Ebola response, has been used to understand transmission dynamics as well as the significance of different control interventions [8, 10]. For example, a recent agent-based model found that Ebola treatment centers (ETCs), safe burials, and protection kits played a defining role in case reductions in Liberia [11]. However models are invariably simplifications of complex realities. In a recent commentary, Chowell and Viboud [12] descripted the tendency to simplify mobility patterns, social behavior, and differences between urban and rural settings. Others have discussed how psychological and socio-cultural factors remain unaccounted for in most Ebola transmission and control models [13]. Local mitigation strategies undertaken by communities themselves often remain unaccounted for in causal explanations.
Modeling the effectiveness of specific Ebola-related interventions requires attention to local community and household-based grassroots efforts. Epidemiologists are increasingly recognizing the role of local social networks in controlling infectious disease outbreaks outside of formal public health and biomedical interventions, and they are attempting to account for them in complex models [14–15]. Reports from news outlets, however, anticipated this response, suggesting well in advance of epidemiologists that declining rates of Ebola were due, in large part, to local acceptance of safe burials and the mobilization of communities to isolate and refer infected individuals to Ebola response teams, ETUs, and community-care centers (CCCs). For example, one case-finder in West Point, Monrovia’s largest slum, stated “The virus is in the community, and the best way to take it from the community is for the community itself to take charge” [16]. Joel Montgomery, a CDC team leader in Liberia also noticed that, “Communities are doing things on their own, with or without our support” [17].
Using data collection methods informed by participatory rural appraisal models [18–20], this study shows how community-based responses both supported and discouraged formal efforts to manage the epidemic; and how community-based informal responses may have contributed to containing the outbreak. Specifically, we provide insight into community leaders’ tactics and strategies for managing the presence of Ebola in their communities, and we analyzed these tactics and strategies in order to develop conceptual models of local responses informed by local narratives and descriptions. When reviewed using an inductive approach informed by ethnographic contextualization (see methods section below), the data revealed how communities were engaging in self-reliance—in the absence of health, infrastructural, and material supports—in order to contain the epidemic at the micro-social level.
Community leaders described how they proposed to engage in prevention efforts through training and awareness, hygiene, surveillance, and the creation of local infrastructures; how they proposed to conduct treatment and response through a process of isolation, quarantine, and triage; and how they proposed to manage the sequelae of the presence of Ebola, especially among orphans, survivors (people who had been ill but had recovered), and for memorialization. They also identified critical epidemic-related and long-term structural barriers inhibiting the utilization of public health and medical infrastructures. Importantly, these data are analyzed in this article to present how local caregiving efforts served as both pathways for disease containment and as socially-mediated pathways for potential new infections.
Our findings demonstrate how communities showed resilience, innovation, and rapid response to the Ebola crisis. They also show how under conditions of extreme stress, culture can be flexible and supple in response to extreme circumstances and the arrival of new information (like public health messages), and make allowances in extraordinary conditions [21]. It contributes to a small but growing literature on local understandings and responses to hemorrhagic fever outbreaks [22–26] that challenges conventional thinking about the role that “culture” plays in epidemics. As Hewlett has demonstrated in Ebola outbreaks in the Democratic Republic of the Congo, Uganda, and Gabon [22, 26], and this research demonstrates for Liberia, culture is not a fixed entity. Local knowledge can shift rapidly in response to public health information and local epidemiological realities [27]. Funerary and caregiving practices can be suspended or altered [28]. Families will break with convention to protect uninfected individuals.
The attributes of culture that public health experts often attend to in the Ebola outbreak (like hygiene practices, food practices, and death rituals) may not been the most important factors in the cultural principles that shaped community-centered Ebola responses. Attention needs to shift to the culture of caregiving that exists in Ebola-affected cities and towns. We need to better understand how strong and dense the emotional ties that bind families and communities together are and can be, and precisely how Ebola, and the failed Ebola response, is doing violence to those social ties [29–33].
The reports of community leaders presented below are based upon hypothetical questions and conditional inquiries about best practices for containing Ebola, and what they would do if Ebola emerged in their communities. The community leaders’ responses, however, emerged directly from their lives and plans as they attempted to prevent or contain the epidemics in and around their own Monrovia communities and broader social networks. Therefore, the data reported here are indicative of transient emergent plans then in circulation in Monrovia communities during the height of the outbreak.
These data were collected as part of a Government of Liberia/World Health Organization GOL/WHO rapid assessment of community leaders’ perceptions of appropriate management practices for addressing the incidence of Ebola in their communities. The research teams were trained and directed by an applied medical anthropologist and conducted data collection from September 1-20th, 2014 in 15 communities of varying economic, ethnic, and population characteristics in Monrovia and Montserrado County, Liberia. Data are drawn from focus groups, qualitative field notes, and supporting literatures. Liberian research teams conducted 15 focus groups, one in each community, consisting of 15–20 people of mixed gender, for a total of 368 participants. All participants were community leaders, drawn from women’s groups, youth groups, local zonal heads, political groups, clinics, church-based organizations, non-governmental organizations, and recreational clubs. The tone of the meetings was widely reported as cooperative and participants were positively engaged.
Field-level ethnographic contextualization was provided through the fieldnotes and qualitative observations of the lead expatriate anthropologist directing the eight-person Liberian research team and the eight Liberian researchers themselves. Field-level observations were based on community-based participant observation, informal interviews, and discussions with various agencies in the Ebola response, as well as focus group data. An additional layer of ethnographic contextualization was provided by the lead author, who has nine years of ethnographic field experience in Monrovia, Liberia [34]. A team of public health and anthropological researchers at the University of Florida and at Yale University analyzed, coded, and thematically clustered de-identified focus group data in October 2014. The inductively derived social structural models (see Figs 1–4) emerge from focus group findings and were informed by the broader literature on the current Ebola outbreak.
Because the data were collected anonymously for the purposes of program evaluation (as part of the World Health Organization’s Ebola response activities) and the authors analyzed data that was de-identified and redacted, the study received an expedited review and exemption under the University of Florida Institutional Review Board (IRB) for the Protection of Human Subjects (IRB-02) #2014-U-1117. In accordance with the protocol detailed during a WHO research ethics review of the proposed data collection initiative, all research participants provided verbal informed consent and their names were not collected with the data. Participants were compensated for their time with US $5.00, and were informed that there would be no other direct benefit for participation in the study.
Community leaders shared with the research team their opinions regarding “best practices” concerning local community responses to the Ebola outbreak. Using a grounded theory approach, their feedback was analyzed for common themes, which generated an “ideal-type” [35] framework—a synthesis of commonly agreed-upon elements—for community-based response to the presence of Ebola in and near urban Liberian communities (Fig 1). This included three sequential phases of action and response: prevention, response and treatment, and sequelae. Both the findings and the figures represented in this section detail community responses, and suggest implications that concatenate with the existing literature on social structure, gender roles, healthcare capacity, and conflict histories in the region [34, 36–40].
Community leaders agreed that prevention was the best strategy to curtail the Ebola outbreak. During the Ebola outbreak in Monrovia, prevention was seen as including six key areas: (1) a sharp increase in the quantity and specificity of community-based training to prevent Ebola infection, (2) improved hygiene, sanitation, and the distribution of cleaning and protective materials, (3) creating a system of surveillance, (4) safely transporting infected individuals from the community into hospitals and Ebola Treatment Units (ETUs), (5) removal of the dead, and (6) establishing a community-based infrastructure to care for people who were sick with Ebola.
Training, said some, would counter the presence of fear in the communities and change the minds of those who continued to deny the existence of Ebola. Community leaders sought training to address the following key technical challenges in Ebola management:
Notably, they were not seeking the basic information about Ebola then offered during health communications campaigns (ex. “What is Ebola? Have you ever heard of Ebola?”). Community leaders felt that they had a strong grasp on the basics of Ebola etiology and transmission, and the challenges they confronted or anticipated pertained to their uncertainty about how to respond in the context of health sector collapse. They called for training methods that would make public health messages more palatable and effective (by using local languages, video, door-to-door education, or billboards), but their concern was that people believed in Ebola and knew enough to be afraid, but not enough to respond effectively in the absence of a functioning Ebola medical response system or general healthcare sector. As one person reported, “We have heard the messages, but most people do not know how to practicalize them.” Continuity of messages and continuity of message delivery was highlighted as a critical issue. Community leaders emphasized the need to engage with communities on a daily basis, and for the messages they were relating to be correct and similar. Proposed training models included youth awareness workshops, community-based volunteer/training-of-trainer (ToT) workshops, and trainings for community leaders on the management of community-based Ebola responses.
To prevent the incidence of Ebola in their communities, community leaders argued that heightened attention to sanitation would prohibit the spread of the virus through bodily fluids. Governmental, NGO, and bilateral support was requested to sponsor heightened sanitation in private and public latrines, to include the distribution of buckets, ash, and bleach for washing, and locally obtainable personal protective equipment (PPE) gear like raincoats, rain boots, and plastic bags. As some demanded, “we need the same PPEs given to the medical doctors and nurses to be given to the community.”
Sanitation alone would not get the job done, however. In order to protect the community from the introduction of Ebola, and deter the spread of infection within neighborhoods, community leaders called for heightened surveillance efforts, and reported those in which they were already engaged. There was a strong community-based ethos informing control measures. As one community member said, “as a community we keep watch over each other.”
Community leaders’ accounts of optimal community-based Ebola surveillance included a four-tiered system of surveillance that was designed to prevent introduction of the virus into the community, facilitate reporting, disseminate information, sustain house-to-house monitoring, and support whistle blowing when community-members were non-compliant with EVD prevention protocols (Fig 2). The first level of surveillance involved the exclusion of “strangers” from the community, prohibiting visitors from sleeping in one’s home (for fear that they might be running from the presence of Ebola infection in their own home), and mandating a 21-day waiting period for those who wished to move into the community to insure that they were Ebola-free. The second level of surveillance included the formation of a community task force that would enforce the exclusion of strangers, and would also assume a leadership role in prevention (like keeping community members away from sick people or the dead). The community task force was also suggested to be responsible for alerting community members to the presence of Ebola, monitoring the health of the sick and their family members, engaging in reporting, and managing resource provisions for community-based quarantines and isolation. At a third level of surveillance was the block watch team. Community leaders suggested that block watches could go house-to-house to monitor the sick, refer new cases to health facilities, and identify efforts to conceal sickness or burials. Fourth, individuals within households were expected to invest in their own domestic surveillance for Ebola by reporting cases of illness within their household, removing themselves or their family members from the possibility of contagion upon discovering sick individuals, and even isolating themselves so as not to infect family members.
The division of labor suggested in community surveillance was implicitly-–and sometimes explicitly—gendered. Women and men were both included in community leadership focus groups, and their reports and ethnographic evidence suggests that men were expected to serve on community task force teams, block watch teams, or community action teams to keep strangers out and engage in reporting and whistle blowing. There was some concern about remilitarization, violence, and destabilization due to this trend, as was born out during the West Point riots in Monrovia, and in armed confrontations between male youth and police and Ebola response teams in Sierra Leone. The mobilization of young men in these communities can, and often does, involve a range of martial and surveillance-like behaviors that can turn rather quickly into a remilitarization of social organization [41–42].
As we will further elaborate below, women, on the other hand, were expected to engage in domestic surveillance, to monitor the physical wellness or illness of family members while they washed, clothed, and fed children, spouses, siblings, and elderly persons, and care for the sick. As a counterpoint, the domestication of surveillance among women caring for the bodies of others within households might have put women at a greater risk of infection within the household, especially under quarantine and isolation conditions, while men were more likely to be infected outside of the household (e.g. through transportation and porterage activities).
Most importantly, community leaders argued that substantial investments in local infrastructure and systems were required to prevent the spread of the epidemic, recalling Paul Farmer’s much circulated call for “staff, stuff, and systems” [9]. They requested government and other organizational support to create community “holding centers” to serve as interim sites for the sick and dead while waiting for Ebola response teams and/or burial teams to arrive. They demanded a hotline system that prioritized rapid response to local communities’ calls to place sick people in hospitals and ETUs and remove bodies. Community leaders also recommended a broader local communications infrastructure, including a better-staffed call center, more ambulances, the establishment of mobile clinics or the reopening of local community clinics that had closed their doors, more testing centers, and finally, the training of additional health workers and burial teams. These health workers, they insisted, need to be paid and given adequate benefits.
Nearly all of the demands for the infrastructure improvements noted above derived from the experience of the failed international and national Ebola response observed and known to community leaders in the months of July, August, and September 2014. When community leaders called for the creation of a communications infrastructure, it was because their calls to hotlines had gone unanswered. When they called for the training of community members to provide care to the sick, manage holding centers, administer quarantines, and isolate or bury the dead, it was because they had experienced the social, medical, and ecological consequences of response teams not arriving in a timely manner or failing to arrive all.
Communication, or the lack thereof, constituted a critical part of the response’s failure at the juncture of prevention and response—the removal of the sick and the dead from communities. The anonymity involved in the process of removing the sick and the dead terrified people, and played a role in their decisions to avoid ETUs and hospitals and to conduct secret burials. In one case, a community leader reported losing an infected individual who had left the community for an “unknown destination.” After going from hospital to hospital, there was no record of that individual’s registration, death, or departure. This was seen as evidence of the failure of the international and national Ebola response.
In the absence of open clinics and hospitals, residents tried to assume responsibility for all aspects of healthcare in their local communities. Everything that follows in this section builds upon this premise. When community leaders engaged in discussions of “best practices” regarding response and treatment, community leaders agreed that the true “best response” was to obtain care in a hospital or ETU, to seek the removal of sick individuals by healthcare teams working for the government, and to engage in proper burials that reduced disease transmission. But, if resources were not in place, community leaders engaged in creative planning and response by innovating alternatives to how a community might best manage the presence of Ebola infection and dispose of infectious corpses.
This section details how community members prioritized the triage and treatment/care of unknown sickness and unexplained death in their communities. In order to assess how community members were likely to respond to illness in local communities, researchers asked a series of questions pertaining to their identification and management of illness in family, friends, and neighbors (Fig 3). Community leaders were quick to highlight the unnecessary morbidity and mortality due to preventable and treatable diseases, injuries, and basic health problems that were caused by the widespread closures of medical clinics and hospitals [43]. Most community leaders concurred that community members would be impelled to provide the best care they could offer for their families within their homes. Demands for guidance abounded: “We need to know how to protect ourselves while taking care of the sick.” “We want training and materials for how we can handle ourselves and the dead.” In order to align local realities with public health messages that were becoming increasingly unreasonable due to their disconnect with the lack of available services, they asked that community taskforce members be trained in basic healthcare provision, treatment of symptoms, and corpse removal.
When someone within a household fell ill, community leaders reported that they were first cared for within the home with palliative care. The person was administered “first-aid treatment,” including locally available pharmaceuticals, herbal remedies, locally accessible oral rehydration solutions or therapies, the provision of fluids, and the early administration of anti-malarials. It was not expected that any single household would be able to provide all of these interventions.
If the sick individual did not recover, community leaders expected caregivers to take a sick person to a hospital, or call the health team. There was a general consensus that if a person did fall ill, both health teams and the community must be notified so that they could respond and take precautions. However, the technical details regarding the proper handling of sick community members at home and in transit was ambiguous among respondents because it was ambivalent in the community. Transportation itself was a vector for infection. Leaders recounted examples of people who had walked to the hospital or carried sick children in their arms rather than travel by taxi or ask for help with transportation, in order to prevent the spread of infection.
Caregiving in all aspects demanded physical contact, but the public health messages regarding physical contact failed to recognize this reality. As some respondents noted, one message said, “Don’t touch,” while another said, “Touch, but use plastic gloves.” The lack of consistency may have resulted in experimentation and innovation, but it also elevated local perceptions that the message “Don’t Touch” was impractical and unhelpful in the context of immediate need. This was particularly significant in a social milieu in which nearly every able-bodied individual functioned as at least a part-time caretaker for children, elderly, or physically or mentally disabled friends and relatives.
Instead, community-based quarantine was identified as the best available strategy or approach. The process of quarantine required careful oversight and supply of resources, and community-leaders gave careful thought to how they might best support individuals and families in isolation and quarantine. In community leaders’ discussions, it was apparent that they sought to position the community at the center of the Ebola treatment response by managing the health and safety of quarantined families through food supply, illness surveillance and oversight, reporting, the provision of medical supplies, and communication and information. There was a strong willingness on the part of the community to serve as a central axis for interaction between the state and local individuals and families by doing the work of organizing food, medical, hygiene, and PPE distribution, case identification and surveillance, multi-level communication and reporting, and patient and corpse conveyance.
Community-based quarantine and home-based healthcare, however, caught communities in a Catch-22. Many community leaders were afraid of continued epidemic spread in their communities, and believed that health workers’ training, medicine, and materials prepared them to support and treat the sick [44]. Conversely, as one research team member commented in her field notes from one focus group, if a relative or neighbor was ill, just half of respondents were likely to call a health team hotline or encourage that person to go to a hospital or Ebola Treatment Unit (ETU). They had good reasons not to. By the time of this study, many of those who had gone to hospitals and ETUs had never returned, or had been turned away from multiple facilities due to lack of beds. Some were concerned that the failure to report the deaths of Ebola patients to their families and communities meant that loved ones’ bodies had been dissected for body parts after death, and that foreign NGOs were trying to keep the practice a secret [44].
This research offered direct insight into the fore-planning process of women as they consider how to respond if and when Ebola arrives in their households, families, and social networks. A broad subset of respondents—mainly women—reported that they would care for their sick family members on their own, and that they preferred to do so inside the home. They described a plan for isolating themselves with a sick family member[s] and for providing the best locally available appropriate care they could offer, using available resources (Fig 4). As one woman noted, “It will be impossible that my child or husband is sick and I refuse to touch them. I do not have the courage or heart to do that.” Referencing a widely circulated video of a nurse who had made her own personal protective equipment (PPE) from garbage bags, rain coats and boots, and gloves, an elderly woman reported her intention of making her own PPE from locally available materials. “I will find my own PPE (using a raincoat, plastic bags on hands) and care for sick relatives like I saw on television. If the person is not getting better, I will hold them (with the plastic still on my hands) and take them to the hospital.” Women showed an intense conviction that they should care for their families, and showed a desire to do so, despite risks to their own health.
Childcare, eldercare, home-based healthcare and the graduated triage approach were gendered activities; and home-based care constituted a zone of risk for both predominantly female caregivers and their dependents [45], in contrast to community-based surveillance, education, and transport roles, which functioned as a zone of risk for men. This may have also had unrecognized repercussions on infant and child mortality, and on the now-confirmed explosive chains of Ebola transmission within kinship networks and families during the Ebola crisis [46].
Community leaders reported that, after parents infected with Ebola had been removed to hospitals or had died, their children were placed under community quarantine for 21 days. During this time:
The reported deaths of young children under quarantine paint a challenging picture to communities’ descriptions of providing care (food and water) to local families and children throughout the quarantine period. It also facilitated the impression that women’s greater involvement in the direct support of children’s diet and healthcare and in traditional burial practices was resulting in a gender differential in levels of exposure, although this has not been borne out by the official data [47–48, 22]. A community-based approach lends itself to the interpretation that, although male and female case incidences and fatalities were roughly equal [2], women’s risk of mortality may have been impacted by their greater reluctance to seek early treatment for Ebola out of fear for children’s lives, and due to concern for children’s and dependents’ well-being under the quarantine that would follow their hospitalization at an Ebola treatment unit (ETU) or temporary quarantine at community-care centers (CCCs).
Other responses indicated the existence of a path for Ebola response in the community that diverged from both formal healthcare seeking and home-based care. The retreat and isolation of sick individuals was characterized as a protective measure for the community, and as an act of generosity from sick individuals to the community. Survivors characterized their own self-isolation as a strategy for honoring and caring for the families of people who had died. Unsupported quarantine entailed the isolation of an infected individual, without providing treatment to him/her, or referring him/her to a healthcare facility. As one person noted, “If my son is sick, I will run away from him. I am not a health worker to tell whether it is Ebola or malaria. It will be better he dies alone and I be left behind to care for his sisters and brothers.” The intention of this statement sounds chilling, but its meaning is complex, and considers the futures of the people surrounding the individual who present as terminally ill, with little likelihood of treatment or survival.
The legacies of Ebola in Liberia appear to be drastic and long-lasting [48]. This study identified three critical issues pertaining to the sequelae of Ebola: (1) the reintegration of Ebola survivors into local communities, (2) the care and management of “Ebola orphans”—or children who had lost one or both parents to Ebola, and (3) memorialization of individuals who have died of Ebola.
At the height of the outbreak, when these data were collected, focus group participants reported support and acceptance for Ebola survivors in the community. Ebola survivors were widely recognized as being an asset in the fight against Ebola. They were seen as being an embodiment of positive messages suggesting that early treatment could allow one to survive Ebola, and they were referred to many times as “ambassadors” of Ebola awareness, as “living testimony to the Ebola crisis,” and as positive role models. Within focus groups, survivors already appeared to be accepting their positive role model status, and offered vignettes like the following:
Although some community leaders mentioned that they were afraid of survivors, most indicated that they welcomed the return of Ebola survivors into their community. They acknowledged that they understood that they were no longer infectious, and that these individuals could not be re-infected with the virus. Concern was voiced that the Ebola virus was found in semen for up to three months following infection, and that previously infected individuals must take care to avoid sexual relations. Others conflated this three-month time period with the perception that these individuals were still infectious, and recommended that they be placed in a halfway house for three months following an infection.
The care of children orphaned by Ebola was widely regarded as a communal responsibility. Some community members mentioned that children would be brought into their homes and families, noting, “The children become our children. These children are our own because their parents are no more.” Another commented that the community leaders must, “encourage people to take children whose parents have died of Ebola as their own, because we have lived in the same community for years and they are like family to us.” Recalling the war and the fragility of life in urban Monrovia [34], one individual noted, “We have done it before. We will take care of the children. Education, feeding, and shelter.” Several respondents mentioned that children required psychosocial counseling to recover from “traumatization” due to having lost their parents, and they recommended that NGOs make counseling available to Ebola orphans in communities and in orphanages. They also stipulated that orphans should not be forced to endure stigma or discrimination because their parents had died of Ebola.
Despite the will to care for the children, there was considerable concern about the economic, emotional, and residential burden that additional children would impose. When children had extended families elsewhere, communities felt that they should be united with them, and that NGOs must assist with “family reunification.” Others called for the creation of orphanages within their communities so that community members could oversee the development of the children, while delegating financial and educational costs to NGOs, UNICEF, the World Health Organization, and the MOHSW. The majority of individuals, however, seemed to hope that children would be able to stay within the community, but would receive financial support for their clothing, food, and education from governmental, non-governmental, ecumenical, and charitable sources.
The data from this research indicated the need for some form of widespread and public memorialization of the lives lost to Ebola. Community members called for: a National Memorial Day; the construction of a statue in the middle of the community; a formal memorial service at the end of the outbreak; a parade; or a day called “Black Day” to be recognized by law. Community members also recommended that a mass grave be built with a headstone inscribing the names of all who had died of Ebola, and had been cremated. The goal of this memorial is to provide family members with a space to “remember” and pay tribute to their loved ones by visiting the grave and laying flowers. Additional suggestions were practical, and included providing scholarships, financial aid, general support, and counseling to Ebola orphans.
While the research reported here takes considerable strides towards helping to understand how local communities in Liberia responded, and envisioned their response, to Ebola, this information must not be mistaken as an indication of community political, medical, or social empowerment or institution-building—although this, too, was present. These communities were not empowered, they were desperate and often abandoned. They found resources from within their communities to compensate for the collective failure of state and international institutions to implement systems of surveillance, treatment, and response. What we are observing here is a community-based response to a condition of medical statelessness and structural violence [49–50].
Even so, health sensitization efforts continued to emphasize the ‘low-hanging fruit’ of public health communications [51] throughout the response, repeating messages like: “What is Ebola? How is it spread? What are the symptoms? How long does it last?” But community health messaging essentially failed to provide the kinds of ‘higher-order,’ practical information and training that communities were desperate for, like “How do I manage a family of children, including infants and toddlers, in quarantine?” “How do I transport someone to a hospital or clinic without promoting infection?” “What capacities need to be built to support a holding center?” “What does my community do with an exposed and infectious body when the health teams do not come to collect it?” “What can I do to make sure that you don’t lose or steal my father/brother/sister/mother at your health facility?”
In the future, engaging local communities in epidemic response will require answering their challenging questions about their encounters with systemic failures in real time. Communities sought guidance for triaging a sick person when he or she had been turned away from hospitals, for building and supporting holding units in communities, and for reporting deaths when their calls to hotlines went unanswered. The global health community needs to consider what it would mean to put into place surveillance and reporting mechanisms in which community-based leaders have the ability to directly account for health, illness, or death of every individual in the population. This could be done through the creation of health identification numbers, the creation of health census lists, or other mechanisms of reporting and marking. In a context in which every death is an Ebola death because there are no community-based testing facilities for Ebola, every death needs to be counted as worthy of being reported. (And when everyone has a number, everyone counts.)
But can locally affected populations, in effect, govern themselves by engaging in medical self-surveillance, self-management, and self-triage? The ethnographic evidence suggests that they can indeed do so [22,52–53], whilst the public health literature has previously examined the importance of local surveillance in other hemorrhagic fever outbreaks in Africa [54–55]. In the short-term, risk can be moderated, in part, by ensuring that required daily resources like food, medicine, housing, PPEs, and other resources are in abundance, and that informational demands meet local populations “where they are.” In the long term, efforts need to focus on equipping local communities with the material and knowledge resources to respond to Ebola and to help build a surveillance infrastructure that can inform a stronger post-epidemic local and state governance architecture.
The study also suggests that the gendered distribution of morbidity and mortality in this Ebola outbreak is strongly associated with existing relations of caregiving and with the distribution of labor in community surveillance and response. The most important thing to understand about culture and caregiving is that women are not going to abdicate the role of primary caregivers. Indeed, the data collection exercise offered direct insight into the fore-planning process of women as they consider how to respond if and when Ebola were to arrive in their households, families, and social networks. Resources must be set into motion to support men and women in their community-allocated surveillance roles and to support women in their caregiving roles, in order to engender support for local-international collaboration and connect the Ebola response effort to the lived experiences of local persons.
This study had two major limitations. First, the number of participants in the focus groups was large (15–20). In order to address this issue, the PI stationed four research assistants around the group in order to capture the responses of all participants. Second, the data collected were partly conjectural, and questions were posed as hypothetical or focused on best practices, rather than on direct local experiences and actions, in order to avoid issues of concealment or avoidance during data collection. As a result, community leaders’ feedback is regarded by the researchers as an “ideal-typical” representation of what a community-based response to Ebola should have been like, rather than a factual account of how these same communities actually responded to the incidence of Ebola. Respondents shared information that was based on their knowledge of Ebola, on community and government messages that they had received about Ebola, on resources that they sought for their communities, and on their experiences with Ebola and non-related Ebola morbidity and mortality.
|
10.1371/journal.pcbi.1006874 | Genesis of the αβ T-cell receptor | The T-cell (TCR) repertoire relies on the diversity of receptors composed of two chains, called α and β, to recognize pathogens. Using results of high throughput sequencing and computational chain-pairing experiments of human TCR repertoires, we quantitively characterize the αβ generation process. We estimate the probabilities of a rescue recombination of the β chain on the second chromosome upon failure or success on the first chromosome. Unlike β chains, α chains recombine simultaneously on both chromosomes, resulting in correlated statistics of the two genes which we predict using a mechanistic model. We find that ∼35% of cells express both α chains. Altogether, our statistical analysis gives a complete quantitative mechanistic picture that results in the observed correlations in the generative process. We learn that the probability to generate any TCRαβ is lower than 10−12 and estimate the generation diversity and sharing properties of the αβ TCR repertoire.
| Receptors on the surface of T-cells recognize pathogens and initiate an immune response. Analyzing the sequences of human T-cell receptors we draw a detailed quantitative picture of the generation process of the two receptor chains allowing us to estimate the diversity of the complete repertoire. We discuss which elements of the receptor production processes are correlated and which are independent, proposing mechanistic models at the origin of the correlations. We discuss the implications of our findings for the functional role of each of these cells, and the diversity of the repertoire.
| The adaptive immune system confers protection against many different pathogens using a diverse set of specialized receptors expressed on the surface of T-cells. The ensemble of the expressed receptors is called a repertoire and its diversity and composition encode the ability of the immune system to recognize antigens. T-cell receptors (TCR) are composed of two chains, α and β, that together bind antigenic peptides presented on the multihistocompatability complex (MHC). High-throughput immune sequencing experiments give us insight into the repertoire composition through lists of TCR, typically centered around the most diverse region, the Complimentary Determining Region 3 (CDR3) of these chains [1–5]. Until recently most experiments and analyses focused on only one of the two chains at a time, and studies of TCR with both chains were limited to low-throughput methods [6–8]. Recent technological and analytical breakthroughs now allow us to simultaneously determine the sequences of both α and β chains expressed on cells of the same clone in a high-throughput way [9] (see also analysis of unpublished data obtained by single-cell sequencing in [10]). These advances make it possible to study the repertoires of paired receptors, and to revisit the questions of the generation, distribution, diversity and overlap of TCR repertoires previously studied at the single-chain level [11–17], but also to gain insight into the mechanisms of T-cell recombination and maturation.
TCR receptor diversity arises from genetic recombination of the α and β chains of thymocytes in the thymus. Each chain locus consists of a constant region (C), and multiple gene segments V (52 for the human β chain and ≈ 70 for α), D (2 and 0) and J (13 and 61). Recombination proceeds by selecting one of each type of segment and joining them together, with additional deletions or insertions of base pairs at the junctions. TCRβ is first recombined and expressed along with the pre-T cell receptor alpha (a non-recombined template gene) on the surface of the cell to be checked for function. T cells then divide a few times before TCRα recombination begins, at which point the thymic selection process acts on the complete receptor. The recombination of each chain often result in non-productive genes (e.g. with frameshifts or stop codons). Subsequent rescue and selection mechanisms ensure that all mature T cells express at least one functional receptor. Recombination of the β chain on the second chromosome may be attempted if the initial recombination was unsuccessful. By contrast, the α chain is recombined on both chromosomes simultaneously [18], and proceeds through several recombination attempts that successively join increasingly distal V and J segments (Fig 1). Taken together, recombination events can potentially produce up to 4 chains (2 α and 2 β) in each cell. In principle, allelic exclusion ensures that only one receptor may be expressed on the surface of the cell, but this process is leaky: 7% of T-cells have two productive β-chains [19, 20], and %1 express both of them on the surface [21–23]. Allelic exclusion in the α chain is less well quantified as it relies on different mechanisms [24, 25], with estimates ranging from 7% [8] to 30% [22] of cells with two functionally expressed α chains.
Despite the partial characterization of the various mechanisms underpinning the recombination, rescue and selection of the two TCR chains, a complete quantitative picture of these processes is still lacking. For instance, the probability of recombination rescue, the probability for a chain to pass selection, or the extent of allelic exclusion, have not been measured precisely. Here we re-analyse the data from [9] to link together each of the 4 α and β chains of single clones, and study α-α and β-β pairs as well as α-β pairs. Using these pairings, we propose a mechanistic model of recombination of the two chains on the two chromosomes, inspired by [26], and study the statistics of the resulting functional αβ TCR.
We analysed previously published data on sequenced T-cell CDR3 regions obtained from two human subjects (PairSEQ), as described by Howie and collaborators [9]. In the original study, sequences of α and β chain pairs associated to the same clone were isolated using a combination of high-throughput sequencing and combinatorial statistics. Briefly, T cell samples were deposited into wells of a 96-well plate, their RNA extracted, reverse-transcribed into cDNA with the addition of a well-specific barcode, amplified by PCR, and sequenced. αβ pairs appearing together in many wells were assumed to be associated with the same T-cell clone, and thus expressed together in the same cells. Because the method relies on the presence of cells of the same clone in many wells, the method can only capture large memory T cell clones present in multiple copies in the same blood sample. Naive clones which have a population size of around 10, or concentration of 10−10 [27], are not expected to be paired in this way.
We generalized the statistical method of [9] to associate α-α and β-β pairs present in the same clone. Along with α-β pairings, this allowed us to reconstruct the full TCR content of a cell. Two additional difficulties arise when trying to pair chains of the same type. First, truly distinct pairs of chains must be distinguished from reads associated with the same sequence but differing by a few nucleotides as a result of sequencing errors. We set a threshold of 11 nucleotide mismatches on the distribution of distances between paired chains (S1 Fig) to remove duplicates while minimizing the loss of real pairs. Second, because of allelic exclusions, one of the two chains of the same type is typically expressed in much smaller amounts than the other. As a result, we find much fewer α-α and β-β pairs than α-β pairs.
Table 1 summarizes the numbers of pairs found in each experiment, with a significance threshold chosen to achieve a 1% false discovery rate (see Methods). This method can then be used to recreate the complete TCR content of a given clone, and set apart clones expressing multiple TCR receptors.
Correlations between the features of the recombination events of the chains present in the same cells are informative about the rules governing the formation of a mature αβ TCR in the case of α-β pairings, and also about the mechanisms and temporal organization of recombination on the two chromosomes in the case of α-α and β-β pairings. We computed the mutual information, a non-parametric measure of correlations (see Methods), between pairs of recombination features for each chain: V, D, and J segment choices, and the numbers of deletions and insertions at each junction (Fig 2). Because recombination events cannot be assigned with certainty to a given sequence, we used the IGoR software [28] to associate recombination events to each sequence with a probabilistic weight reflecting the confidence we have in this assignment (see Methods). We have shown previously that this probabilistic correction removes spurious correlations between recombination events [12, 28]. Correlations within single chains recapitulate previously reported results for the β [12] and α [29] chains. Inter-chain correlations, highlighted by red boxes, are only accessible thanks to the chain pairings.
We find no correlation between the number of insertions in different chains across all pair types. Such a correlation could have been expected because Terminal deoxynucleotidyl transferase (TdT), the enzyme responsible for insertions, is believed to correlate with the number of inserted base pairs [30], and is expected to be constant across recombination events in each cell. The lack of correlation between different insertion events thus suggests that the there is no shared variability arising from differences in TdT concentration across cells.
We report generally weak correlations between the α and β chains (Fig 2A and S2 Fig for an analysis of statistical significance), with a total sum of 0.36 bits, about 10 times lower than the total intra-gene correlations of the α chain. The largest correlation is between the choice of Vα and Vβ genes (0.036 bits) and Jα and Vβ genes (0.033 bits), in agreement with the analysis of [10] on unpublished single-cell data. These correlations probably do not arise from biases in the recombination process, because recombination of the two chains occurs on different loci (located on distinct chromosomes) and at different stages of T cell maturation. A more plausible explanation is that thymic selection preferentially selects some chain associations with higher folding stability or better peptide-MHC recognition properties. Distinguishing recombination- from selection-induced correlations would require analysing pairs of non-productive sequences, which are not subjected to selection, but the number of such pairs in the dataset was too small to extract statistically significant results. An analysis of the correlations between gene segments (S3 Fig) does not show any particular structure.
Pairs of β chains show almost no correlations (Fig 2C and S2 Fig for an analysis of statistical significance). Looking in detail at the correlations between gene segments reveals a strongly negative correlation of TCRBV21-01 and TCRBV23-01 (both pseudogenes) with themselves (S4 Fig), which is expected because at least one of the two β chain must have a non-pseudogene V. More generally, correlations are likely to arise from selection effects, since the two recombination events of the two β chains are believed to happen sequentially and independently. The fact that at least one of the chains needs to be functional for the cell to survive breaks the independence between the two recombination events.
By contrast, the α-α pairs have very strong correlations between the V and J usages of the two chromosomes, and none between any other pair of features (Fig 2B). These correlations arise from the fact that the two α recombination events occur processively and simultaneously on the two chromosomes, as we analyse in more detail below.
We wondered whether the detailed structure of the observed correlations between the α chains on the two chromosomes could be explained by a simple model of recombination rescue. The correlations of the Vα segments on the two chromosomes and of the Jα segments show a similar spatial structure as a function of their ordering on the chromosome (see Fig 3A): proximal genes are preferentially chosen together on the two chromosomes, as are distal genes. The correlations between the Vα gene segment on the first chromosome and the Jα on the second chromosome also show a similar diagonal structure (S5 Fig).
The two chromosomes recombine simultaneously, and proceed by successive trials and rescues. If the first recombination attempt fails to produce a functional chain, another recombination event may happen on the same chromosome between the remaining distal V and J segments, excising the failed rearranged gene in the process. The recombination of a functional chain on either of the chromosomes immediately stops the process on both chromosomes. By the time this happens on one chromosome, a similar number of recombination attempts will have occurred on the other chromosome. We hypothesize that this synchrony is the main source of correlations between the Vα and Jα gene usages of the two chains.
To validate this hypothesis, we simulated a minimal model of the rescue process similar to [26] (Methods), in which the two chromosomes are recombined in parallel. If recombination happens to fail on both chromosomes, repeated “rescue” recombinations (which we limit to 5) take place between outward nearby segments (Fig 3C). The covariance matrices obtained from the simulations for both Vα and Jα (Fig 3B) show profiles that are very similar to the data, with positive correlations along the diagonal, in particular at the two ends of the sequence. However, the actual distributions of V and J genes segments (see S6 Fig) are much more heterogeneous than the slowly decaying distribution implied by our simple model: the question of gene usage is further complicated by other factors, such as gene accessibility and primer specificity.
We wondered if the paired data could be used to estimate the percentage of cells with two recombined chains of the same type. However, since pairing was done based on mRNA transcripts through cDNA sequencing, silenced or suppressed genes are not expected to be among the identified pairs, leading to a systematic underestimation of double recombinations. While the authors of [9] also provided a genomic DNA (gDNA) dataset that does not have this issue, the number of sequences was too small to resolve statistically significant pairings. Nonetheless, we can derive strict bounds from the proportion of productive sequences found in this (unpaired) gDNA dataset. Following recombination, using IGoR we estimate p nc α = 69 . 5 % of the α sequences, and p nc β = 73 . 5 % of β sequences are non-coding or contain a stop codon. We collectively refer to as “non-coding” sequences. The remaining sequences, called “coding”, make up a fraction p c α , β = 1 - p nc α , β of random rearrangements. We denote by p f α and p f β the probability that a coding sequence can express a functional α or β chain that can ensure its selection.
The number of observed non-coding sequences depends on whether the second chromosome attempts to recombine following the recombination of the first one. We call pr the probability that a second recombination happens when the first recombination fails to produce a functional chain, and p r ′ when the first recombination succeeds. Then, the proportion f nc of observed non-coding sequences can be written as (see tree in Fig 4 and Methods):
f nc = ( p r + p r ′ ) p nc 1 + p r ′ + 2 ( 1 - p f p c ) p r . (1)
Note that this formula assumes that the presence of more than one functional chain does not affect its selection probability. Comparing the proportion of observed non-coding β chain sequences calculated from Eq 1 with the values from gDNA data (f nc β = 18 ± 1 % in [9] and 14% in [11]), allows us to constrain the values of p r β and p r β ′. The probability of a second recombination, even if the first recombination failed, is always lower than 65% (Fig 4A). By constrast, the observed fraction of non-coding sequences in the α chain, f nc α = 40 ± 1 %, constrains the the rescue probabilities p r α and p r α ′ to be close to 100% (Fig 4B), in agreement with the fact that both chromosomes are believed to recombine independenly. Assuming strict independence, p r α = p r α ′ = 1 puts bounds on the probability that a random coding α sequence is functional, 70 % ≤ p f α ≤ 100 %.
Can we learn from pairing data what fraction of cells expressed two chains of the same type? gDNA pairings do not allow us to do that, because they are severely limited by sequencing depth: most chains cannot be paired because of material losses, and estimating the fraction of cells with several chains is impossible. While cDNA pairings are in principle less susceptible to material loss, non-functional sequences are much less expressed than functional ones [23, 25], lowering their probability of beingfound and paired and introducing uncontrolled biases in the estimate of fractions of cells with different chain compositions. However, we can use this difference in expression patterns by examining the distribution of read counts for each type of chain. We use both the second and the third experiments, which are more data-rich than the first one. The total read count of each sequence is obtained by summing its read count in individual wells.
Sequences of chains paired with a non-coding chain of the same type must be functional and expressed on the surface of the cell. These sequences are more expressed than non-coding ones, and their distributions of read counts are markedly different (S7A Fig).
Comparatively, coding sequences paired with a coding sequence of the same type can be either expressed or silenced, depending on their own functionality and the status of the other chain. Thus, their read count should follow a mixture distribution of both expressed and silenced sequences (S7B and S7E Fig) the latter being assumed to follow the same distribution as noncoding sequences. The best parameter fit of this mixture to the read counts of paired coding sequences (S7C and S7D Fig) yields the total proportion p e of functional sequences among the coding sequences coupled with another coding sequence. For α sequences, we found p e , exp 2 α = 0 . 66 ± 0 . 03 for experiment 3 and p e , exp 3 α = 0 . 69 ± 3, meaning that around 2 p e α - 1 ∼ 0 . 35 ± 0 . 1 of cells express two different α chains (see Methods). This number is consistent with older results [31], but higher than a recent estimate of 14% based on single-cell sequencing [8]. Another estimate from the same data [31], but taking into account material loss (see Methods), suggests that 24 ± 5% of cells have two functional and expressed α chains, more consistent with our own estimate. Of course, one of the major limitations of this method is that it only applies to relatively large clones, that can be paired by the PairSeq method, and it cannot be excluded that this ratio differs in naive cells for example. It should also be noted that the (fitted) mixture distribution and the original distribution do not coincide exactly (S7C and S7F Fig), this could be due to imperfect silencing or to a difference in the expression levels of non-coding and silenced sequences.
For β chains, the fit is noisier, because non-coding sequences are much more suppressed and therefore scarcer than for the α chain (only 4.5% of sequences are non-coding). We estimate that there are 8-10 times more silenced coding sequences than non-coding sequences, but the fit does not allow us to estimate the fraction of cells with two expressed β chains, although this number is consistent with 0 according to the data.
It is often assumed that all coding sequences must be functional, and previous studies have used the difference between coding and non-coding sequences to quantify the effects of selection [14, 32, 33]. However, some fraction of coding sequences may actually be disfunctional, silenced, or not properly expressed on the cell surface. By contrast, sequences that can be paired with a non-coding sequence of the same type must be functional and expressed on the cell surface, lest the cell that carries them dies. These sequences represent a non-biased sample of all functional sequences, and their statistics may differ from those of ‘just coding’ sequences. In Table 2 we report the differences between the two ensembles in terms of CDR3 length (defined from the conserved cystein of V and the conserved phenylalanine or tryptophan of J, corresponding to IMGT positions 105 to 117) and gene usage. All comparisons are with sequences that could be paired with another one to remove possible biases from the pairing process.
We find that functional sequences are on average slightly larger (by 1-2 nucleotides) than coding and non-coding sequences (Table 2 and S8 Fig). More markedly, the variance of their length is smaller, implying stronger selection towards a prefered length in the functional ensemble than in the coding and non-coding ensembles. These observations, which hold for both the α or β chains, indicate that the functional ensemble (as defined here using pairing information) is more restricted than ‘just coding’ sequences, and gives a more precise picture of the selected repertoire.
The impact of selection can also be measured by how much gene usage departs from the unselected ensemble using the Kullback-Leibler divergence (see Methods and Table 2), offering a more contrasted view. Vβ and Jα usages are similar in functional and coding sequences in terms of their divergence with non-coding sequences. For Jβ however, this divergence is higher in functional than in simply coding sequences, while the opposite is true for Vα.
Ignoring small correlations between features of the α and β chains reported in Fig 2, we can assume that the probability of generating a αβ pair is given by the product of the probabilities of generating each chain independently. These probabilities can be calculated using the IGoR software [28] for each paired chain in our datasets. The distribution of the pair generation probabilities obtained in this way (Fig 5A) shows an enormous breadth, spanning more than 20 orders of magnitude. We self-consistently validated the assumption of independence by showing that random assortments of α and β chains yielded an identical distribution of generation probabilities (green curve).
The maximum TCR generation probability is <10−12, meaning that generating the same pair twice independently is extremely unlikely. This suggests that, without strong antigenic selection, only a negligible number of full TCR sequences will be shared in samples obtained from distinct individuals. To make that prediction more quantitative, we simulated a computational model of sequence generation followed by thymic selection. α and β chains were generated by IGoR, and then each TCRαβ amino-acid sequence was kept with probability q to mimick thymic selection [17]. We further assume that selection acts on each chain independently, so that the ratio q is given by qα qβ, where qα,β are the selection probabilities infered from the analysis of single chains. These selection factors can be obtained by fitting the curve giving the number of unique amino-acid sequences as a function of unique nucleotide sequences [17], yielding qβ = 0.037 and qα = 0.16 (S9 Fig).
Using the model, we can make predictions about the expected number of TCRαβ nucleotide sequences shared between any of 10 individuals (Fig 5B) for which a million unique synthetic TCRαβ were obtained. We find that, while a substantial fraction of sequences of each chain are expected to be shared by several individuals, sharing the full TCRαβ is very unlikely, and drops well below 1 for more than 2 individuals. This suggests that the existence in real data of any TCRαβ shared between several individuals should be interpreted as resulting from strong common selection processes, probably associated with antigen-specific proliferation, leading to convergent selection of the shared sequences. A concomitant question concerns the total number of TCR sequences shared between two individuals. This number does not depend on selection or sample size, but rather on the total number of different clonotypes in an individual. While this last quantity is not precisely known, estimates range between 108 and 1011 [13, 34]. Using the analytical formulas and numerical procedure described in [17] with these estimates of the repertoire size, we predict the proportion of shared clonotypes between two individuals to fall between 0.001% and 0.1% of their full repertoires (see Methods for details).
To further investigate the effects of convergent selection, we quantified how often the same α chain was associated with distinct β chains in different clones (S10A Fig), and vice versa (S10B Fig). While association of a β with 2 distinct α chains could happen in the same cell because of the existence of two copies, we found a substantial fraction (3%) of all paired TCRβ that could be associated with three or more TCRα.
Convergent recombination of β can create clones that shares their β but not their α chains. This effect can be quantified using the generation and thymic selection model introduced in the previous paragraph. Simulations with the same sample sizes as the data show that such convergent recombination is predicted to happen with a rate of 0.5%, and thus cannot explain the data. However, there is another effect at play: cells divide around 5 times between β and α recombination, which leads to clones with the same β chain but with up to 25 ∼ 30 distinct α chains. A simulation considering these two effects together (see Methods) predicts a sharing fraction of 3%, consistent with the fraction observed empirically.
Analysing computationally reconstructed pairs of TCR α and β chains, as well as α-α and β-β pairs, allowed us to quantify the various steps of sequence generation, rescue mechanisms, convergent selection, and sharing that were not accessible from just single-chain data.
Pairing α chains in single cells revealed correlations that were suggestive of a parallel and processive mechanism of VJ recombination in the two chromosomes. These signatures were well recapitulated by a simple computational model of successive rescue recombinations. Our model is similar to that of [26], but differs in its details and parameters, as the original model could not reproduce the correlation pattern of the data.
We estimated that ∼35 ± 10% of cells express two α chain, higher than a recent report of 14% using single-cell sequencing [8]. However, this fraction is very hard to assess directly, as material loss can lead to its underestimation—in fact, correcting for this loss with minimal assumptions gives a fraction ∼24% instead of 14%. While our estimate is indirect, we expected it to be more robust to such loss.
Our finding that the statistics of the two chains are largely independent of each other—with only a weak correlation between Vβ and (Vα, Jα) usage—is in agreement with recent observations using direct single-cell chain pairing [10]. While independence between the α and β recombination processes is perhaps expected because they occur at different stages of T-cell development, it is worth emphasizing that the absence of correlations reported here involves coding TCRαβ sequences, which are believed to be largely restricted by thymic selection. This restriction can introduce correlations, notably through negative selection which could forbid certain αβ combinations. Our results do not exclude such joint selection, but suggests that it does not introduce observable biases. The independence between the two chains implies that the entropies of the two generation processes can be simply summed to obtain the entropy of the full TCRαβ. Taking the values previously reported in [15] of 26 bits for the α chain, and 38 bits for the β chain, yields 64 bits for the TCRαβ, i.e. a diversity number of 264 ≈ 2 ⋅ 1019.
The independence between the chains also allowed us to make predictions about the amount of TCR repertoire overlap one should expect between samples from different individuals. Our analysis predicts that sharing of αβ pairs between two samples should be rare, and that sharing between more than two is exceptional. In a recent report [10], 26 TCRαβ pairs were found to be shared between any 2 of 5 individuals. Our result indicate that such a high level of sharing cannot be explained by convergent recombination alone: by simulating samples of the same size as in [10], we estimated a total expected number of 0.001 sequences between all their pairs (see Methods). The much higher number of shared sequences reported in the original study may result from over-correcting for sequencing errors, or alternatively from strong convergent selection in all 5 donors. A clonotype expansion of 104 (not unexpected in the context of an immune response, see e.g. [35]) would be sufficient to explain this result.
Future studies collecting the αβ repertoires of more individuals, as promised by the rapid development of single-cell sequencing techniques, will help us get a more detailed picture of the diversity and sharing properties of the TCRαβ repertoires. Our analysis provides a useful baseline against which to compare and assess the results of these future works.
The generation model was obtained and used through the IGoR software [28]. The IGoR software is able to learn, from out-of-frame receptor sequences, the statistics of a V(D)J recombination process. We don’t use IGoR in its inference capacity here, but rather rely on the pre-inferred recombination model for TRA and TRB chains in humans supplied with IGoR, as the recombination process is widely shared between individuals [12]. Briefly, the probabilities of recombination of α and β chains factorize as:
P recomb α = P ( V , J ) P ( del V | V ) P ( del J | J ) P ( ins V J ) ∏ i ins V J P V J ( n i | n i - 1 ) , (2)
P recomb β = P ( V , D , J ) P ( del V | V ) P ( ins V J ) P ( del D 5 ′ del D 3 ′ | D ) P ( ins D J ) (3)
× P ( del J | J ) ∏ i ins V D P V D ( m i | m i - 1 ) ∏ i ins D J P D J ( r i | r i - 1 ) , (4)
where (ni), (mi), (ri) are the inserted nucleotides at the VJ, VD, and DJ junctions. IGoR infers these probabilities through an Expectation-Maximization algorithm as described previously.
We rely on IGoR for:
We use the data and method of [9] to infer pairing from sequencing data of cells partitioned in W = 95 wells (instead of 96 as erroneously reported in the original paper, as one of the wells did not provide any results). We calculate the p-value that two sequences each present in w1 and w2 well are found together in w12 wells, under the null model that they are distributed randomly and independently:p ( w 1 , 2 , w 1 , w 2 , W ) = ∑ u ≥ w 12 ( w 1 u ) ( W − w 1 w 2 − u ) / ( W w 2 ).
We first select all the pairs under a given p-value threshold (10−4). For α−α and β−β pairs, we apply a threshold on their Levenshtein distances in order to remove most of the false pairings (pairing of near identical sequences due to sequencing errors). Then for each pair of well occupation numbers (w1, w2), we set the p-value threshold so that false discovery rate (using the Benjamini—Hochberg procedure) is always less than 1%. Compared to the analysis of Ref. [9], where the discreteness of the p-value distribution was taken into account by using a permutation algorithm, our approach is more conservative, as we worried about the potential effect of fake pairings on the false discovery rate. Thus our reported number of pairs (Table 1) slightly differs from that reported in the original study.
The mutual information (in bits) of two variables X, Y with a joint distribution p(x, y) is defined by: I(X, Y) = ∑x,y p(x, y)log2[p(x, y)/(p(x)p(y))]. We estimated it from the empirical histogram of (x, y) using a finite size correction [36], (nX nY − nX − nY + 1)/2N log(2), where N is the sample size, and nA is the number of different values the variable A can take.
In the specific case of sequences in paired cells, a better correction can be obtained by computing the mutual information between shuffled sequences, where the two chains are assorted at random.
The Kullback-Leibler divergence between two distribution p(x) and q(x) of a variable X is given by: DKL(p∥q) = ∑x p(x)log2(p(x)/q(x)).
The V and J genes are indexed by i and j from most proximal to most distal along the chromosome: Vi, i = 1, …LV and Jj, j = 1, …LJ. In the first recombination attempt of the first chromosome, the model picks the V and J gene indices i1 and j1 from a truncated geometric distribution, P(i1 = i) ∝ (1 − p)i−1 (and likewise for j1), with p = 0.05. The same process is simulated for the second chromosome. With probability 2/3 for each chromosome, the recombination fails. If both chromosome fail, a second recombination takes place on each between more distal genes indexed by i2 > i1 and j2 > j1, distributed as P(i2 = i) ∝ (1 − p)i2−i1−1 (and likewise for j2), to reflect observations that successive recombination occur on nearby genes in the germline [37]. If recombination repeatedly fail on both chromosomes, the process is repeated up to 5 times [38], after a success however, on any of the two chromosomes, it stops. This model is similar to that of [26], where a uniform instead of a geometric distribution was used.
Non coding sequences can only appear in the TCR repertoire if they share a cell with a functional sequence. The probability of such a cell to appear in the selection process is A = p nc ( p r + p r ′ ) p c p f. The probability for a cell to possess only one functional receptor is B = p c p f ( 1 - p r ′ ), while the probability to possess two receptors and at least one functional one can be written as C = p c p f [ p r ( 1 - p c p f ) + p r ′ ]. The proportion of non-coding reads is thus A/(B + 2C), which gives Eq 1.
We have shown that the pairs Vα − Vβ and Jα − Vβ were not independent (Fig 2). In this section we define the simplest model that can reproduce these correlations. The marginal distributions p V α , J α and p V β, coupled with the experimental pairing data can be used to obtain selection factors q V α , J α V β:
p ( V α , J α , V β ) = p V α , J α p V β q V α , J α , V β (5)
By adding a tunable temperature, we can modify the level of selection we want to observe:
p ( V α , J α , V β ) ∝ p V α p V β ( q V α , V β ) 1 / T (6)
When T → 0, the selection conserves only a few specific pairs of V, while for T → ∞ there is no selection. This modifies the mutual information between Vα and Vβ in the same cell, but also, because V and J on the same chromosome are not independent, the mutual information between Vα and Jβ. In S11 Fig, we show the evolution of the mutual information between Vα, Jα, Vβ and Jβ as a function of T. The model underestimates the mutual information between Vα and Jβ which hints that it may be necessary to also include Jβ in the selection model.
In order to estimate the ration of expressed sequences in a set of coding chains, we fit the empirical distribution of reads per coding chain, ρc, with a mixture of two distributions (S7 Fig): ρe, corresponding to chain sequences that could be paired with a non-coding sequence of the same type and thus believed to be expressed; and ρnc corresponding to non-expressed sequences and learned from non-coding sequences. Each distribution is estimating by taking histograms with bin size chosen using the Freedman-Draconis rule. For a given parameter pe, the mixture distribution is obtained by sampling N p e expressed chains and N(1 − pe) non-coding sequences, with N large. The fit is done by minimizing the (two-sample) Kolmogorov-Smirnov (KS) distance between the two distributions, and the error bars are obtained through bootstrapping. The result of the fit is a parameter pe corresponding to the proportion of expressed sequences among the chains. Applying this method to coding chains paired with another coding chain, we can infer p2α, the proportion of cells with two expressed α. The relation between p2α and pe should be p e = ( 2 p 2 α + ( 1 - p 2 α ) ) / 2, hence p2α = 2pe − 1. A different approach to estimate pe from the data consists in comparing the mean of the distributions. While this gives poor results with raw data due to the long tail of the distributions, it matches the distance-minimization result when the distributions are log-transformed. We find a value p e α = 28 % ± 10 %, not compatible with the value of 14% ± 3% obtained in [8] (19 out of 139 cells in which at least one productive sequence was found). But the authors of [8] make their estimate by sequencing cDNA, which can lead to different drop-out rates depending on the nature of the sequence. Silenced productive sequences or non-productive sequences are less expressed and their drop-out rates are higher. They find two TCRA (productive or not) in only 58% of cells, while both TCRA are expected to recombine [39]. In this context the 14% rate can only be understood as a lower bound. Assuming that non-productive and silenced sequences are expressed in similar quantities, we obtain an estimate for p e α of 24% ± 5% (19 out of the 80 cells which had two sequences, productive or not) from their data, which is consistent with our result.
We follow the methods of [17]. A large number of productive α and β chain pair sequences are generated through a stochastic model of recombination using IGoR [28]. Each TCRαβ amino-acid sequence is then kept if its normalized hash (a hash is a deterministic but maximally disordered function) is ≤ q = qαqβ, so that a random fraction q of sequences passes selection. The values of qα and qβ are learned from rarefaction curves showing the number of unique amino-acid sequences of each chain as a function of the number of unique nucleotide sequences (S5 Fig), using the analytical expressions given in [17]. The predictions for the number of shared TCRαβ nucleotide sequences reported in Fig 5B, as well as the estimation of the sharing between the full repertoire of two individuals, are computed using the analytical expressions of [17]. If N sequences are sampled in m individuals, the expected number of sequences which will be found in exactly k individuals is:
M k , m ( N ) = ∫ 0 ∞ d p P ( p ) ( m k ) e N p ( m − k ) ( 1 − e − N p ) k (7)
Without selection P(p) is the probability density function for of sequences probabilities. We used this formula with p = Pgen/q for selected sequences, and p = 0 otherwise. The integral in Eq 7 is evaluated using a Monte Carlo simulation. Derivations and details about the Monte Carlo simulation can be found in [17]. We use this formula to estimate the proportion of full receptors shared between two individuals.
The results of [17] can also be used to estimate the theoretical proportion of clonotypes sharing a β in a sample of size N. This sharing is due to two phenomena: the possibility of generating twice the same β sequence and the division stage between the recombination of β and α. To simulate the first mechanism we can, following [17], generate an important number of β sequences (in-frame, no-stop codons) with IGoR, associate to each of them a hash between 0 and 1 and then only keep the sequences whose hash is lower than qβ to simulate the selection. The cellular division between β and α recombination creates 30 cells with the same β and different α. Some of these cells won’t have a functional α receptors, while others will not pass selection, while there is no precise way to quantify how many cells survive, we can consider an estimate of roughly nd ≈ 10 cells. Because the probability p(s) of generating a given sequence is so low, this increase in cell number multiplies p(s) by nd, hence corresponds to a change qβ → qβ/nd. Then, for 105 sequences and nd = 10, we find that ≈ 3% of clonotypes are expected to share their β sequence with another TCR.
|
10.1371/journal.pmed.1002143 | Orthostatic Hypotension and the Long-Term Risk of Dementia: A Population-Based Study | Orthostatic hypotension (OH) is a common cause of transient cerebral hypoperfusion in the population. Cerebral hypoperfusion is widely implicated in cognitive impairment, but whether OH contributes to cognitive decline and dementia is uncertain. We aimed to determine the association between OH and the risk of developing dementia in the general population.
Between 4 October 1989 and 17 June 1993, we assessed OH in non-demented, stroke-free participants of the population-based Rotterdam Study. OH was defined as a ≥20 mm Hg drop in systolic blood pressure (SBP) or ≥10 mm Hg drop in diastolic blood pressure (DBP) within 3 min from postural change. We furthermore calculated within participant variability in SBP related to postural change, expressed as coefficient of variation. Follow-up for dementia was conducted until 1 January 2014. We determined the risk of dementia in relation to OH and SBP variability, using a Cox regression model, adjusted for age; sex; smoking status; alcohol intake; SBP; DBP; cholesterol:high-density lipoprotein ratio; diabetes; body mass index; use of antihypertensive, lipid-lowering, or anticholinergic medication; and apolipoprotein E genotype. Finally, we explored whether associations varied according to compensatory increase in heart rate. Among 6,204 participants (mean ± standard deviation [SD] age 68.5 ± 8.6 y, 59.7% female) with a median follow-up of 15.3 y, 1,176 developed dementia, of whom 935 (79.5%) had Alzheimer disease and 95 (8.1%) had vascular dementia. OH was associated with an increased risk of dementia (adjusted hazard ratio [aHR] 1.15, 95% CI 1.00–1.34, p = 0.05), which was similar for Alzheimer disease and vascular dementia. Similarly, greater SBP variability with postural change was associated with an increased risk of dementia (aHR per SD increase 1.08, 95% CI 1.01–1.16, p = 0.02), which was similar when excluding those who fulfilled the formal criteria for OH (aHR 1.08, 95% CI 1.00–1.17, p = 0.06). The risk of dementia was particularly increased in those with OH who lacked a compensatory increase in heart rate (within lowest quartile of heart rate response: aHR 1.39, 95% CI 1.04–1.85, p-interaction = 0.05). Limitations of this study include potential residual confounding despite rigorous adjustments, and potentially limited generalisability to populations not of European descent.
In this population predominantly of European descent, OH was associated with an increase in long-term risk of dementia.
| Orthostatic hypotension is a common cause of transient cerebral hypoperfusion that is associated with subclinical brain disease, as well as increased risk of stroke.
Hypoperfusion has been implicated in the pathophysiology of dementia, but whether orthostatic hypotension affects the risk of dementia is uncertain.
Back in 1990, we measured orthostatic hypotension in 6,204 community-dwelling individuals participating in the Dutch population-based Rotterdam Study, and we followed these people for occurrence of dementia until 2014.
Having orthostatic hypotension at start of the study increased the risk of developing dementia over the next 25 years by 15%, with similar results for all-cause dementia and Alzheimer disease.
The risk of dementia was particularly increased when the orthostatic blood pressure drops were not compensated for by a sufficient increase in heart rate.
These findings suggest that transient cerebral hypoperfusion plays a role in the aetiology of dementia and that further studies are warranted to investigate the effects of hypoperfusion and treatment of orthostatic hypotension on markers of neurodegenerative disease and cognition.
| Cardiovascular health is now well-established as a key determinant in the prevention of dementia, including Alzheimer disease [1,2], but the mechanisms by which vascular damage leads to cognitive decline remain largely unknown. As cerebral hypoperfusion is widely implicated in dementia [3,4], cerebral haemodynamics have been suggested as a potential link between vascular risk factors and dementia [5]. Two important mechanisms for maintenance of proper and continuous cerebral perfusion are local vasoreactivity and autonomous nervous system function. Cerebral vasoreactivity has indeed been found to be associated with the risk of developing dementia in the general population [6], but the role of autonomous nervous system function in the onset of dementia has been less well studied.
Autonomic dysfunction may result in orthostatic hypotension (OH), which affects 20%–30% of the elderly population [7,8]. OH is characterised by a marked drop in blood pressure following postural change, insufficiently compensated for by sympathetic and parasympathetic mechanisms. This drop in blood pressure may elicit transient cerebral hypoperfusion, especially in the absence of a compensatory increase in heart rate. OH is associated with an increased risk of cardiovascular events, stroke, and mortality [9]. Moreover, OH is highly prevalent among patients with dementia and mild cognitive impairment, compared to healthy controls [10–13], but only one study has assessed the longitudinal relation between OH and the risk of dementia in initially healthy participants. In this Swedish population, OH was associated with an increased risk of dementia at re-examination after 6 y, but the investigators were unable to adjust for cardiovascular risk factors aside from hypertension, and attrition was substantial, with 37.5% of participants lost to follow-up between examination rounds [14]. These limited data regarding OH and cognition prompted a recent review and meta-analysis to conclude that longitudinal studies using standardised criteria are needed to elucidate whether OH is an independent risk factor for developing dementia [9,15]. We therefore aimed to determine the association between OH and the risk of dementia, in a long-term, ongoing population-based study.
The Rotterdam Study has medical ethics committee approval per the Population Study Act Rotterdam Study, executed by the Ministry of Health, Welfare and Sports of the Netherlands. Written informed consent was obtained from all participants. For the current study, the analysis plan was drafted in June 2015.
This study is embedded within the Rotterdam Study, a large ongoing population-based cohort study in the Netherlands, with an initial study population of 7,983 participants (78% of invitees) aged ≥55 y from the Ommoord area, a suburb of Rotterdam. The Rotterdam Study methods have been described in detail previously [16]. In brief, participants were interviewed at home and subsequently examined at the research centre for baseline assessment from October 1989 to July 1993. OH was determined during baseline assessment. Of 7,983 interviewed participants, 7,157 (89.7%) visited the research centre for the baseline physical examination. As of 2015, five follow-up examinations have been carried out.
Blood pressure and heart rate were measured using an automatic recorder (Dinamap, Critikon). The baseline blood pressure reading was the mean of two measurements on the right upper arm with the participant in supine position, after 5 min of rest. Measurements were repeated in the standing position after 1, 2, and 3 min. OH was defined as ≥20 mm Hg decrease in systolic blood pressure (SBP) or ≥10 mm Hg decrease in diastolic blood pressure (DBP) after postural change at any of the three measurements, in accordance with the Consensus Committee of the American Autonomic Society and the American Academy of Neurology [17,18]. We defined severity of OH by degree of blood pressure drop, i.e., ≥20/10 but <30/15, ≥30/15 but <40/20, and ≥40/20 mm Hg. We calculated continuous measures of blood pressure change in response to postural change, expressed as the coefficient of variation of within participant variability, defined as the ratio of the standard deviation (SD) to the mean of all measurements (i.e., measurements in supine and upright position combined). Furthermore, we determined the maximum increase in heart rate within 3 min after postural change. Directly afterwards, participants were asked whether they had felt unwell within the minutes following postural change.
Participants were screened for dementia at baseline and follow-up examinations using a three-step protocol [19]. Screening was done using the Mini-Mental State Examination (MMSE) and the Geriatric Mental State Schedule (GMS) organic level. Those with MMSE < 26 or GMS > 0 subsequently underwent examination and informant interview using the Cambridge Examination for Mental Disorders of the Elderly (CAMDEX). Additionally, the total cohort was continuously monitored for dementia through computerised linkage of medical records from general practitioners and the regional institute for outpatient mental healthcare with the study database. Available neuroimaging data were used when required for establishing a diagnosis. For all suspected cases of dementia, a consensus panel led by a consultant neurologist (P. J. K.) decided on the final diagnosis in accordance with standard criteria for dementia (DSM-III-R), Alzheimer disease (NINCDS-ADRDA), and vascular dementia (NINDS-AIREN). Follow-up until 1 January 2014 was near complete (94.0% of potential person years), and participants were censored within this follow-up period at date of dementia diagnosis, date of death, date of loss to follow-up, or 1 January 2014, whichever came first.
We assessed smoking status (i.e., current, former, never), alcohol intake, use of antihypertensive medication, use of lipid-lowering medication, and use of anticholinergic medication at baseline by interview. Anticholinergic medication included antipsychotic and antidepressant medication, but also drugs prescribed against parkinsonism, urinary incontinence, or obstructive pulmonary disease that can have anticholinergic side effects. Fasting serum lipid levels were measured at baseline. Hypertension was defined as the use of antihypertensive medication and/or elevated systolic or DBP (>140/90 mm Hg). Body mass index (kg/m2) was computed from measurements of height and weight. Diabetes mellitus was defined as the use of blood-glucose-lowering medication at baseline or a random serum glucose level ≥11.1 mmol/l [20]. Myocardial infarction and atrial fibrillation were assessed by direct questioning and presence of abnormalities on a 12-lead electrocardiogram as determined by study physicians and reviewed by a cardiologist. Heart failure was determined using a validated score, similar to the definition of heart failure of the European Society of Cardiology [21]. APOE genotype was determined using polymerase chain reaction on coded DNA samples.
Analyses included all non-demented, stroke-free participants attending the study centre for examination. Of 7,157 participants attending the study centre, 531 were ineligible due to prevalent dementia (n = 312), stroke (n = 168), or both (n = 51). Missing covariate data (maximum 17.6%), except for APOE genotype, were imputed using 5-fold multiple imputation, based on determinant (presence of OH and postural SBP variability), outcome, and included covariates. Distribution of covariates was similar in the imputed versus non-imputed dataset.
We determined the association between presence of OH and incident dementia, using Cox proportional hazard models. We repeated the analysis with dementia or death as a joint outcome measure, to reduce selection due to competing risk (upon referee’s request). Subsequently, we analysed categories of increasing severity of orthostatic blood pressure drop, and OH with and without feeling unwell. Because of right-skewedness, we performed a natural logarithmic transformation of SBP variability to obtain a roughly normal distribution (mean −2.52, SD 0.58). Z-scores were computed by dividing the difference between the individual value and the population mean by the population SD. We determined the association between SBP variability related to postural change and incident dementia, per quartile and continuously per SD increase, using a Cox model. To eliminate a paradoxical impact of high blood pressure variability in those with excessive increases, we repeated analyses after excluding those with a ≥20 mm Hg increase in SBP or ≥10 mm Hg increase in DBP within 3 min (upon referee’s request). Furthermore, we determined whether associations extended to those without a formal diagnosis of OH. We then assessed whether the risk of dementia in relation to orthostatic blood pressure drops was modified by response in heart rate after postural change, by testing for multiplicative interaction in the above Cox model and providing risk estimates of OH for dementia per quartile of response in heart rate. We verified that the proportional hazard assumption was not violated in these models by plotting the partial (Schoenfeld) residuals against follow-up time.
All analyses were adjusted for age and sex (model I), and additionally in a second model for smoking status, alcohol intake, systolic and diastolic blood pressure, use of antihypertensive medication, ratio of serum total cholesterol to high-density lipoprotein, use of lipid-lowering medication, diabetes mellitus, body mass index, use of anticholinergic medication, and APOE genotype (model II).
We repeated the analyses for Alzheimer disease and vascular dementia separately, after censoring participants at time of incident stroke, after excluding those with Parkinson disease at baseline, after excluding those with heart disease (i.e., coronary heart disease, heart failure, atrial fibrillation; upon referee’s request), and after excluding those with possible postural tachycardia syndrome (defined as a ≥30 beats per minute increase in heart rate or any heart rate of ≥120 beats per minute). Finally, we performed several sensitivity analyses: (1) for men and women separately, (2) for persons above and below the median age (68.5 y), (3) excluding the first 5 y of follow-up to assess for reverse causality, (4) for those with and without heart failure at baseline, (5) for those with and without a history of hypertension, (6) distinguishing use of antihypertensive drugs, and (7) for those with and without diabetes (upon referee’s request). Finally, we repeated analyses for OH, SBP variability, and dementia risk in a subset of participants without significant comorbidity (i.e., excluding those with heart disease, Parkinson disease, and diabetes; upon referee’s request).
All analyses were done using IBM SPSS Statistics version 23.0. Alpha level (type 1 error) was set at 0.05.
Of 6,626 eligible participants, 6,303 (95.1%) underwent examination for OH. No baseline blood pressure measurement was obtained in eight individuals, and no measurement after postural change in 91 individuals, leaving a total of 6,204 (93.6%) cases for analysis. Baseline characteristics of participants are shown in Table 1.
Overall, 1,152/6,204 (18.6%) participants had OH. The prevalence of OH steeply increased with age, to 30.6% of those aged ≥75 y. Although prevalence in the elderly was similar for men and women, there was a slightly higher prevalence in women at younger ages (Fig 1). Of all patients with OH, 160 (13.9%) reported feeling unwell along with their blood pressure drop.
During a median follow-up time of 15.3 y (interquartile range 8.3–20.8), 1,176 individuals developed dementia, of whom 935 (79.5%) were diagnosed with Alzheimer disease, 95 (8.1%) with vascular dementia, 43 (3.7%) with Parkinson dementia, and 30 (2.6%) with another type of dementia, and in 73 (6.2%) no definite subdiagnosis could be made. Of all incident dementia cases, 129 were preceded by a stroke a median 3.7 y (interquartile range 1.2–7.2) before diagnosis of dementia.
OH at baseline was associated with an increased risk of dementia during follow-up (adjusted hazard ratio [aHR] 1.15, 95% CI 1.00–1.34, p = 0.05; Table 2). Similarly on a continuous scale, variability in SBP related to postural change was associated with an increased risk of dementia (aHR per SD increase: 1.08, 95% CI 1.01–1.16, p = 0.02). This association was similar when excluding those who fulfilled the formal criteria for OH (aHR 1.08, 95% CI 1.00–1.17) and unaltered by excluding those with a marked increase in blood pressure following postural change (see S1 Table). Results were similar for Alzheimer disease only. For vascular dementia, we observed higher risk estimates with OH than for Alzheimer disease in the age- and sex-adjusted model (aHR 1.53, 95% CI 0.97–2.43), but these were largely explained by cardiovascular risk factors, so that fully adjusted estimates were similar to those for Alzheimer disease (aHR 1.20, 95% CI 0.73–1.96; Table 2). We did not observe a clear exposure-response relation for severity of OH, because of lower effect estimates for participants with the most severe blood pressure drops (Fig 2). In contrast, the risk of dementia strongly increased per quartile of blood pressure variability (Fig 2; for a full table with risk estimates per quartile and per SD, see S2 Table). Risk estimates were similar when modelling dementia or death as a joint outcome (OH: aHR 1.17, 95% CI 1.08–1.27; blood pressure variability: aHR 1.08, 95% CI 1.04–1.12). Estimates for both OH and SBP variability were attenuated when incorporating these simultaneously in a model (OH: aHR 1.07, 95% CI 0.90–1.27; blood pressure variability: aHR 1.06, 95% CI 0.99–1.14).
Results for OH where individuals reported feeling unwell along with the blood pressure drop were similar to results for OH without feeling unwell (aHR 1.20, 95% CI 0.86–1.66, versus aHR 1.15, 95% CI 0.98–1.34, respectively). The risk of dementia related to OH was most profound in participants who lacked a compensatory increase in heart rate (aHR for lowest quartile of heart rate response 1.39, 95% CI 1.04–1.85, p-value for interaction = 0.05; Fig 3). This risk was similar after excluding all participants taking beta-blockers (see S2 Table).
Sensitivity analyses showed similar results after censoring for incident stroke, excluding participants with prevalent Parkinson disease, excluding those with possible postural tachycardia syndrome, or omitting the first 5 y of follow-up (Table 3). A history of hypertension or use of any antihypertensive medication did not modify the risk of dementia associated with OH (Table 3). Amongst 177 participants with heart failure at baseline, risk estimates for OH were higher than in those without heart failure, albeit not statistically significantly (aHR 1.52, 95% CI 0.63–3.66, p-value for interaction = 0.07). Risk estimates for both OH and SBP variability were slightly stronger when excluding participants with cardiac disease, neurodegenerative comorbidity, or diabetes all together (see S3 Table).
In this large population-based study, OH was present in nearly one in five participants and was associated with a 15% increase in long-term risk of dementia. The risk of developing dementia was highest in those with OH lacking a compensatory increase in heart rate. Similarly, higher variability in blood pressure related to postural change was associated with a higher risk of dementia, even in individuals without a formal diagnosis of OH.
Prevalence of OH in our study was high and increased steeply with age, in line with previous studies among individuals within a similar age range in the community [7,8]. A few studies have investigated OH in relation to cognitive test performance. In the ARIC study, OH was associated with decline on two cognitive tests, but this decline was largely explained by cardiovascular risk factors [22]. Two smaller studies found no overall association between OH and decline on the MMSE after 2 y [7,23]. Conversely, OH was found to increase the risk of conversion from mild cognitive impairment to dementia after 3 y [24], as well as the risk of dementia in patients with Parkinson disease [25]. Only one other study has assessed the relation between OH and the risk of dementia in initially healthy individuals. In a sample of 1,480 individuals in the Swedish general population, OH was associated with the risk of having dementia at re-examination after 6 y [14]. However, the investigators were unable to fit survival models or adjust for cardiovascular risk factors aside from hypertension, and attrition was substantial, with 37.5% of participants lost to follow-up [14]. We found OH to be associated with long-term risk of dementia on continuous follow-up, independent of various other risk factors.
The most apparent explanation for our findings is that OH causes brain damage due to recurrent transient cerebral hypoperfusion. Autonomic nervous system function is responsible for maintaining continuous cerebral perfusion, together with local vasoreactivity, which has previously been associated with the risk of dementia in the general population [6]. Brief episodes of hypoperfusion elicited by sudden blood pressure drops may lead to hypoxia, with detrimental effects on brain tissue via, for instance, neuroinflammation and oxidative stress [26]. These mechanisms have been suggested to be of particular relevance in the pathogenesis of small vessel disease [27], and orthostatic blood pressure drops in patients with dementia have been associated with deep white matter and basal ganglia hyperintensities [28], albeit not with overall white matter lesion volume [29]. The reduction in cerebral blood flow with autonomic failure has also been reported to predominantly affect the hippocampus [30], possibly linking hypoperfusion to early Alzheimer pathology. Another potential explanation for our findings might be that OH serves as a marker of other detrimental consequences of autonomic dysfunction, such as blood pressure variability [31,32], response to Valsalva manoeuvre [13,33], cardiovascular reflex and heart rate variability [34,35], and 30/15 ratio [35]. Several of these measures may be linked to dementia via hypoperfusion, but other pathways could be involved also. For instance, decreased arterial wall compliance with hypertension likely contributes to OH by diminishing baroreceptor sensitivity [36]. Arterial stiffness and OH are both associated with increased burden of cerebral white matter lesions and vascular disease including stroke [9,28,37,38]. As these are established risk factors for dementia, these conditions may act as mediators or indicate shared aetiology between diseases. Nevertheless, we did not see risk estimates for all-cause dementia attenuated after adjustment for cardiovascular risk factors, censoring for clinical stroke, or restricting the population to those without cardiac or neurodegenerative comorbidities. Alternatively, sympathetic failure can occur with diabetic neuropathy, and although we adjusted for clinical history of diabetes, some patients with impaired fasting glucose might already have had autonomic derailment together with subclinical small vessel disease. We had no other direct measures of autonomic dysfunction to ascertain whether OH is the main driving force behind our findings. Blood pressure variability in our study was measured following postural change and therefore reflects a distinct orthostatic response, although it might in part occur as a manifestation of wider autonomic failure, unrelated to postural change. Accordingly, risk estimates for both OH and blood pressure variability attenuated after incorporating both in the same model, but it remains unclear whether this is because they are both computed from postural-change-related measurements or because they reflect the same underlying autonomic dysfunction. Given the interaction we found between OH and the lack of a compensatory increase in heart rate, the impact of OH may well vary with the degree of overall autonomic failure, and future studies are encouraged to incorporate various measures of autonomic dysfunction collected in the same individuals simultaneously (for which reported risks and effect estimates may serve as a guidance). Although autonomic dysfunction may even reflect early signs of neurodegeneration, the long follow-up duration of our study renders reverse causality less likely.
The risk of dementia associated with OH in our study was independent of whether participants reported feeling unwell along with the blood pressure drop, and the vast majority of patients with OH did not have symptoms during testing. Although for blood pressure variability we observed an exposure-response association, we did not find this for severity of OH itself. As OH is also associated with mortality [9], this finding may be attributable to competing risk, causing the most severely affected participants to die at a younger age, prior to developing dementia. Alternatively, rather than the degree of blood pressure drop, the lack of compensatory increase in heart rate may better reflect the severity of consequences of orthostatic drops in blood pressure. Taken together, our findings suggest that formal assessment of OH is necessary to have sufficient test sensitivity, and incorporation of heart rate response in the definition of OH may contribute to evaluating aetiology and clinical severity. Hypotension might be harmful even without accompanying clinical symptoms such as light-headedness. This lack of symptoms with orthostasis was previously observed in patients with dementia [39] and may warrant caution in view of studies linking low blood pressure in late life to cognitive decline and dementia [40].
OH most commonly arises due to autonomic dysfunction in the absence of neurological disease, but may be provoked by synucleinopathies (e.g., Parkinson disease), small fibre peripheral neuropathy, volume depletion (e.g., due to diuretics), and diminished cardiac pump function. In addition, several drugs can cause or aggravate OH, including antihypertensive agents and antidepressants. Participants in our study with heart failure at baseline seemed particularly affected by OH, possibly due to the lack of a compensatory increase in stroke volume. OH has been associated with the development of structural cardiac changes, including left ventricular hypertrophy [41], which may function as a mediator towards dementia [42,43]. However, the subgroup of participants with heart failure in our study was too small to draw any firm conclusions. We found similar associations between OH and dementia after excluding those with Parkinson disease, and in users versus non-users of antihypertensive medication.
Although we believe our findings are valid, there are certain limitations to our study to take into account. First, measures of OH were not available for all eligible participants. Although this was largely due to logistic reasons, we cannot completely rule out selection bias. Second, despite adjustment for many potentially confounding factors, residual confounding may still occur. However, given the lack of attenuation of our results in our second, more fully adjusted model, residual confounding is unlikely to result from the most relevant, included covariates. Third, we continued blood pressure measurements for up to 3 min after postural change, and while this approach is in line with international guidelines, it may have resulted in missed orthostatic blood pressure drops beyond this time window [44]. However, any misclassification (i.e., missed diagnosis of OH) would likely have led to underestimation of the true effect. Fourth, subtypes of dementia were based on clinical diagnosis, and mixed pathology (e.g., Lewy bodies) in patients with clinical Alzheimer disease may contribute to the observed associations. Fifth, we were unable to adjust for the fact that OH predisposes for falls, which may contribute to cognitive decline due to traumatic brain injury. Finally, the majority of our study population was of European descent, and findings may not be applicable to other ethnicities.
In conclusion, OH is associated with an increased risk of dementia in the general population. This finding supports an important role for maintaining continuous cerebral perfusion in the prevention of dementia.
|
10.1371/journal.pgen.1002613 | The Functions of Mediator in Candida albicans Support a Role in Shaping Species-Specific Gene Expression | The Mediator complex is an essential co-regulator of RNA polymerase II that is conserved throughout eukaryotes. Here we present the first study of Mediator in the pathogenic fungus Candida albicans. We focused on the Middle domain subunit Med31, the Head domain subunit Med20, and Srb9/Med13 from the Kinase domain. The C. albicans Mediator shares some roles with model yeasts Saccharomyces cerevisiae and Schizosaccharomyces pombe, such as functions in the response to certain stresses and the role of Med31 in the expression of genes regulated by the activator Ace2. The C. albicans Mediator also has additional roles in the transcription of genes associated with virulence, for example genes related to morphogenesis and gene families enriched in pathogens, such as the ALS adhesins. Consistently, Med31, Med20, and Srb9/Med13 contribute to key virulence attributes of C. albicans, filamentation, and biofilm formation; and ALS1 is a biologically relevant target of Med31 for development of biofilms. Furthermore, Med31 affects virulence of C. albicans in the worm infection model. We present evidence that the roles of Med31 and Srb9/Med13 in the expression of the genes encoding cell wall adhesins are different between S. cerevisiae and C. albicans: they are repressors of the FLO genes in S. cerevisiae and are activators of the ALS genes in C. albicans. This suggests that Mediator subunits regulate adhesion in a distinct manner between these two distantly related fungal species.
| In this study, we compared the roles of Mediator, a central transcriptional regulator in all eukaryotes, between the pathogenic fungus Candida albicans and the non-pathogenic model yeasts Saccharomyces cerevisiae and Schizosaccharomyces pombe. We discovered that Mediator has both shared and species-specific functions in the three yeasts. The shared functions include regulation of genes required for cell separation after cell division by the Middle domain subunit Med31. The species-specific functions include transcriptional regulation of the cell wall adhesins, which play key roles in the pathogenesis of C. albicans. In C. albicans, the Mediator subunits Med31, Med20, and Srb9/Med13 are activators of the ALS cell wall adhesins. In S. cerevisiae, our results and previous reports suggest an opposite, repressive role in the expression of the FLO genes and in adhesion-dependent phenotypes. The C. albicans Med31, Med20, and Srb9/Med13 contribute to processes highly important for disease: the switch to filamentous morphology and biofilm formation. Moreover, Med31 impacts on virulence in an invertebrate infection model. Our study has implications for understanding the regulation over virulence-associated genes in C. albicans and the roles of a key transcriptional regulator in this process.
| The transcription factor complex Mediator is associated with RNA polymerase II and it has essential roles in transcription ([1], reviewed in [2]). The yeast Mediator is composed of 25 subunits, which are structurally and functionally organized into four modules [3]–[8]. The core complex is comprised of the Head, Middle and Tail domains [3]–[6]. A fourth, Kinase domain is associated with Mediator under some conditions ([9]–[11]; reviewed in [2]). The core Mediator has a positive role in transcription, while the Kinase domain mainly functions in repression [2].
The roles of Mediator in transcription are complex [2], [12]. Mediator interacts with gene-specific transcription factors and RNA polymerase II and mediates polymerase-activator interactions and formation of the pre-initiation complex (reviewed in [2], [12], [13]). In addition to activated transcription, Mediator also stimulates basal transcription [1], [14], [15]. Further proposed roles for Mediator are in post-initiation steps [12], [16]–[19], re-initiation during multiple rounds of transcription [20] and regulation of chromatin structure [12], [21], [22]. Two recent reports showed an additional role for the core Mediator in sub-telomeric gene silencing [23], [24]. In addition to these versatile roles in gene transcription, Mediator also appears to be a central “integrative hub” for the regulation of gene expression by physiological signals [12]. Examples from yeast include regulation of the Kinase domain by the Ras/PKA pathway via phosphorylation of the Srb9/Med13 subunit [25], and control over the expression of iron-responsive genes by an interplay between the Tail subunit Med2, which has a positive role, and the Kinase domain that phosphorylates Med2 to inhibit its function [7], [26].
The multisubunit Mediator complex emerged early in the evolution of eukaryotes, and the versatility of its functions and its role as an integrative platform for cell physiology could have contributed to the shaping of gene expression programs in different species, for adaptation to specific environments and life styles [27]. Fungi represent an excellent model system for exploring these questions. A comparative analysis in model yeasts Saccharomyces cerevisiae and Schizosaccharomyces pombe showed remarkable conservation of the roles of Mediator in spite of the fact that these two yeasts are highly divergent [28]. The conserved functions include a broad role in stress responses, and specific, distinct roles of the Mediator domain in the regulation of cell wall dynamics and cell morphology. The Head and Middle domains of Mediator are required for the expression of the cytokinesis genes under the control of the transcription factor Ace2 [23], [28], while the Kinase domain represses transcription of the cell wall adhesins [7], [9], [25], [28], [29]. Furthermore, studies of the Srb11/Ssn8 cyclin subunit of the Mediator Kinase domain in human and plant fungal pathogens (Cryptococcus neoformans, Candida albicans, Fusarium vertisilloides and Fusarium gramineaurum) suggest conserved roles in the repression of nutrient responsive functions and genes required for the production of toxins and pigments, as well as a conserved role in stress responses and regulation of cell wall integrity [9], [30]–[37]. Moreover, in Candida glabrata, the Mediator Tail subunit Med15/Gal11 plays a conserved role with S. cerevisiae in drug resistance mediated by the transcription factor Pdr1 [38], [39].
Here we report the first study of Mediator functions in the human pathogen C. albicans. We show that the C. albicans Mediator has some conserved functions with S. cerevisiae and Schizo. pombe, but also has additional roles in the expression of virulence-related genes, most notably the ALS adhesins. Phenotypic analysis showed roles for Mediator subunits in phenotypes of C. albicans associated with pathogenesis – filamentous growth and biofilm formation. Our data presented here and previous reports [25], [40] show that control of the cell wall adhesins by Mediator subunits Med31 and Srb9/Med13 differs between S. cerevisiae and C. albicans, suggesting distinct Mediator-dependent control of adhesion in these two yeast species.
To start delineating the function of the Mediator complex in C. albicans, we made homozygous deletion mutants in the Middle domain subunit Med31. In C. albicans, Med31 is encoded by orf19.1429 and it displays 48.2% and 39.80% sequence identity with its S. cerevisiae and Schizo. pombe orthologs respectively. Transcriptome-wide profiles of the med31ΔΔ mutant were obtained and compared to those of a complemented med31ΔΔ+MED31 strain. Routine manipulations during mutant strain construction can result in gross chromosomal rearrangements, such as aneuploidies [41], which could profoundly affect the results of transcriptome analysis. Inspection of the med31ΔΔ transcriptional profile in the chromosomal context did not reveal a colour distribution associated with aneuploidies (Figure S1A), indicating that the med31ΔΔ mutant has the same chromosomal structure as the complemented strain. Additionally, gene sets representing 50 kb fragments were included in the Gene Set Enrichment Analysis (see below), and this analysis also did not reveal any gross chromosomal alterations (data not shown).
In agreement with a general role for Med31 in gene expression, and consistent with data from S. cerevisiae and Schizo. pombe [7], [28], [42], [43], 7.8% of the genome (510 genes) was differentially expressed in the absence of MED31 (cut-off of 1.5 fold, p<0.05, Dataset S1). Out of the genes differentially expressed in med31ΔΔ cells, 61.7% (315) were down-regulated and 38.2% (195) were up-regulated. This is consistent with a predominantly positive role of Med31 in transcription, and is in line with reports in S. cerevisiae and Schizo. pombe [7], [28], [42], [43]. To reveal the cellular pathways regulated by Med31 in C. albicans, we performed Gene Set Enrichment Analysis (GSEA) [44], [45]. GSEA compares a list from the transcript profile of interest created by ranking all of the genes according to the change in their expression (in this case that of a med31ΔΔ mutant) to a predefined gene set, and asks if a specific gene set is enriched in the top (up-regulated genes) or the bottom (down-regulated genes) of the ranked list [44], [45]. A ranked list of genes from the transcript profile of med31ΔΔ cells was compared to a custom database of 8123 gene sets (http://candida2.bri.nrc.ca/andre/GSEA/index.cfm; Sellam and Nantel, submitted) constructed using GO annotations and protein interaction data from CGD (PMID: 19808938), SGD (http://www.yeastgenome.org) and BioGRID [46], most currently published C. albicans transcriptional profiling and ChIP-CHIP experiments, our own TF motif database (PMID: 18342603), and S. cerevisiae genetic-association data (PMID: 20093466). Since profiles can exhibit correlations with hundreds of overlapping gene sets, significantly enriched gene sets (p<0.005, FDR<25%) were further organized and visualized using the Cytoscape: Enrichment Map plug-in (PMCID: PMC2981572), which produces networks of gene sets that share significant overlaps with each other (Figure 1A shows the most prominent networks of genes; the complete network is shown in Figure S2 where the details can be visualised by using the “zoom in” function in the pdf document). Figure 1B shows examples of enrichment plots for selected gene sets. The complete GSEA output can be found at http://dl.dropbox.com/u/7211133/Med31%20GSEA%20Results.zip.
GSEA detected enrichment for nucleolar functions, rRNA and ribosome biogenesis genes, and genes involved in nucleotide biosynthesis in the set of genes down-regulated in the med31ΔΔ mutant, while genes required for mitochondrial function were up-regulated (Figure 1A and 1B). Enrichment was also found in gene sets important for virulence-promoting function in C. albicans. Those include genes differentially expressed during C. albicans-host interactions with mouse macrophages [47], reconstituted human oral epithelial cells [48] and polymorphonuclear leukocytes [49], as well as genes differentially expressed in conditions which alter cellular morphogenesis, such as the induction of hyphal growth [50], [51], mutations in the Ras-cAMP morphogenesis pathway (ras1 and cdc35/cyc1) [52], and inhibition of cell cycle progression that causes pronounced polarised growth (treatment with hydroxyurea or down-regulation of the polo-like kinase CDC5) [53] (Figure 1A and 1B). Genes expressed at the G1-S and S-G2 transition of the cell cycle [54] were up-regulated in the med31ΔΔ mutant, as were those required for membrane and cell wall biosynthesis (Figure 1A and 1B). Genes required for cytokinesis were down-regulated (as shown by the black arrow in Figure 1A, and in the enrichment plot in Figure 1B).
Modulation of several gene sets enriched in the med31ΔΔ mutant, for example down-regulation of genes required for protein synthesis and up-regulation of those required for cell wall biogenesis, is part of a more general stress response in C. albicans [55]. Mediator has been previously implicated in stress responses in yeasts [7], [28], and it is therefore possible that some of the differences in the med31ΔΔ transcriptome are due to activation of stress responses upon loss of Med31 function. However, our analysis indicates that this is unlikely to be the cause for much of the differential gene expression in the mutant. There was little correlation between the med31ΔΔ transcriptional profile and our large database of transcriptional profiles produced from stressed cells when analysed by GSEA (of note, profile to profiles comparisons such as those done by GSEA tend to produce the strongest correlations and therefore if a correlation existed it is very likely that it would have been detected by GSEA). We further used scatter plots to directly compare the med31ΔΔ profile with stressful conditions, such as osmotic or oxidative stress, and these comparisons confirmed lack of extensive correlation between the med31ΔΔ transcriptome and differential gene expression upon stress (Figure S1B).
Analysis of gene ontology terms using the GO term finder tool at the Candida Genome database and the genes up- or down-regulated by at least 1.5 fold in med31ΔΔ cells (see Dataset S1) confirmed that genes related to morphogenesis, mitochondrial function and the cell wall were differentially expressed (Table 1 and Dataset S2). Interestingly, several central regulators of filamentous differentiation, such as the transcription factors Tec1, Efg1, Cph1 and Nrg1, were amongst the down-regulated genes, as were six out of the eight genes from the FGR6 (Filamentous Growth Regulator) family located in the RB2 repeat sequence (Dataset S2). The FGR6 family is one of the gene families found to be enriched in pathogenic yeast species, and specifically expanded in C. albicans [56]. While GSEA scored the cell wall gene set as up-regulated, we noticed that there were also several genes in this group that appeared at the bottom of the list, in the down-regulated group. In fact, another Candida-specific gene family expanded in pathogens was down-regulated in the med31ΔΔ mutant, that encoding the ALS cell wall adhesins [56] (Table 2). The major C. albicans adhesin ALS1 was one of the most down-regulated genes in the mutant (5 fold down-regulation, Table 2 and Dataset S1). ALS5 and ALS6 were also down-regulated (Table 2), but of note, these genes are expected to be expressed at very low levels in the wild type. Additionally, several other genes encoding cell wall proteins were down-regulated in the mutant, as were genes necessary for cell wall construction and remodelling, in particular those required for cytokinesis and regulated by the transcription factor Ace2 (e.g. the chitinase CHT3 and the endoglucanase ENG1) [57] (Table 2; notably GSEA also scored the cytokinesis genes as down-regulated and this is shown in Figure 1). The existence of several down-regulated cell wall genes indicates that the up-regulation of genes with roles in cell wall integrity that is detected in the med31ΔΔ mutant (Figure 1 and Table 2) likely reflects a compensatory feedback regulation due to a defective cell wall structure in the absence of Med31. That med31ΔΔ mutants have altered cell walls is supported by phenotypic analysis demonstrating changes in sensitivity to the cell wall targeting drugs congo red and calcofluor white (Table 3).
In conclusion, the transcriptome analysis indicated a broad role for Med31 in cell physiology in C. albicans, with functions in morphogenesis and cell cycle progression, growth and metabolism, cell wall integrity, the expression of the cytokinesis genes under the control of the transcription factor Ace2 and those regulated by the interaction of C. albicans with host cells. Finally, two gene families enriched in pathogenic yeasts, the FGR family of filamentous growth regulators and the ALS adhesins, required Med31 for wild type expression levels.
Down-regulation of Ace2 target genes in the absence of Med31 in C. albicans is in agreement with a role for Med31 and other Mediator subunits in Ace2-dependent gene expression that is conserved between C. albicans, Schizo. pombe and S. cerevisiae (this study and [28]). Comparing more broadly the genes affected in med31ΔΔ cells with those reported to be differentially expressed in the C. albicans ace2 mutant [57] revealed that differential expression of 35 genes is shared between these two transcription factors (Table S1). Genes involved in cytokinesis and cell wall functions were predominant in the shared “down-regulated” group, whereas mitochondrial biogenesis genes were predominant in the shared “up-regulated” group (Table S1). This analysis suggests that the functions of Med31 in cell wall integrity and metabolism are mediated, at least in part, by Ace2-dependent roles.
To explore this further, we used quantitative PCR (qPCR) to directly compare the expression levels of candidate genes in the med31ΔΔ and ace2ΔΔ mutants (Figure 2A). Under yeast growth conditions (as was done for the transcriptome analysis) the mRNA levels for the chitinase CHT3 and the cell wall proteins PIR1, EAP1 and ALS1 were reduced in both med31ΔΔ and ace2ΔΔ cells (albeit to a different degree). CHT3 and PIR1 have been previously reported as Ace2-targets, while ALS1 and EAP1 were not [57]. The expression of the transcription factor TYE7 was reduced in med31ΔΔ cells (consistent with the microarray data), but not in the ace2ΔΔ mutant.
In C. albicans, the expression of cell wall proteins is activated upon filamentous differentiation [50], [51], [58]–[61]. This includes ALS1 and other ALS and non-ALS adhesins. Therefore, we next tested if Med31 was required for the expression of the ALS1, and two hypha-specific adhesins ALS3 and HWP1, during filamentous growth. For these experiments cells were grown in filament-inducing Spider media at 37°C. The mRNA levels of both ALS1 and ALS3 were down-regulated in med31ΔΔ cells (Figure 2A). Ace2 was not required for ALS1 expression during hyphal growth, while the levels of ALS3 were lower in cells lacking Ace2, but the effect was less than in the absence of Med31 (Figure 2A). HWP1 was up-regulated in both mutants (1.8–3 fold) (Figure 2A). Collectively, the qPCR analysis suggests that Med31 and Ace2 have common, but also independent roles in the expression of the cytokinesis genes and the cell wall adhesins during yeast and hyphal growth.
Given that we found novel cell wall protein targets that require Ace2 for wild type expression (ALS1, ALS3 and EAP1), we searched the upstream regulatory regions of these genes for putative Ace2 binding sites as defined in [57]. There are three Ace2-binding motifs within 1.5 kb upstream of the start codon for ALS3 (Figure 2B). No motifs that strictly conform to the consensus sequence were found in the promoters of ALS1 and EAP1, although variant motifs could be found (data not shown). The motifs in the ALS3 promoter included one at −467 bp, which is in a region found to be essential for ALS3 activation in hyphae (the so-called A1 region) [59]. This suggests that ALS3 could be a direct target of Ace2.
We next performed phenotypic analysis of the C. albicans med31ΔΔ mutant to address the biological relevance of the observed changes in gene expression. Cells lacking Med31 displayed a cytokinesis defect (Figure 3A), consistent with a role in Ace2-dependent gene expression. ∼40% of cells from med31ΔΔ cultures showed a cell-chain phenotype typical of mutants that cannot undergo cytokinesis. This phenotype was observed in two independently constructed homozygous deletion mutants and was partially complemented by re-introduction of a wild type copy of MED31 into the mutant genome (Figure 3A and 3B).
The med31ΔΔ mutant also displayed phenotypes consistent with altered cell membrane and cell wall integrity that were suggested by transcriptome analysis. The mutant was sensitive to formamide, SDS, the sterol-binding antifungal drug nystatin, DMSO and growth at 16°C, all phenotypes consistent with defective membrane structure. The mutant was also sensitive to the cell wall-targeting drug congo red, but more resistant to the chitin-binding dye calcofluor white (Table 3 and Figure S4). Furthermore, the med31ΔΔ mutant was sensitive to oxidative and salt stress and ethanol (Table 3 and Figure S4). Some of these phenotypes are also observed in S. cerevisiae and Schizo. pombe med31 mutants, suggesting conserved roles [28], [43]. We also tested the C. albicans ace2ΔΔ mutant side by side with med31ΔΔ for tolerance to the various compounds (Table 3 and Figure S4). Resistance to calcoflour white was also observed in the ace2ΔΔ mutant (Table 3 and [62]), as was a mild sensitivity to nystatin and growth at 16°C, indicating that these phenotypes could be due to the role of Med31 in Ace2-dependent gene expression. However, the other sensitivities of the med31ΔΔ mutant were not shared by the ace2ΔΔ mutant, and are therefore unrelated to Ace2-dependent phenotypes.
The transcriptome analysis indicated a role for Med31 in cellular morphogenesis and we therefore tested whether Med31 was necessary for the yeast-to-hypha morphogenetic switch in response to a variety of inducers in vitro. The med31ΔΔ mutant was unable to produce filaments on solid Spider and M199 media, or on plates containing N-acetylglucosamine (Figure 3C). The mutant was also compromised for filamentation in liquid media, however to a lesser extent than on plates (Figure S5). med31ΔΔ cells could filament in response to serum and in M199 media in culture, but with delayed kinetics and with a proportion of cells remaining in yeast form (Figure S5). In Spider media the mutant had a more pronounced defect, and even after 7 h cells were still largely in yeast form (Figure S5). However, after prolonged incubation of 12 h, filaments were observed in this medium also (data not shown). To address whether the role for Med31 in filamentous growth would be important in a disease context, we further tested the ability of med31ΔΔ cells to filament in vivo, using the C. albicans-Caenorhabditis elegans infection model [63], [64]. This is a well-established host-pathogen system, which recapitulates key elements of disease as seen in vertebrates, most notably, the requirement for filamentous growth [63]. Only 16.5% of the worms infected with the mutant developed filaments after 3 days compared to 49% for worms infected with wild type C. albicans or the complemented strain (Figure 4A and 4B). Moreover, for those med31ΔΔ-infected worms that did develop filaments, there was a significant delay in filamentation, there were many fewer filaments per worm compared to the wild type or complemented strains, and the filaments were much shorter (Figure 4A; a similar phenotype was observed even after 7 days of infection, Figure S6). We also evaluated the pathogenic potential of the med31ΔΔ mutant in the worm. The kinetics of killing of the worm by C. albicans was delayed in worms infected with the med31ΔΔ mutant compared to those infected with the wild type (p<0.02 for all experiments performed) (Figure 4C). This result supports a role for Med31 in C. albicans virulence.
We next sought to show that the regulation of the adhesins by Med31 was biologically relevant. To that end, we tested the ability of the med31ΔΔ mutant to form biofilms, as both ALS1 and ALS3 are key adhesins for biofilm formation by C. albicans [65]. Biofilm formation was tested in vitro, on serum-coated silicone disks as previously described [66]. In the absence of Med31, biofilms were severely compromised in both density and depth, as determined by confocal scanning laser microscopy (Figure 5A). Scanning electron microscopy (SEM) showed a similar defect (Figure S7). Quantitative analysis confirmed the phenotype observed by microscopy (Figure 5B). The med31ΔΔ mutant displayed a strong defect at the earliest time point of 90 minutes (adherence stage). The med31ΔΔ mutant grew somewhat slower than the wild type in planktonic conditions (Figures S4 and S8), however at the 90 minutes time point in the biofilm formation assay we did not observe significant cell growth for any of the strains, including the wild type (data not shown), strongly suggesting that the defect in the med31ΔΔ mutant is due to an adherence defect and not to the observed slower growth.
Rescue by over-expression of target genes has been previously used as a strategy for identification of transcription factor targets relevant for biofilm formation (for example in the C. albicans bcr1 mutant [65]). We employed a similar strategy to test whether ALS1 is a relevant Med31 target gene for biofilm formation. The idea is that, if lower expression of ALS1 in the med31ΔΔ mutant is contributing to the biofilm defect, its ectopic expression under a constitutive promoter should, at least in part, rescue biofilm formation by med31ΔΔ cells. As shown in Figure 5B, expression of ALS1 in the med31ΔΔ mutant under the constitutive TEF1 promoter led to a substantial rescue of the biofilm defect at all time points, including the early adherence and initiation stages. Similar rescue was observed in three independent med31ΔΔ+TEF1-ALS1 clones. Expression of TEF1-ALS1 in the wild type strain did not significantly change biofilm biomass (data not shown). As an independent confirmation that the biofilm formation defect of the med31ΔΔ mutant is not due to slower growth, we tested whether introduction of the TEF1-ALS1 construct was rescuing the growth defect of the med31ΔΔ mutant. We did not observe rescue of the mutant growth defect by TEF1-ALS1 (Figure S8), although biofilm formation was rescued (Figure 5B). This confirms that the biofilm defect is due to lower adherence and not slower growth. Collectively, these results suggest that ALS1 is a biologically relevant gene target of Med31 for biofilm formation by C. albicans.
To address more generally the roles of Mediator in C. albicans, we constructed mutants in the Med20 subunit from the Head domain (orf19.2711.1) and the Kinase domain subunit Srb9/Med13 (orf19.1452; for simplicity the mutant is indicated as srb9ΔΔ in the figures) [27]. In agreement with a general role for Mediator in adhesin gene expression, both MED20 and SRB9/MED13 were required for the expression of ALS1 and ALS3 in hyphae, and SRB9/MED13 was further required for wild type transcript levels of HWP1 (Figure 6A). Consistent with the effects on adhesins, both MED20 and SRB9/MED13 were necessary for wild type biofilm formation (Figure 6B). In regards to filamentous growth, we observed a similar trend as with the med31ΔΔ cells: filamentous growth was severely compromised in med20ΔΔ and srb9/med13ΔΔ mutants on plates (Figure 6C), while in liquid media the effects were much less pronounced (Figure 6D). In liquid media, med20ΔΔ showed a mild defect with a larger proportion of pseudohyphae (for example see the 3 and 5 h time points in serum), while srb9/med13ΔΔ behaved like the wild type (Figure 6D). To address the similarities and differences between the Mediator subunits more broadly, we analysed whether Med31-dependent genes during yeast growth were also dependent on Med20 and Srb9/Med13 for their expression. EAP1 was down-regulated in med20ΔΔ cells, ALS1 was down-regulated in both mutants, and CHT3 was only affected in srb9/med13ΔΔ cells (Figure 7A). The med20ΔΔ mutant did not display a cytokinesis defect, consistent with wild type CHT3 levels (Figure 7B). The srb9/med13ΔΔ cells were slightly elongated and a fraction (between 20–40%) formed what appeared similar to pseudohyphae, perhaps consistent with lower CHT3 levels (Figure 7B). The med20ΔΔ and srb9/med13ΔΔ mutants shared some (but not all) sensitivities to antifungal compounds and stress conditions with the med31ΔΔ strain (Table 3), and some of these sensitivities are conserved with what has been reported for the homologous mutations in model yeasts [28].
In the model yeast S. cerevisiae Mediator has been implicated in the repression of the FLO cell wall adhesins, due to the inhibitory functions of the Kinase domain subunits (including Srb9/Med13) and the Tail domain component Sin4/Med16 [25], [40]. Repression of the cell wall adhesins by the Mediator Kinase domain has also been reported in Schizo. pombe [28]. Our data in C. albicans showed that Srb9/Med13 is a positive regulator of the genes encoding adhesins (Figure 6), thus indicating that the roles of this Mediator Kinase domain subunit in adhesion are different in C. albicans. To probe this notion further, we tested the role of Med31 in adhesion in S. cerevisiae using the Σ1278b strain background, which expresses the FLO11 adhesin and is capable of adhesion-dependent phenotypes such as biofilm formation on polystyrene [67]. The wrinkled colony morphology in the med31Δ mutant was enhanced compared to the wild type, a phenotype that is indicative of higher FLO11 levels (Figure 8A). Indeed, the expression of FLO11 was significantly up-regulated in the absence of MED31 (40 fold), and consistently, the mutant was hyper-adherent in the S. cerevisiae biofilm model (Figure 8B and 8C). The levels of the other FLO family members, which are silent in the wild type, were not up-regulated under the conditions assayed (data not shown). These results are consistent with a repressive role for Med31 in the expression of FLO11 and biofilm formation in S. cerevisiae.
So far, most comparative analysis of transcription factor function in fungi have focused on DNA binding gene-specific transcription factors, and only very few studies have addressed the functions of transcriptional co-regulator complexes, such as Mediator. With Mediator, conserved roles have been reported for the core complex and the Kinase domain, including distinct roles of the sub-domains in cell wall dynamics shared by the divergent species S. cerevisiae and Schizo. pombe [9], [23], [25], [28], [29]. Our study is the first comprehensive analysis of Mediator components in the pathogen C. albicans. We identified conserved functions for Mediator subunits in Ace2-dependent gene expression and stress responsive phenotypes with the model yeasts, but also novel roles in the expression of genes related to virulence attributes of C. albicans. In particular, our study uncovered roles in the expression of the cell wall adhesins and functions in biofilm formation, which are shared between the Middle domain subunit Med31, the Head domain subunit Med20 and Srb9/Med13 from the Kinase domain.
Our data in C. albicans is consistent with reports in model yeasts in that Mediator subunits show overlapping, but also some specific roles [7], [28], [43]. In S. cerevisiae and Schizo. pombe, mutants in the Head and Middle domain subunits tend to correlate in regards to gene expression and cellular phenotypes [7], [28], [43]. We also find this for the C. albicans med31ΔΔ and med20ΔΔ mutants. For example, the expression of the cell wall adhesins ALS1, ALS3, EAP1 is down-regulated in the absence of either MED31 or MED20, both mutants display biofilm defects and a stronger filamentation defect on plates than in liquid media, and they share sensitivities to compounds such as formamide, DMSO, SDS, ethanol and congo red (Table 3). However, our data also supports differences in the roles of Med20 and Med31: Med20 is not required for the expression of the Med31-regulated genes CHT3, PIR1 and TYE7, it shows a milder biofilm formation and filamentous growth defect than Med31, and there are differences in the sensitivities to some compounds between the two mutants. Shared but also distinct roles of Med20 and Med31 are supported by results in S. cerevisiae showing a 0.41 Pearson correlation coefficient between the transcript profiles form the two mutants [43]. In Schizo. pombe, in addition to Med31, Med20 and other Head domain subunits also control the expression of Ace2-dependent genes and the mutants show defects in cell separation [28]. Deletion of MED20 in C. albicans did not result in lower expression of CHT3 or a cytokinesis defect (Figure 7). Together with Med18 and the C-terminal domain of Med8, Med20 forms a structural and functional sub-complex within the Head domain, the Med8C/18/20 submodule [28], [68], [69]. Genetic analysis in Schizo. pombe supports the idea that Med18 can compensate for the loss of Med20 [28]. It is therefore possible that Med18 would need to be inactivated in C. albicans to uncover the roles of the Mediator Head domain in Ace2-dependent transcription and cytokinesis.
In contrast to the core Mediator complex which functions predominantly in transcriptional activation, the Kinase domain is mainly a repressor of transcription and the phenotypes of the Kinase mutants in model yeasts tend not to correlate with mutants in the core Mediator (for example see [7], [28]). Our data in C. albicans supports this, as the med31ΔΔ and med20ΔΔ mutants were more similar to each other in terms of gene expression and phenotypes than to the srb9/med13ΔΔ strain. However, the srb9/med13ΔΔ strain shared some phenotypes with med31ΔΔ and med20ΔΔ including the effects on ALS gene expression, biofilm defects and sensitivities to compounds that affect the cell wall, such as congo red and SDS. The Mediator kinase domain consists of four subunits: Srb8/Med12, Srb9/Med13 and the kinase-cyclin pair Srb10/Srb11. The C. albicans mutants in Srb10/Srb11 were in the Kinase collection constructed by Blankenship et al. [35]. Our srb9/med13ΔΔ mutant shares the sensitivity to oxidative stress with cells inactivated for SRB10 or SRB11 (Table 3 and [35]), however the srb10 and srb11 mutants were reported to display wild type biofilm formation [35]. The Kinase domain subunits share many functions, but can also have different roles, in particular positive functions in the transcription of some genes have been reported that are not shared by the whole domain [70], [71]. The data reported here (Figure 6) and in [35] support distinct roles for the C. albicans Mediator Kinase domain subunits in biofilm formation.
Our data indicates that in C. albicans the transcriptional activator Ace2 could be modulating a number of Med31-dependent effects on gene expression, in particular transcriptional activation of the cytokinesis genes, the expression of the adhesins EAP1 and ALS3 and the cell wall protein PIR1, and the regulation of genes with mitochondrial functions (Figure 2 and Table S1). The ALS3 promoter contains a putative consensus Ace2 binding site within a region known to be required for activation [59], suggesting it could be a direct target of Ace2. Consistent with shared functions, both ace2ΔΔ and med31ΔΔ mutants display a cytokinesis defect, as well as adherence and biofilm formation phenotypes (this study and [72]). In Schizo. pombe Mediator interacts directly with Ace2, via the Head domain subunit Med8, and therefore Mediator plays a direct role in Ace2-dependent transcription [73]. By analogy to Schizo. pombe, and given that we did not observe a change in transcript levels for ACE2 in the med31ΔΔ mutant that would indicate an indirect effect (Dataset S1), we propose that in C. albicans Ace2 also interacts directly with Mediator for transcriptional activation. Our data also suggests that Med31 and Ace2 have roles in transcription that are independent of each other. Unlike Ace2, which is indispensable for cytokinesis in C. albicans, a milder phenotype is observed in the absence of Med31 (Figure 3B). There is only a partial overlap between genes differentially expressed in ace2 and med31 mutants respectively ([57] and this study, Table S1), and the degree of the effects on gene transcription differs between the two mutants (for example CHT3 is much more affected by inactivation of ACE2 than MED31, while MED31 has a stronger effect on PIR1, Figure 2). ALS1, which we show is a key target of Med31, does not require Ace2 during high transcription in hyphae (Figure 2). As in Schizo. pombe [28], [73], it could be that in C. albicans Mediator subunits additional to Med31 are involved in co-activating Ace2-dependent transcription. Conversely, Med31 certainly acts through additional DNA binding transcription factors, which remain to be identified.
Med31, Med20 and Srb9/Med13 all positively regulate the expression of the ALS1 and ALS3 cell wall adhesins in C. albicans, and Srb9/Med13 is further required for wild type expression of HWP1. Consistent with the transcriptional defects, the mutants are defective for biofilm formation, an adhesin-dependent phenotype [65], . Moreover, our genetic data supports the notion that the regulation of ALS1 by Med31 is biologically relevant for biofilm development. The activators that mediate the effects of Mediator on adhesin transcription remain to be characterised. The results that Med31 is required for Ace2-dependent transcription, and that Ace2 has a role in ALS3 expression suggest a potential role for Ace2 in Med31-dependent regulation of ALS3 (Figure 2). The srb9/med13ΔΔ mutant showed parallels in regards to both morphogenesis and gene expression phenotypes with the mutant inactivated in the activator Bcr1. Both mutants are required for the expression of ALS1, ALS3 and HWP1, and, as was seen for the bcr1 mutant, the srb9/med13ΔΔ mutant is proficient for hyphal growth in liquid media, but shows biofilm defects (Figure 6 and [65], [66]).
The biochemical, genetic and gene expression data in yeast supports the notion that Mediator is universally required for RNA polymerase II-dependent transcription (reviewed in [12]), and we therefore suggest that Mediator is directly involved in the transcription of the adhesin genes. Chromatin immunoprecipitation (ChIP) experiments to address occupancy of the ALS gene promoters by Med31 remained inconclusive, as we observed variability between biological replicates, from minimal to large 20–30 fold enrichments (data not shown). We suspect that this variability could be due to the exact timing of the crosslinking of Med31 to the promoters in respect to the transcriptional activation of the ALS genes in hyphae and/or how uniformly filamentous growth/ALS gene transcription is induced in the population of cells in the culture. ChIP studies in S. cerevisiae have yielded different results between labs in regards to Mediator occupancy, from modest (albeit functionally important) enrichments at some constitutively transcribed genes [75], no enrichment of Mediator subunits on the majority of transcribed genes [76], [77], to detectable enrichment of Mediator subunits upstream of many active, as well as inactive genes, and even in the coding regions of some genes [78]. More prominent enrichment for Mediator subunits is seen on genes that are responsive to stress (e.g. heat shock, or change of carbon source from glucose to galactose) [75]–[77]. Mediator does not bind directly to DNA, which is likely to be a factor in ChIP experiments. It has also been proposed that the interactions of Mediator with promoters could be transient [77]. Moreover, different Mediator subunits can yield different fold enrichments over the background (for example see [76], [77]), and it is therefore possible that a Mediator subunit other than Med31 needs to be assayed to detect consistent Mediator occupancy on promoters in C. albicans. In addition to a direct role for Mediator in adhesin gene expression, an alternative (and not mutually exclusive) possibility is that Mediator regulates the expression of transcription factors, which then in turn regulate the adhesins. The expression of several transcription factors was lower in med31ΔΔ mutants (Table 1, Dataset S2), for example that of EFG1, TEC1 and CPH1, which have been previously shown to regulate the expression of Med31-regulated adhesins ALS1, ALS3 and EAP1 [59], [79].
In S. cerevisiae Med31 is a repressor of the adhesin FLO11 (Figure 8), suggesting different functions for this Mediator subunit in the expression of cell wall adhesion molecules and regulation of adhesion-dependent phenotypes in comparison to C. albicans. In the S. cerevisiae med31Δ mutant, the reminder of the Mediator complex is intact [43], supportive of a specific role for Med31 in the repression of the FLO11 gene that does not result simply from the loss of other repressor subunits from the Mediator complex. Previous reports in S. cerevisiae for Srb9/Med13 also show repression of the FLO adhesins [25], which is again opposite to the activating function for Srb9/Med13 in the expression of ALS1/3 and HWP1 that we observed in C. albicans. Other subunits of the Mediator Kinase domain as well as Sin4/Med16 from the Tail also inhibit expression of the S. cerevisiae FLO genes [25], [40]. In Schizo. pombe, the Mediator Kinase domain subunits Srb10 and Srb8/Med12 also repress the cell wall adhesins, despite the fact that some of the affected adhesins are not related to the FLO genes [28].
While the FLO and ALS gene are not related, there is conservation in terms of the pathways and transcription factors that regulate their expression in S. cerevisiae and C. albicans respectively. Examples include positive regulation by activators Flo8, Ste12/Cph1 and Tec1, and negative regulation by repressors Nrg1, Tup1 and Sfl1 [50], [59], [80]–[83]. Mediator interacts with gene-specific transcription factors to regulate gene expression [76], [84]. It is possible that the key transcription factor(s) through which Med31 and Srb9/Med13 act to regulate the adhesin-encoding genes differ between the two yeasts. How Mediator subunits repress the S. cerevisiae FLO genes is not well understood, but a suggested mechanism involves functional interactions with the repressors Sfl1 and Tup1 [85]. Both Sfl1 and Tup1 are required for repression of the ALS genes in C. albicans [50], [81], but the requirement for specific Mediator subunits in the repression of adhesins by these factors in C. albicans could differ from what has been suggested in S. cerevisiae. Another possibility relates to silencing mechanisms. The FLO genes in S. cerevisiae are subject to position-dependent silencing regulated by histone deacetylases ([86]; reviewed in [87]). The C. albicans adhesins are not known to be regulated by silencing. Mediator has been implicated in the regulation of chromatin structure and gene expression in regions affected by silencing, such as the sub-telomeres [12], [21]–[24]. Based on work in S. cerevisiae, it has been proposed that Mediator contributes to the establishment of a repressive chromatin structure by binding to the silenced regions and influencing the recruitment of the histone deaceytlase Sir2 and the histone acetyltransferase Sas2 and thus the acetylation status of histone H4 K16, a mark of active chromatin [24]. While this has not been tested directly, an interesting possibility is that the regulation of chromatin contributes to the repressive role of the Mediator subunits in the expression of the S. cerevisiae FLO genes. For example, the Tail subunit Sin4/Med16, which is a repressor of FLO11 and of the sub-telomeric FLO1 adhesin [40], [85], contributes to sub-telomeric silencing [24]. Med7 has also been found to affect chromatin structure, by regulating H4K16 acetylation and the presence of Sir2 and Sas2 at the sub-telomeres [24]. Med7 is closely functionally linked with Med31 - the N-terminal part of Med7 together with Med31 forms a structural and functional sub-module of Mediator, Med7N/31 [43]. It will be interesting to study whether differences in silencing mechanisms that operate on the FLO and ALS genes contribute to determining how Mediator subunits control their expression.
The C. albicans strains used in this study are derivatives of BWP17 [88]. The med31ΔΔ, srb9ΔΔ and med20ΔΔ strains were constructed by standard methods based on PCR and homologous recombination, using ARG4 and URA3 as selective markers. The complemented strains were constructed by re-introducing a wild type copy of MED31, SRB9 or MED20 under own promoter and terminator into the HIS1 locus of the respective mutants using the integrative plasmid pDDB78. To make matched HIS1+ mutant strains, an empty pDDB78 vector was integrated into the genome of the respective strains. The ace2 mutant is a homozygous mutant in the BWP17 strain background and was a generous gift from Aaron Mitchell (this strain is also URA3+ ARG4+ HIS1+). The med31ΔΔ +TEF1-ALS1 overexpression strain was constructed using the plasmid pCJN498, as described in [65]. The S. cerevisiae med31Δ mutant was constructed in the ∑1278b strain using the KANMX4 cassette. The wt and flo11Δ mutant of Σ1278b were a generous gift from Todd Reynolds and are described in [67].
Standard growth conditions were YPD (2% glucose, 2% peptone, 1% yeast extract), at 30°C, 200 rpm. For ura− strains the media was supplemented with 80 µg/ml uridine. The mutants were selected using minimal media lacking the appropriate amino acids. The TEF1-ALS1 overexpression strains have a nourseothricin resistance cassette (NAT) and were selected on 400 µg/ml NAT plates (nourseothricin was from Werner Bioagents). The S. cerevisiae mutant was selected on 200 µg/ml G418 plates.
For cell morphology analysis (Figure 3A), cells were classified as being in a chain if 3 or more cells were attached. An average of 200 cells per sample were counted, and the experiments were repeated at least with 3 independent cultures. The average and the standard error are shown in the figure. To observe the mother-bud junctions, cells were stained with calcofluor white (1 mg/ml) for 8 min in the dark, followed by washes in phosphate buffered saline (PBS). Imaging was done using an Olympus IX81 microscope with the Olympus cell∧M software, using the 100× objective with DIC or the DAPI filter for calcofluor white stained cells.
Filamentous growth was tested by dilution of cells from overnight cultures grown to OD600 = 0.1–0.2 into pre-warmed YPD+10% calf serum, Spider media (1% nutrient broth, 1% D-mannitol, 2 g K2HPO4), M199 or N-acetylglucosamine media (9 g NaCl, 6.7 g yeast nitrogen base and 0.56 g N-acetylglucosamine per liter) and incubated at 37°C for the times indicates in the figures. All cell imaging was done using an Olympus IX81 microscope with the Olympus cell∧M software. For testing filamentation on plates, C. albicans strains were re-streaked on plates containing filamentous-growth inducing media. Plates were incubated for up to five days at 37°C and colonies examined and photographed with a stereo dissecting microscope (Olympus SZX 16).
For analysis of sensitivities to various drugs and chemicals, ten fold serial dilutions of cultures from wild type and mutant strains were dropped on control plates, or plates containing the compounds indicated in Figure S4 and Table 3. Plates were incubated at 30°C for three days (unless growth was assessed at 37°C or 16°C), and photographed.
C. albicans biofilms were grown in 96-well microtiter plates or on silicone disks for quantitative or qualitative analysis respectively. Quantitative biofilm assays were performed as described [89], [90]. 100 ml of cultures of C. albicans wild type or mutant strains (107 cells/ml in Spider media) were added to wells and incubated at 37°C with gentle shaking (75 rpm) for 90 min (adhesion phase). Non-adherent cells were discarded and 100 µl of fresh Spider media were added to each of the wells. Biofilms were allowed to develop for a future 4.5 h (6 h in total), 24 h, and 48 h, representing the early, intermediate or mature stage of biofilm development respectively. The medium was replenished after 24 h by aspiration and addition of fresh medium. Biofilm biomass was determined at the different time points using crystal violet staining. Wells containing only Spider medium with no yeast served as negative controls. For qualitative studies, biofilms were formed in vitro on serum-treated silicone disks, which is a well-established system for biofilm analysis [66]. Sterile silicone disks were pretreated with fetal bovine serum (Sigma) overnight at 37°C with gentle shaking (75 rpm). The silicone disks were then washed twice with PBS and transferred to a 12-well plate containing 2 ml of freshly prepared cell suspensions (107 cells/ml in Spider media). The plate was incubated for 1.5 h at 37°C with gentle shaking (75 rpm) to allow the yeast to adhere to the disk surfaces. The silicone disks were then washed in PBS and transferred to a new 12-well plate with Spider media followed by incubation for 48 h at 37°C with shaking at 75 rpm. The established biofilms were examined with SEM or CLSM. For SEM, biofilms were fixed with glutaraldehyde (2.5%, v/v, in 0.1 M cacodylate buffer, pH 7.0) and 1% osmium tetraoxide at room temperature, and dehydrated with gradually increased ethanol (50%, 75%, 95%, 100%, and absolute 100%) and hexamethyldisilazane (HMDS) (50%, 75%, 95%, 100%, and absolute 100%). Samples were coated with gold with a Balzers SCD005 sputter coater and viewed under a Hitachi S570 scanning electron microscope. For CSLM, biofilms were stained with FUN-1 (10 µM, Molecular Probes) and Concanavalin A–Alexafluor488 conjugate (Con A, 25 µg ml−1; Molecular Probes) for 45 min at 37°C. Stained biofilms were observed with a Leica SP5 CLSM, and images were captured and processed using the softwares Leica LAS AF and Amira 5.2.1. All experiments for quantitative analysis were repeated at least three times in triplicate, and qualitative assays were repeated at least 3 times. One-way ANOVA was used to compare the difference in biofilm biomass produced by different C. albicans strains. p values of <0.05 were considered to be statistically significant.
For S. cerevisiae biofilms (Figure 8), overnight cultures were growth in synthetic complete media supplemented with 2% glucose. The cells were then resuspended into 96 well polystyrene plates to an OD600 = 1.0 using synthetic 0.2% glucose media [67]. Adherence was assayed at the time points indicated in the figure using crystal violet staining as described above.
RNA was extracted using the hot-phenol method from cultures grown in YPD at 30°C to log phase (OD600 = 1). Following hot-phenol extraction, 100 µg of RNA samples were further purified using the RNAeasy kit and following the manufacturer's instructions. The microarray analysis was performed as described [91], on microarrays spotted with 6459 70-mer oligonucleotides (GEO Platform GPL9818). Data normalization and analysis was conducted in GeneSpring GX version 7.3 (Agilent Technologies). The microarray data set has been deposited in GEO, under accession number GSE31632. Genes that were up- or down-regulated in the med31ΔΔ mutant by 1.5 fold or more (p≤0.05) were selected from a Volcano Plot and considered to be differentially expressed. Gene ontology analysis was performed at the Candida genome database (CGD, candidagenome.org) [92]. GSEA analysis ([93] and Sellam et al, submitted) was performed using the GseaPreranked tool and the weighted enrichment statistics on 6387 (for C. albicans) gene sets each containing 5–500 genes. Statistical significance was estimated from 1000 permutations. Enrichment maps were constructed with Cytoscape 2.8 (http://www.cytoscape.org; [94]) and the Enrichment Map 1.1 plug-in (http://baderlab.org/Software/EnrichmentMap) using the default settings. For quantitative PCR, reverse transcription was performed using the Transcriptor High Fidelity cDNA synthesis kit from Roche. qPCR reactions were prepared using Fast-Start Sybr Green Master (Roche) on an Eppendorf Realplex master cycler and analysed by absolute quantification. The expression levels of the mRNAs were normalized to the level of ACT1 or the GAPDH encoding gene TDH3. Three independent cultures were analyzed, with two technical replicates each. qPCR primers that enable differential amplification of the ALS1 and ALS3 genes were from Green et al [95]. Sequences for all qPCR primers used in this study are listed in Table S2. For analysis of gene expression in YPD, cultures were grown under the same condition as for the microarrays analysis, to mid log phase (OD600 = 1). For assaying expression of genes under hyphal growth, strains were grown in Spider media as described in [65].
The worm-C. albicans infection assay was performed as described previously [64]. Briefly, young adult nematodes were allowed to feed for 4 h on lawns of C. albicans grown on solid BHI media (Difco) containing ampicillin (100 µg/ml), kanamycin (50 µg/ml) and streptomycin (200 µg/ml). Worms were washed with M9 media and transferred into wells of a six-well microtiter dish (Corning) containing 2 ml of liquid media (80% M9 and 20% BHI) at 60 to 80 worms per well. The plates were incubated at 25°C, and worms were qualitatively assessed at 24 h intervals for penetrative C. albicans filamentation using a DIC microscope and photographed using an Olympus IX81 microscope with the Olympus cell∧M software. The % of worms with penetrative filamentation was determined from four independent experiments at day 3 of the infection. Means and the standard error were calculated, and the p value was determined using the student t-test. The killing assays (Figure 4C) were performed three times and equivalent results were obtained.
|
10.1371/journal.pcbi.1006610 | On variational solutions for whole brain serial-section histology using a Sobolev prior in the computational anatomy random orbit model | This paper presents a variational framework for dense diffeomorphic atlas-mapping onto high-throughput histology stacks at the 20 μm meso-scale. The observed sections are modelled as Gaussian random fields conditioned on a sequence of unknown section by section rigid motions and unknown diffeomorphic transformation of a three-dimensional atlas. To regularize over the high-dimensionality of our parameter space (which is a product space of the rigid motion dimensions and the diffeomorphism dimensions), the histology stacks are modelled as arising from a first order Sobolev space smoothness prior. We show that the joint maximum a-posteriori, penalized-likelihood estimator of our high dimensional parameter space emerges as a joint optimization interleaving rigid motion estimation for histology restacking and large deformation diffeomorphic metric mapping to atlas coordinates. We show that joint optimization in this parameter space solves the classical curvature non-identifiability of the histology stacking problem. The algorithms are demonstrated on a collection of whole-brain histological image stacks from the Mouse Brain Architecture Project.
| New developments in neural tracing techniques have motivated the widespread use of histology as a modality for exploring the circuitry of the brain. Automated mapping of pre-labeled atlases onto modern large datasets of histological imagery is a critical step for elucidating the brain’s neural circuitry and shape. This task is challenging as histological sections are imaged independently and the reconstruction of the unsectioned volume is nontrivial. Typically, neuroanatomists use reference volumes of the same subject (e.g. MRI) to guide reconstruction. However, obtaining reference imagery is often non-standard, as in high-throughput animal models like mouse histology. Others have proposed using anatomical atlases as guides, but have not accounted for the intrinsic nonlinear shape difference from atlas to subject. Our method addresses these limitations by jointly optimizing reconstruction informed by an atlas simultaneously with the nonlinear change of coordinates that encapsulates anatomical variation. This accounts for intrinsic shape differences and enables rigorous, direct comparisons of atlas and subject coordinates. Using simulations, we demonstrate that our method recovers the reconstruction parameters more accurately than atlas-free models and innately produces accurate segmentations from simultaneous atlas mapping. We also demonstrate our method on the Mouse Brain Architecture dataset, successfully mapping and reconstructing over 1000 brains.
| Recent advances in brain imaging [1, 2], methods to label neurons [3], and computational methods have brought about a new era of neuroanatomical research, with a focus on comprehensively mapping brain circuits [4]. Mapping whole-brain circuitry is important for three distinct reasons: scientific understanding of how the brain works, mechanistic understanding of neurological and neuropsychiatric disorders, and as a comparison point for artificial neural networks used in machine learning [5, 6].
Circuit mapping is technique limited, and falls into three broad scales corresponding to distinct imaging modalities—indirect mapping at a macroscopic scale corresponding to MRI-based methods [7], and direct mapping at light (LM) and electron microscopic (EM) scales. For MRI and LM data, atlas mapping is an important step in the analysis. Several approaches exist for gathering LM data at the whole brain level [8–10]. For some of these approaches (two-photon serial block-face imaging, knife edge scanning microscopy and light sheet microscopy for cleared brains) two-dimensional (2D) optical sections are acquired in three-dimensional (3D) registry with each other, so that the only computational step required is 3D volumetric registration of the individual brain data set to a canonical atlas. However, for classical neurohistological approaches using tissue sectioning followed by histochemical processing, the 2D sections are gathered independently and each section can undergo an arbitrary rotation and translation compared to the block face. This may be considered a disadvantage of the classical neuroanatomical workflow, however the physical sectioning method followed by conventional histochemical analysis has certain important advantages. This allows for the full spectrum of histochemical stains, acquisition of physical sections for downstream molecular analyses, and processing for larger brains (upto and including whole human brains). Therefore it is necessary to perform an intermediate 2D to 3D registration step, where the individually acquired 2D sections are mutually co-registered into a 3D volume.
This paper develops a joint stack reconstruction and atlas mapping procedure that simultaneously restacks the 2D histology sections, applying a sequence of rigid motions to the sections, and estimates the diffeomorphic correspondence between the registered histology stack and the 3D atlas. We apply these algorithms to data sets from the Mouse Brain Architecture Project (MBAP), for which the experimental workflow generating the data utilizes a tape transfer technique [11], allowing for the sections to maintain geometrical rigidity within section and also allowing for physically disjoint components to maintain their spatial relations. The tape method ensures that the number of missing sections is minimal, with serial sections cut at a thickness of 20 μm and alternate sections subjected to Nissl staining alongside staining with histochemical or fluorescent label. These Nissl stained sections form the basis of alignment to a Nissl whole-brain reference atlas.
The histological reconstruction problem has been explored by several groups previously. Malandain first described the ill-posedness of reconstructing 3D sections and object curvature without prior knowledge of the shape of the object [12]. Rigid transformations for stack reconstruction have been estimated via block-matching of histological sections in [13], with point information based on landmarks introduced to guide volume reconstruction [14]. Dense external reference information such as MRI has been applied to guide reconstruction via registration of corresponding block-face photographs and for histology to MRI mapping [15, 16].
The principal contribution of this work is to rigorously solve the problem when an external resource of identical geometry (such as an MRI of the same mouse) is not available, while accommodating for the innate anatomical variation from atlas to subject. The lack of a same-subject reference volume is often the standard in mouse brain histology and other large scale histology studies. This places us into the computational anatomy (CA) orbit problem for which constraints are inherited from an atlas that is diffeomorphic but not geometrically identical. With the availability of dense brain atlases at many resolution scales [17–20], methods to map atlas labels onto target coordinate systems are being ubiquitously deployed across neuroscience applications. Since Christensen’s early work [21], diffeomorphic transformation has become the de-facto standard as diffeomorphisms generate one-to-one and onto correspondences between coordinate systems. Herein we focus on the diffeomorphometry orbit model [22, 23] of computational anatomy [24], where the space of dense volume imagery is modelled as a Riemannian orbit of an atlas under the diffeomorphism group. We use the large deformation diffeomorphic metric mapping (LDDMM) algorithm first derived for dense imagery by Beg [25] to retrieve the unknown high-dimensional reparameterization of the template coordinates.
Of course, for the histological stacking problem solved here, the interesting twist is the augmentation of the random orbit model with 3 rigid motion dimensions for each target section. At 20 μm, this implies as many as 500 sections augmenting the high-dimensionality of the diffeomorphism space to include as many as 1500 extra dimensions for planar rigid motions for restacking. Here lies the crux of the challenge. To accommodate the high-dimensionality of the unknown rigid motions, the space of stacked targets is modelled to have finite-squared energy Sobolev norm, which enters the problem as a prior distribution restricting the roughness of the allowed restacked volumes. The variational method jointly optimizes over the high-dimensional diffeomorphism associated to the atlas reparameterization and the high-dimensional concatenation of rigid motions associated to the target.
Fig 1 shows the components of the model for the histology stacking problem. We define the mouse brain to be sectioned as a dense three-dimensional (3D) object I ( x , y , z ) , ( x , y , z ) ∈ R 3, modelled to be a smooth deformation of a known, given template I0 so that I = I0 ∘ φ−1 for some invertible diffeomorphic transformation φ. The Allen Institute’s mouse brain atlas [26] (CCF 2017) is taken as the template. Distinct from volumetric imaging such as MRI which delivers a dense 3D metric of the brain, the histology procedure (bottom row, Fig 1) consisting of sectioning, staining, and imaging generates a jitter process which randomly translates and rotates the stack sections. Denote the rigid motions acting on the 2D sectioning planes R i : R 2 → R 2,
R i ( x , y ) = ( cos θ i x + sin θ i y + t i x , − sin θ i x + cos θ i y + t i y ) , ( x , y ) ∈ R 2 , (1)
with θi the rotation angle and ( t i x , t i y ) ∈ R 2 the translation vector in section i. The histology stack J i ( x , y ) , ( x , y ) ∈ R 2 , i = 1 , … , n, is a sequence of 2D image sections with jitter under smooth deformation of the atlas in noise:
J i ∘ R i ( x , y ) = I 0 ∘ φ − 1 ( x , y , z i ) + noise ( x , y ) , ( x , y ) ∈ R 2 . (2)
Modeling the photographic noise as Gaussian and conditioning on the sequences of jitters Ri, i = 1, …, n and atlas deformation I = I0 ∘ φ−1, φ ∈ Diff, the photographic sections Ji are a sequence of conditionally Gaussian random fields with log-likelihood
ℓ ( v , R ; J ) = ∑ i ( − α i ∫ R 2 | J i ∘ R i ( x , y ) − I 0 ∘ φ v , − 1 ( x , y , z i ) | 2 d x d y ) . (3)
Here αi is a weighting factor dependent on the noise of each section such that damaged sections can be weighted; v denotes the vector field which indexes the deformation as a diffeomorphic flow (see below).
The parameterization of the histology pipeline augments the standard random orbit model of computational anatomy with the rigid-motion dimensions of the random jitter sectioning process. The unknowns to be estimated become ( R 1 , … , R n , φ ) ∈ R 3 n × Diff for n−sections. At 20 μm then n = 500 implying the nuisance rigid motions are of high dimension O(1500). The solution space must be constrained. We use priors on the deformations and on the rigid motion stacking of the images.
Model the random sectioning with section-independent jitter as a product density π ( R ) = ∏ i π ( θ i , t i x , t i y ), the priors centered at identity, with the priors on θ circular Gaussian with standard-deviation σθ and translation with means μ c x , μ c y at the center of the sections with σ c x = σ c y:
π ( θ , t x , t y ) = 1 2 π σ θ e − θ 2 2 σ θ 2 1 2 π σ c x e − ( t x − μ c x ) 2 2 σ c 2 1 2 π σ c y e − ( t y − μ c y ) 2 2 σ c 2 . (9)
We choose our standard-deviations so that they are small relative to the center of the image, and a small rotation, roughly 5 percent of the total range of each. Generating MAP estimates of the rigid motions generates the MAP estimator of the histology restacking problem denoted as
I R ( x , y , z i ) = J i ∘ R i ( x , y ) , ( x , y ) ∈ R 2 , i = 1 , … , n .
Since the diffeomorphisms are infinite dimensional, the maximization of the log-likelihood function with respect to a function with the deformation penalty is termed the “penalized-likelihood estimator”. Conditioned on the known atlas, the augmented random variables to be estimated are ( R 1 , … , R n , φ ) ∈ ( R 3 n × Diff ).
Problem 1 (MAP, Penalized-Likelihood Estimator).
Given histology stack Ji(x,y),(x,y)∈ℝ2,i=1,… and reconstructed stack IR(⋅, zi) = Ji ∘ Ri(⋅), i = 1, …, n modelled as conditionally Gaussian random fields conditioned on jitter and smooth dormation of the template. The joint MAP, Penalized-Likelihood estimators arg maxR,v log π(R, v|J) given by
argmaxR,v−12∫01‖vt‖V2dt−12∑i‖DhIR(·,zi)‖22+∑i(logπ(Ri)−αi‖IR(·,zi)−I0∘φv,−1(·,zi)‖22). (10)
The MAP, Penalized-Likelihood estimators satisfy
{ R*=argmaxRi,i=1,…∑i(logπ(Ri)−12‖DhIR(·,zi)‖22−αi‖IR(·,zi)−I0∘φv*,−1(·,zi)‖22),v*=argmaxv−12∫01‖vt‖V2dt−∑iαi‖IR*(·,zi)−I0∘φv,−1(·,zi)‖22
with ‖ · ‖ 2 2 denoting the norm per z-axis section:
‖ f ( · , z i ) ‖ 2 2 = ∫ R 2 f ( x , y , z i ) 2 d x d y . (11)
We call this the atlas-informed model. The first two prior terms of (10) control the smoothness of template deformation and the realigned target image stack, with the third keeping the rigid motions close to the identity. The last term is the “log-likelihood” conditioned on the other variables.
The optimization for the R* rigid-motions is not decoupled across sections because of the smooth diffeomorphism of the LDDMM update and the Sobolev metric represented through the difference operator across the z− sections. Clearly, the smooth diffeomorphism is able to interpolate through the measured target sectioning data when the restacking solution gives a relatively smooth target, as diffeomorphisms are spatially smooth with at least one derivative. The optimization of the vector field v* corresponds to the LDDMM solution of Beg [25].
The principal algorithm used for solving this joint MAP-penalized likelihood problem alternates between fixing the rigid motions and solving LDDMM and fixing the diffeomorphism and solving for the rigid motions. This is described below in the following section.
When there is no atlas available this is equivalent to setting αi small and becomes a MAP rigid motion restacking of the sections:
argmax R i , i = 1 , … ∑ i ( log π ( R i ) − 1 2 ‖ D h I R ( · , z i ) ‖ 2 2 ) .
We term this the atlas-free model. The gradient of the rigid motions with respect to the components of translations tx, ty and rotation θ is defined in S3 Text. The registration is not independent across sections due to coupling through the Sobolev metric.
Here we describe the details of the algorithm used for solving for the MAP/penalized–likelihood problem described above. The algorithm alternately fixes the set of rigid motions while updating LDDMM and fixes the diffeomorphism while updating the rigid motions.
Algorithm 1.
0. Initialize φnew, Rnew ← φinit, Rinit, Iold ← J ∘ Rinit:
1. Update φold←φnew,Riold←Rinew, Iold(⋅, zi) ← Inew(⋅, zi), i = 1,….
2. Update LDDMM for diffeomorphic transformation of atlas coordinates:
v n e w = argmax v − 1 2 ∫ 0 1 ‖ v t ‖ V 2 d t − ∑ i α i ‖ I R − o l d ( · , z i ) − I 0 ∘ φ 1 v − 1 ( · , z i ) ‖ 2 , φ n e w = ∫ 0 1 v t n e w ∘ φ t n e w d t + id . (12)
3. Deform atlas I0 ∘ φnew−1 and generate new histology image stack:
R n e w = arg max R i , i = 1 , … ∑ i ( log π ( R i ) − 1 2 ‖ D h I R ( · , z i ) ‖ 2 2 − α i ‖ I R ( · , z i ) − I 0 ∘ φ n e w − 1 ( · , z i ) ‖ 2 2 ) ; I R − n e w ( · , z i ) = J i ∘ R i n e w ( · ) , i = 1 … (13)
4. Return to Step 1 until convergence criterion met.
The form of the gradients for the rigid motions is given in S4 Text. The LDDMM update solutions are given by Beg [25].
The algorithm described above is applied to Nissl histological stacks using the Allen Institute’s mouse brain atlas as a template. The Allen Mouse Brain Atlas is a micron-scale atlas that includes annotated Nissl-stained images at 10, 25, 50, and 100 μm voxel resolution, with 738 labeled compartments in the annotation.
Atlas mapping is computed on the Nissl-stained histological image stack showing the clear definition of anatomical boundaries. The associated fluorescent tracer images are transformed to the Nissl stack so that the atlas subvolume labels can be cast onto the new modality. The fluorescent and Nissl images are registered within animals by applying rigid registration based on a mutual information cost function.
A software pipeline which performs start-to-finish registration operations was implemented on a high performance computing cluster for atlas-mapping and histology restacking on the Mouse Brain Architecture data. To date, the pipeline has been successfully run on over 1000 MBAP brains. The general pipeline workflow is illustrated in Fig 2. In our application, we apply a two channel LDDMM [32] algorithm for the optimization with respect to φ, where the first channel is the Nissl-stained grayscale image, and the second channel is a mask of the brain tissue with ventricles and background set to a pixel value of zero. The brain mask for each brain stack is automatically generated by thresholding at an estimated background intensity value and applying morphological opening and closing for denoising. The threshold value is estimated by a RANSAC-like procedure over the image histogram, assuming a normal distribution of intensity values in the image foreground. A first-order Sobolev-norm (see below) is used for the smoothness constraint regularization of the histology stack. In order to accommodate for sections damaged by the histology process or structures excluded from imaging, the objective functions in all parts of the algorithm are optimized with respect to only the image data that exists. Essentially, this is a masking procedure on the cost function that allows matching between a whole atlas brain and some target which is a partial, or subset of a whole brain.
After registration of the structural Nissl image, the fluorescence volume is registered to its corresponding Nissl volume. The registration is restricted to rigid motions on each individual section. The optimization bears a similar form to Eq (13) with the squared error matching term replaced with mutual information in order to account for the different modalities of the template and target histology stack. Once fluoro-to-Nissl registration is complete, the Nissl segmentation can be applied to the fluorescence image.
The LDDMM algorithm that maps the atlas image to an aligned stack of sections is implemented in C++. Images and other data are stored as basic arrays, and there are no dependencies other than for FFTs (we use FFTW or Intel MKL depending on availability). The remainder of code is written in Matlab (Natick, MA).
The run-time/complexity for the volume LDDMM algorithm has complexity order nT Nvoxlog(Nvox), where nT is the number of steps for integrating the time varying velocity field, and Nvox is the total number of voxels. The slice based portion of the code is order Nvox. While the FFTs are order NlogN, in practice most computation time is spent during linear interpolation (order N). The end-to-end running time from initial stack alignment to completed atlas registration is approximately 6-8 hours using 8 cores on an Intel Xeon E5-2665 processor for target and template image volumes of approximately 200 × 300 × 300 voxels. Jobs are performed in parallel on a high performance cluster at CSHL. The fluoro-to-nissl cross registration running time is approximately 1 hour on the same environment and volume size.
The following hyper-parameters are required by our model, with sample values provided for the MBAP dataset:
the weights between the matching term (1.0), the regularizing prior (0.001), and the Sobolev norm (1.0) on the rigid objective function
the variances of the priors on rotation (π 9) and translation (7.0) in each stacking plane
the weight between the matching term (0.4) and the regularizing term in LDDMM (1.0)
the LDDMM kernel size (a cascade of 0.05, 0.02, and 0.01)
the initial gradient descent step size (0.000025 for rigid parameters and 5e-13 for LDDMM parameters)
The hyper-parameters were selected by grid search on a predefined range of parameter values, testing the rigid stack alignment and LDDMM parameters separately.
A final experiment was conducted on brain data sampled from the MBAP database, using the Allen mouse brain as the atlas. We selected specific targets which were prone to poor registration results due to image intensity local minima. In particular, structures like the cerebellum tend to be difficult to register accurately due to their folded nature; one fold can easily be mistaken for the adjacent fold, and if the target and atlas are not well initialized, the deformation required to flow one fold onto another can have a high metric cost. We are also interested in inspecting lower-contrast structures like the corpus callossum, which may be poorly registered due to local minima in other nearby bright structures. We also evaluate our mapping quality in the hippocampal region, which is one of the most relevant regions for the study of neurodegenerative diseases.
The reconstructed histological target stack in the atlas-informed model shown in Fig 8a takes on the shape of the atlas but is prone to reconstruction artifacts. The deformation grids produced by the atlas-informed mapping is much smoother and has many fewer wrinkles than the atlas-free mapping. This is seen clearly in Fig 9.
Fig 10 shows examples of improved segmentations in selected regions of the brain. The atlas-informed model generates more accurate segmentation results and produces smoother mappings as exhibited by the less wrinkled and distorted grids (bottom row b), showing more consistent results throughout the MBAP dataset.
This paper examines the CA random orbit model at the mesoscale for the stacking of sectioned whole brains coupled with mapping to annotated atlases. The standard CA model has been expanded to include the O(3 × n) extra rigid motion dimensions representing the planar histology sections. The estimation procedure solved here simultaneously estimates the diffeomorphic change of coordinates between atlas and target histological stack, as well as the “nuisance” rigid motion parameters for each section in stack space. This requires the introduction of a smoothness constraint on the target jitter simultaneous with LDDMM, which is enforced via a Sobolev metric, encouraging the reconstructed stack to be smooth by controlling the derivative along the cutting axis.
Results are shown demonstrating that the introduction of an atlas into the estimation scheme and simultaneously accommodating for the nonlinear atlas-to-target shape difference via diffeomorphism solves several of the classic problems associated with volume reconstruction, including the recovery of the curvature of extended structures. Since the atlas gives a priori indication of the global shape, the tendency to remove distortions along the section axis is balanced against the desire to minimize the amount of deformation of the atlas onto the reconstruction. The algorithm is shown to mediate this tension well.
The clear limitation of this method is that we model sections that are out of order, folded upon themselves, or damaged by censoring from the mapping solution using the weighting coefficient αi and removing their impact from the overall deformation. This is a global censoring, but we do not apply shearing deformations within plane and we do not include in the algorithm an automatic solution to detecting the folding problem. Although we do not currently include correction beyond rigid motion within the plane of each section, one could imagine attempting to add such shearing distortions to the model, which would remain stable as the number of new dimensions would remain low. The global censoring solution requires human quality control within the pipeline for detection of globally deformed or damaged sections.
The use of dense large deformation diffeomorphic image matching is being used extensively for magnetic resonance imaging in the brain at 1 millimeter scale for both T1 and DTI [23, 25, 32, 33] as well as for human anatomy [22] including for transferring the geometries of Cardiac fibers in dense Cardiac imaging [34, 35] and for radiation treatment planning [36]. These technologies form the basis of many implementations such as Ashburner’s important SPM [37, 38]. The aforementioned applications have not included complex prior distributions to encode distortions such as the Sobolev derivative prior introduced here that may have be required due to the distortions introduced in the imaging and stacking process.
|
10.1371/journal.ppat.1002221 | APOBEC3A Is a Specific Inhibitor of the Early Phases of HIV-1 Infection in Myeloid Cells | Myeloid cells play numerous roles in HIV-1 pathogenesis serving as a vehicle for viral spread and as a viral reservoir. Yet, cells of this lineage generally resist HIV-1 infection when compared to cells of other lineages, a phenomenon particularly acute during the early phases of infection. Here, we explore the role of APOBEC3A on these steps. APOBEC3A is a member of the APOBEC3 family that is highly expressed in myeloid cells, but so far lacks a known antiviral effect against retroviruses. Using ectopic expression of APOBEC3A in established cell lines and specific silencing in primary macrophages and dendritic cells, we demonstrate that the pool of APOBEC3A in target cells inhibits the early phases of HIV-1 infection and the spread of replication-competent R5-tropic HIV-1, specifically in cells of myeloid origins. In these cells, APOBEC3A affects the amount of vDNA synthesized over the course of infection. The susceptibility to the antiviral effect of APOBEC3A is conserved among primate lentiviruses, although the viral protein Vpx coded by members of the SIVSM/HIV-2 lineage provides partial protection from APOBEC3A during infection. Our results indicate that APOBEC3A is a previously unrecognized antiviral factor that targets primate lentiviruses specifically in myeloid cells and that acts during the early phases of infection directly in target cells. The findings presented here open up new venues on the role of APOBEC3A during HIV infection and pathogenesis, on the role of the cellular context in the regulation of the antiviral activities of members of the APOBEC3 family and more generally on the natural functions of APOBEC3A.
| Macrophages and dendritic cells represent important targets for HIV-1 and the understanding of the complex relationship established between these cells and the virus is of the outmost importance. Here, we show that APOBEC3A, the least known member of the APOBEC3 family of cytidine deaminases, restricts HIV-1 specifically in these cells. The antiviral effect of APOBEC3A is exerted at viral DNA accumulation during de novo infection through a mechanism that may involve deamination. For the first time, our results indicate that APOBEC3A, rather than being devoid of inhibitory activity against HIV-1, plays an important role in blocking not only HIV-1, but more generally primate lentiviruses. This antiviral effect is specific to myeloid cells in which this factor is naturally expressed. Among the viral proteins capable of opposing APOBEC3A, we have found that Vpx, a protein coded by members of the SIVSM/HIV-2 lineage provides a partial protection against this factor by inducing its degradation. Overall, these results shed new light on the regulation of members of the APOBEC3 family in primary cells and open up new venues on the role that APOBEC3A, a restriction factor specifically expressed in myeloid cells, may play during viral pathogenesis.
| The apolipoprotein B mRNA editing enzyme, catalytic polypeptide-like 3 family (APOBEC3s) comprises 6 members of highly related cytidine deaminases [1]. The prototype of the family, APOBEC3G (or A3G), has been identified on the basis of its ability to inhibit HIV-1 infection in the absence of the Vif protein [2]. In this case, the antiviral effect of A3G is exerted via its incorporation into virion particles in virus-producing cells. This incorporation leads to the deamination of newly synthesized viral DNA during the subsequent cycle of infection [3]–[6], although a non-deaminase dependent mechanism of inhibition has also been described [7], [8]. In the presence of Vif, A3G is targeted to an E3-ubiquiting ligase complex and is thus degraded in virus-producing cells [9]–[17]. Even if this seems the major mechanism with which the virus protects itself, retroviruses can use a Vif-independent manner that does not involve the degradation of A3G, but results in the exclusion of the protein from assembling viral particles [18]–[21]. In the case of HIV-1, Vif has also been proposed to promote structural changes in A3G that negatively affect its ability to be incorporated into virion particles [22], [23].
While this host-pathogen struggle takes place in producing cells, the role of the pool of A3G molecules present in target cells and thus welcoming incoming viral particles remains unclear and in large part unexplored [24], [25]. The possibility that APOBEC3 members might exert an inhibitory effect directly on incoming viruses is interesting, because in these steps Vif is absent (or present only in trace amounts in virion particles), so that the virus would be denuded of protection against them [26], [27]. A3G has been reported to block the early phases of HIV-1 infection of quiescent T cells and DCs, but this activity has been recently ascribed to more general modifications of the activation status of quiescent lymphocytes, rather than to a specific effect of A3G silencing during the early phases of infection [24], [25].
All the members of the APOBEC3 family coded by primates (A through H) exhibit antiviral properties, although their potency may vary (for review see [28], [29] and [30] for the specific case of A3H). The exception to this rule is APOBEC3A (A3A) that is to date the only member of the APOBEC3 family whose ectopic expression in established cell lines (HeLa or HEK293T cells) bears no consequence for retroviruses [21], [31]. APOBEC3A is a single-domain cytidine deaminase, it is capable of editing exogenous and endogenous DNA upon expression in HeLa cells and has been shown to be responsible for the strong editing activity observed in primary monocytes [32]–[35]. A3A is capable of inhibiting the replication of LTR and non-LTR retroelements [36]–[39], of Parvoviruses [36] and has more recently been shown to moderately inhibit Alpharetroviruses (more specifically the Rous Sarcoma Virus, RSV, upon incorporation into virion particles, [40]). Increased expression of A3A, along with other APOBEC3 members, has been observed in keratinocytes and skin of precancerous cervical biopsies of human papillomavirus (HPV)-positive patients, presence that has been correlated to an increased evidence of HPV editing [41]. Although certain studies suggest that the editing activity of A3A may play a role in the host cell genome integrity and advance the hypothesis that this activity may lead or contribute to cancerogenesis [34], the inhibition of retroelements and of the adeno-associated virus 2 (AAV2) seems independent from the deaminase activity of A3A [36]–[39].
In the case of HIV-1, no firm evidence links A3A to an antiviral activity in either producing or receiving cells, although one study advanced the hypothesis that the increased replication of HIV-1 in differentiated macrophages over monocytes could be due to their lower content in A3A [42]. Together with a few recent articles indicating that A3A could be preferentially expressed in cells of the myeloid lineage, these reports call for a deeper analysis of the relationship existing between HIV-1 and myeloid cells [43], [44].
Primary circulating blood monocytes, along with differentiated macrophages and dendritic cells (DCs) are key cells in immune responses acting both as sentinels of danger signals, as well as instructors for other cells of the immune system (reviewed recently in [45]–[47]. These cells represent an important target for HIV-1 replication and have been shown to serve both as a vehicle for its dissemination through the body, as well as a viral reservoir. However, myeloid cells are collectively more resistant to HIV-1 infection than other cell types, as for example activated lymphocytes. This resistance varies according to the differentiation status and stimulation present and appears at different steps of the viral life cycle. Among these steps, a major impairment occurs during the early phases of infection, as a number of laboratories including ours have established [48]–[55]. The exact nature of this resistance is unclear and likely to be multifactorial. However, a number of recent evidences point to the existence of cellular restriction factors that affect incoming virus specifically in the peculiar environment of myeloid cells [50], [56]. In this respect, two very recent studies identified one such factor in SAMHD1, whose alternative name is dendritic cells derived interferon γ-induced protein (DCIP) [57] and the deficiency of which causes the Aicardi-Goutières syndrome, a disease in which interferon responses are deregulated [58]. Despite the fact that SAMHD1 is expressed in different cell types, it seems to restrict lentiviral infection specifically in myeloid cells, probably highlighting the need for a particular cellular context, or for specific cellular partners [59], [60].
Given that previous reports indicated that A3A was preferentially expressed in CD14-positive monocytes [42]–[44], we tried to determine whether the pool of A3A molecules present in target cells could inhibit de novo infection of myeloid cells by HIV-1.
To this end, we first determined that A3A is the sole member of the APOBEC3 family specifically expressed in different myeloid cells. Then, using either ectopic expression of A3A in established HeLa cells or specific silencing in primary macrophages and DCs, we determined that A3A does indeed inhibit the early phases of infection specifically in myeloid cells. Indeed, silencing of A3A not only increases the susceptibility of target cells to HIV-1 during single round infectivity assays, but also augments the spread of R5 tropic HIV-1 viruses. This antiviral effect is mediated by the pool of A3A present in target cells and is exerted through a decrease in viral DNA accumulation. The identification of TC editing in vDNA produced during HIV-1 infection of myeloid cells suggests that this antiviral mechanism may at least in part be exerted via deamination of the viral genome. The inhibition mediated by A3A targets also SIVMAC viruses, suggesting a conserved antiviral mechanism directed against primate lentiviruses. Finally, we provide evidence that Vpx, a protein coded by members of the SIVSM/HIV-2 lineage seems to provide partial protection against A3A by driving its degradation.
A few reports have indicated that among circulating white blood cells, A3A was highly expressed in monocytes [42]–[44]. To determine more widely the pattern of expression of A3A and of the different members of the APOBEC3 family, quantitative RT-PCR was performed on quiescent or PHA-stimulated primary blood cells (PBLs, depleted of monocytes), monocytes, macrophages and dendritic cells (DCs) differentiated upon incubation of monocytes for 4 to 5 days with M-CSF and GM-CSF/IL4, respectively (Figure 1). Our analysis reveals that A3A is overexpressed by at least 100 fold in differentiated macrophages over PBLs, irrespectively of their activation status. This difference increases further in DCs and is the greatest in monocytes in which A3A is expressed over 5 logs more than in stimulated PBLs. The expression of the remaining APOBEC3 members did not display such cell type specific variations, with the exception of A3D/E that seems expressed at least 10 fold more in PBLs than in the myeloid cells tested. On the contrary, A3B and to a lower extent A3H are less expressed in non-stimulated monocytes than in stimulated PBLs, although their expression increased during differentiation into either macrophages or DCs. Overall, myeloid cells express consistent levels of all APOBEC3 members, but A3A is the only one whose expression is restricted to them. These results are in line with previous reports and extend them in the absolute quantification of the copy number of the different APOBEC3 members during the differentiation of circulating monocytes into macrophages and DCs. Of note, IFNα treatment increased the expression levels of A3A at both the mRNA and protein level, but the gradient of A3A expression did not change with respect to untreated cells: monocytes, DCs, macrophages and PBLs (from the highest to the lowest, Figure S1A and S1B). At the protein level, we confirm that A3A can be recognized after migration on an SDS-PAGE gel by the ApoC17 antibody initially raised against A3G [23] and Figure S2). In our hands, A3A can be detected as a doublet by WB. Detection of this doublet is however highly variable and unpredictable, probably influenced by the detection limit of A3A in a given experiment and by donor-to-donor variations. In any case, the antiviral effect of A3A seems not linked to the detection of an A3A doublet by WB.
In conclusion, these results indicate that if myeloid cells express comparable levels of APOBEC3 members, A3A is the sole member of the APOBEC3 family whose expression is restricted to cells of myeloid origins, among cells of hematopoietic origins.
To determine whether the levels of A3A were modulated during the course of HIV-1 infection, M-CSF-differentiated macrophages were infected with an R5 tropic HIV-1 virus (Yu2) and viral spread was assessed by determining the amount of virion-associated RT activity released in the cell supernatant (by exo-RT, Figure 2A). At day 7, when consistent viral spread had occurred, cell aliquots were lysed and the levels of APOBEC3 members analyzed at the mRNA level and, for A3A, also at the protein level by WB (Figures 2B and 2C, respectively). Spreading HIV-1 infection was associated to a consistent increase in A3A at both mRNA and protein levels. With the exception of A3B whose expression decreased during HIV-1 infection, the expression of the remaining members of the APOBEC3 family was not significantly modified. This increase was not due to the presence of type I interferon (IFNs) in the cultures, as assessed by the failure of supernatant derived from infected macrophages to inhibit the spread of a GFP-bearing Vesicular Stomatitis Virus in A549 cells (VSV, Figure S3). This assay is not only extremely sensitive, but allows the detection of the presence of all the subtypes of IFN α and β in the culture (in number of 13 and 2, respectively) [61], [62]. Although we cannot exclude the secretion of low levels of type I IFN and subsequent engagement of the IFN receptor, absence of IFN production during HIV-1 infection of dendritic cells has been recently reported [63]. Thus, although the expression of A3A can be induced by type I IFN [35], [43], [64], [65] and Figure S1), its expression during HIV-1 spreading infection of macrophages can also increase in a manner that is largely type I IFN-independent. To further support this argument, the expression of A3G, also modulated by type I IFN, is not significantly modified during infection. Thus, the upregulation of A3A observed here is likely to be part of a more complex response to viral infection that takes place in infected macrophages.
The main described anti-retroviral function of APOBEC3 members involves the pool of APOBEC3s molecules present in virus-producing cells. On the contrary, the role of the pool of APOBEC3 molecules present in target cells has been explored only for A3G and remains highly debated [24], [25]. Among the members of this family, A3A is to date the sole member devoid of known inhibitory activity against HIV-1 when overexpressed in established cell lines [21], [31]. In agreement with these previous results, single cycle infection of HeLa cells overexpressing an HA-tagged version of A3A (that is editing-competent, data not shown) did not significantly modulate the susceptibility of target cells to infection with both complete or minimal HIV-1 vectors (i.e. containing or lacking non-structural viral proteins, Figure 3A). Given that no differences were observed between complete and minimal vectors in the experiments described below, the formers were routinely used for their slight but consistent higher infectivity (data not shown and [66]).
Contrarily to what observed in HeLa cells, silencing of A3A increased the susceptibility of target myeloid cells to HIV-1 infection (Figure 3B). Silencing was established using either miR-shRNAs expressed upon HIV-1-vectors mediated gene transduction [67] or liposome-mediated transfection of siRNAs. Silenced cells were then challenged with GFP-coding HIV-1 vectors. Given that both silencing methods yielded similar results, the formers were routinely used, as they allowed for more potent silencing. Knockdown of A3A increased the susceptibility of all cells of myeloid origins tested to the challenge with HIV-1 by an average of 5 to 7 fold: THP-1 differentiated into macrophage-like cells upon incubation with PMA (Diff. THP-1), as well as macrophages and DCs (Figure 3B). As expected, knockdown of A3A exerted no effect on the infection of PHA-stimulated PBLs (in the presence or absence of IFNα, Figure 3B and Figure S4). This result was expected as A3A is barely expressed in both quiescent and activated PBLs even after IFNα treatment, if compared to myeloid cells. The absence of an effect of the A3A knockdown on the susceptibility of PBLs to HIV-1 infection suggests the lack of non-specific effects (i.e. the A3A target must be present for the observed phenotype). To further control the specificity of our knockdowns, quantitative RT-PCR analysis was carried out on mRNA preparations derived from miR-shRNA-silenced macrophages for all the members of the APOBEC3 family (Figure S5A). This analysis indicated a specific decrease in A3A mRNA, but not in the mRNAs of other members of the family. As further proof of the specificity of the A3A knockdowns, A3A was silenced in macrophages upon liposome-mediated transfection of siRNAs. These siRNAs where chosen to target sequences within the A3A mRNA that were distinct from those targeted by the miR-shRNAs. Again, knockdown of A3A yielded an increase in the infectivity of HIV-1 similar to the one observed for miR-shRNAs (Figure S5B).
Overall, these experiments reveal that A3A plays a specific antiviral role during the early phases of HIV-1 infection in cells of myeloid origins in which it is naturally highly expressed.
To determine whether the antiviral effect of A3A could be observed also during replicative infection, silenced macrophages were challenged with R5 tropic HIV-1 (Yu2) at MOI of 0,05 and viral spread monitored by exo-RT activity (Figure 4). Even in this case, the specific silencing of A3A (shown in the WB side panel) increased viral spread in the culture. Although this result does not exclude an effect of A3A in virus-producing cells (the most classically described antiviral effect of APOBEC3s), this increase does correlate with the one observed during the early phases of infection in single round infectivity assays, indicating that A3A inhibits viral spread in macrophages. In light of the data presented in Figure 2B indicating an increase of A3A during spreading infection of HIV-1, we can hypothesize that the effect of silencing of A3A is mitigated by the overexpression of A3A observed during spreading HIV-1 infection. Similar experiments were attempted in DCs, but were abandoned due to the high mortality of silenced DCs upon viral challenge. This higher mortality is independent from the identity of the silenced gene, as it is observed in control and A3A knockdowns alike and does not seem specific for HIV-1.
To identify the step affected by the depletion of A3A in target cells, silenced macrophages were lysed 24 hours post infection and the accumulation of vDNA analyzed by qPCR. The use of an HIV-1 based miR-shRNA vector precluded the analysis of all the viral DNA products synthesized over the course of infection. However, the presence of GFP in the challenge HIV-1 vector enabled the analysis of vDNA products obtained after the minus strand strong stop, which represent one of the major restrictive step during the infection of myeloid cells. Under these conditions, silencing of A3A increased the accumulation of vDNA by 4 to 5 fold (Figure 5A), in agreement with the increase observed in infectivity. While the analysis of vDNA products by PCR yields quantitative data on their accumulation as a whole, this technique does not reveal the kinetic behavior of truly infectious vDNA. Infectious vDNA constitutes a minor fraction of the total vDNA produced during infection [68], a phenomenon that is exacerbated in primary cells more resistant to infection. Thus, determining the behavior of these genomes represents an important additional parameter during the early phases of infection. To determine the kinetic behavior of infectious vDNA and more specifically to determine how fast it was completed during infection, we employed a technique we had previously described [48]. In this setup, infectious viral genomes are defined as those capable of expressing the viral-coded GFP reporter and their accumulation over time is determined by arresting the reverse transcription process at different times post infection through the addition of the RT inhibitor Nevirapine. The percentage of GFP-positive cells at each time point is then determined by flow cytometry 3 days post infection and values are graphed after normalization to control infections carried out in the absence of RT inhibitors. This assay yields a kinetic measurement of the speed at which completion of infectious vDNA is carried out. In this respect, the assay does not yield information on the total amount of vDNA synthesized during infection as in Fig. 5A, but describes another parameter which is the speed at which infectious vDNA is produced. When control and A3A-silenced macrophages were thus analyzed, no significant changes were observed in the kinetics of reverse transcription of HIV-1 infectious genomes (Figure 5B). These results indicate that A3A does not modify the kinetics of reverse transcription, but affects the overall amount of vDNA synthesized. This decrease is of the same amplitude of the effect observed in viral infectivity.
Inhibition of AAV or retroelements by A3A seems to follow a deaminase-independent mechanism. To determine whether this was the case here, vDNA products accumulated during infection of macrophages were amplified, cloned and sequenced in search of potential cytidine editing. HIV-1 vDNA produced in macrophages exhibited a significantly higher proportion of edited cytidine with respect to vDNA synthesized in HeLa cells (Figure 5C). This deamination signature depended at least in part from A3A, since the number of edited sites decreased upon silencing of A3A. When the dinucleotide motifs surrounding the edited cytidine were examined, a minority was found to be CC (described as specific of A3G), while more than a half were TC (described to be preferentially recognized by the remaining A3 members) (Figure 5D). Cytidine deamination of the latter TC motif depended from A3A as their proportion specifically decreased upon A3A silencing. Overall, these results indicate that a small yet detectable cytidine editing signature is present in vDNA produced during the infection of myeloid cells, a portion of which is clearly influenced by the intracellular levels of A3A.
A previous report indicated that SIVMAC Vpx could degrade A3A [72]. To determine whether this was the case, A3A and A3G were co-transfected along with HIV-1 Vif and SIVMAC Vpx in HeLa cells (Figure 7A). As expected, HIV-1 Vif decreased the steady state levels of A3G and A3A, although this effect was less pronounced for A3A. On the contrary, SIVMAC Vpx affected only A3A, but not A3G. When the stability of A3A was determined, a clear decrease in its stability was observed upon co-expression of SIVMAC Vpx (Figure 7B). This degradation seems to involve a proteasome dependent mechanism since degradation of A3A by Vpx is significantly blocked upon incubation with the proteasome inhibitor MG132 (Figure 7C).
To determine whether Vpx was also able to degrade endogenous A3A in myeloid cells, SIVRCM and SIVMAC Vpx proteins were expressed in DCs after lentiviral-mediated gene transduction. Briefly, HIV-1 vectors coding the 2 Vpx proteins were produced in 293T cells, normalized by protein content and used to transduce DCs. Under these conditions, Vpx was not incorporated into HIV-1 particles and thus did not affect the infectivity of these vectors in DCs (data not shown). When DCs were analyzed 4 days after transduction, a clear decrease in the levels of A3A was observed in DCs expressing SIVMAC Vpx, but not in DCs expressing SIVRCM Vpx. This decrease was specific for A3A, as the levels of A3G were unaffected by the expression of Vpx proteins.
Overall, these results indicate that SIVMAC Vpx may partially protect SIVMAC viruses from the negative effects of A3A by inducing its degradation during the early phases of infection.
In the present study we reveal a novel role for A3A in controlling HIV-1 replication specifically in myeloid cells. Cells belonging to this lineage express all members of the APOBEC3 family, but A3A is the only one restricted to myeloid cells among the white blood cells targeted by HIV-1 in vivo and is also the sole member whose expression augments during spreading infection of primary macrophages. The results presented here indicate that the specific silencing of A3A in differentiated THP-1 cells (that mimic macrophages upon differentiation), primary macrophages and DCs increases their susceptibility to de novo HIV-1 infection, while on the contrary silencing in primary PBLs or ectopic expression of A3A in established cell lines exerts no effect on these phases. Thus, these results indicate that the antiviral effect of A3A is common to cells of myeloid origins. We have been so far unable to silence A3A in monocytes in a robust manner that also preserves their non-stimulated phenotype, so that the effect of A3A on HIV-1 could not be explored. However, in light of its extremely high expression, we may surmise that A3A hampers HIV-1 in non-stimulated monocytes, as suggested previously [42].
Overall, our findings are in agreement with past studies indicating that A3A does not exert an antiviral activity in HeLa or HEK293T cells [21], [31], and extend them by revealing that this property is explicit only in cells in which A3A is naturally expressed, i.e. myeloid cells.
While the restricted expression of A3A indicates a tight transcriptional regulation, the lack of antiviral activity upon ectopic expression in HeLa cells suggests that the mere expression of A3A is not sufficient to confer an antiviral state to recipient cells. This observation suggests that A3A may be regulated also at the post-translational level through modifications that affect its activity, localization, association to co-factors, processivity or all these things together. This cell type dependent regulation is not unprecedented among members of the APOBEC3 family, as an as yet unidentified inhibitory activity has been recently proposed to regulate the editing activity of A3G in T cells [79]. In the case of A3A, these different levels of control may be extremely important to regulate its activity and cytotoxicity. In our hands, the stable expression of A3A is particularly cytotoxic (data not shown), an effect that is likely linked to its ability to edit the cellular genome and to induce a DNA damage response that is followed by cell cycle arrest in A3A-overexpressing cell lines [32], [34], [35]. However, myeloid cells express constitutively high levels of A3A and must have devised manners to protect themselves from these negative effects. Since all the myeloid cells tested here are non-dividing, it is possible that this state makes them less susceptible to editing of the cellular genome or to a DNA damage response than cycling cells. Alternatively, and not mutually exclusive with the hypotheses mentioned above, a specific post-translational regulation might protect myeloid cells. Such modifications, and/or the association with other cellular components may strongly influence the ability of A3A to recognize its target for example by shaping the single-stranded DNA docking groove of A3A, as identified in [80].
The results presented in this study indicate that the pool of A3A present in target cells is directly capable of inhibiting incoming viruses. This possibility had been raised before in quiescent T cells and DCs for A3G, although a recent report tempered enthusiasm over the function of APOBEC3 members during these phases [25]. Our results demonstrate that the specific knockdown of A3A positively impacts the early phases of HIV-1 infection, as clearly observed using vectors capable of a single round of infection. An antiviral effect at this step is conceptually interesting because Vif, that normally inhibits APOBEC3s during the phases of virus production, is absent in virion particles [26], [27]. As a consequence, incoming virus may be largely deprived of one of its major protection against APOBEC3 members present in recipient cells. Indeed, as a further proof that Vif cannot influence the early phases of infection, use of minimal HIV-1 vectors deficient in Vif, did not modify the results obtained here (data not shown). At present, we ignore whether this inhibition during the early phases of infection is specific to the combination A3A-myeloid cells, or if it is more generally present in other primary cells. Despite the fact that a number of studies have characterized the lack of effect of APOBEC3 members during the early phases of infection, rare are those that examined this question in the natural targets of HIV-1 replication. In light of the results presented here, we believe this issue deserves further investigation.
A3A inhibits HIV-1 during reverse transcription, the step that a number of laboratories including ours have determined as a major restriction during the infection of primary myeloid cells [48]–[55]. The process of reverse transcription is extremely sensitive to the intracellular environment and multiple factors have been shown to affect it (core-stability, viral protein processing, dNTPs levels, possibly intracellular trafficking and so forth). So, in principle, A3A could inhibit this process by affecting any of these parameters either more potently than other APOBEC3 members or else more specifically in myeloid cells in which these steps take place at a slower pace with respect to other cell types. Viral DNA produced upon infection of primary myeloid cells presents cytidine editing in higher levels than the one synthesized in HeLa cells. More than 50% of edited cytidines are present within a TC dinucleotide recognized by all members of the APOBEC3 family with the exception of A3G, while most of the remaining edited sites are within a GC dinucleotide that does not seem to be specific for APOBEC3s. Only a minority of edited cytidine is present in a CC dinucleotide context specific to A3G (10 fold less than TC). Among the different sites, the levels of TC editing are specifically influenced by the intracellular levels of A3A, indicating it as a major responsible for TC editing in myeloid cells. The amount of cytidine editing observed here is small and unlikely to account for the antiviral effect of A3A observed here. However, this low amount of editing may represent only the detectable fraction of the total edited vDNA, as edited vDNA may be efficiently removed or repaired in myeloid cells. Given that detection of this small amount of vDNA editing is challenging, previous reports indicating its absence upon A3A inhibition of LINE-1 retrotransposons and AAV2 may mean that either editing is in these cases lower than what observed here, or else that A3A can exert its antiviral effect independently from deamination [36]–[39].
Although our results reveal that A3A plays a role directly in target cells, the finding that A3A is a cell type specific restriction factor suggests that A3A may also exert the most classical of the antiviral activities of APOBEC3s, namely the one of being incorporated into HIV-1 virions and then deaminate newly synthesized vDNA. In light of the exquisite cell type dependent regulation of the activity of A3A this possibility is highly likely and in this case HIV-1 Vif that does degrade A3A, although less efficiently than A3G, may protect the virus against A3A as it does against other APOBEC3s.
The features of HIV-1 inhibition by A3A seem essentially conserved among primate lentiviruses, as the depletion of A3A increases also the infectivity of SIVMAC. However, we have noticed that Vpx-deficient SIVMAC viruses are more responsive to A3A depletion than WT ones, a phenotype that is more marked in differentiated THP-1 cells, although less so in DCs. This may relate to the more drastic infectivity defect of Vpx-deficient SIVMAC viruses in the latters. Vpx is to date the sole non-structural viral protein coded by primate lentiviruses capable of affecting the early phases of infection specifically in myeloid cells [50]. In the past, we have hypothesized that this could be due to the block of a restriction factor specifically expressed in these cells [50], [56]. In our hands, Vpx proteins interact functionally with A3A using three different methods: binding of endogenous A3A obtained from the lysate of differentiated THP-1 cells, co-IP after co-expression and DuoLink immunolocalisation that requires the physical proximity of two proteins for a positive signal to develop (less than 40 nm). This interaction leads to the degradation of A3A, but not of A3G, in HeLa cells and more importantly in primary DCs. Thus, our results are in agreement and extend a past study indicating that Vpx functionally interacts and degrades A3A after co-expression in established HeLa cells [72]. In the past, we have shown that SIVMAC Vpx modifies the susceptibility of myeloid cells to lentiviral infection with both cognate and non-cognate viruses (as HIV-1) [50], [81]. Under these conditions, the major effect of Vpx on the early phases of HIV-1 infection was exerted on the speed at which viral DNA accumulated, a phenomenon that we had hypothesized to protect vDNA from the negative effects of a particularly hostile intracellular milieu [50]. Given that knockdown of A3A does not affect the kinetics of vDNA formation following HIV-1 infection, we believe that A3A may not be the sole effector counteracted by Vpx (Figure S6). However, the fact that A3A silencing increases in a more dramatic manner the infectivity of Vpx-deficient viruses, as well as the fact that A3A is degraded upon expression of Vpx in HeLa cells and DCs indicates that A3A is an important target of Vpx during the infection of myeloid cells. Two recent studies identified SAMHD1 as a cellular factor that is specifically bound and degraded by Vpx [59] [60]. Contrarily to A3A, SAMHD1 is expressed in cells other than myeloid cells, as its isolation from HEK 293T cells indicates [59]. However, its antiviral effect is apparent mainly in myeloid cells, arguing for the need of a particular environment or for specific cellular co-factors. In light of the results presented here, A3A whose expression is truly restricted to the myeloid lineage (among cells of hematopoietic origin) may be a co-factor in the antiviral action of SAMHD1 or conversely, SAMHD1 may be an important cofactor in the action of A3A. The fact that Vpx seems to degrade two factors that hamper the early phases of infection specifically in myeloid cells evokes the possibility that they may act on the same pathway.
In conclusion, our results reveal a novel role for A3A in inhibiting the de novo infection of myeloid cells with HIV-1 and more generally with primate lentiviruses. In light of the extreme importance of myeloid cells during HIV-1 pathogenesis, these results call for a more ample evaluation of the impact of A3A during HIV-1 pathogenesis in cells that play a key instructive role in immune system responses.
293T and HeLa cells were maintained in Dulbecco's Eagle Modified Medium (DMEM) supplemented with 10% of fetal calf serum (FCS) and 100 U of penicillin/streptomycin. The THP-1 monocytic cell line was maintained in RPMI1640 medium supplemented with 10% FCS, HEPES 10 mM, 0,05 mM β-mercaptoethanol and 100 U of penicillin/streptomycin. THP-1 cells were differentiated for 24 hours in presence of 100 ng/ml of phorbol-12-myristate-13-acetate (PMA, SIGMA) to induce their differentiation into macrophage-like cells. Primary blood human monocytes and peripheral blood lymphocytes (PBLs) were purified from the blood of healthy donors obtained at the blood bank of Lyon (EFS-Lyon), as described before [67], yielding cell populations of purity greater than 95%. Monocytes were differentiated into macrophages or immature dendritic cells (DCs) upon incubation for 4 to 5 days in RPMI1640 complete medium supplemented with 100 ng/ml of MCSF or 100 ng/ml of GM-CSF/IL-4, respectively (AbCys). PBLs were either left in the absence of stimulus (quiescent) or stimulated with phytohemmagglutinin (PHA, SIGMA at 1 µg/ml) and IL2 (150 U/ml, AIDS Reagents and Reference Program of the NIH) for 24 hrs prior to infection. When indicated cyclohexamide (CHX, SIGMA) was added at a concentration of 100 µg/ml to arrest cell translation and MG132 (SIGMA) was added for 8 hrs at 10 µg/ml to block proteasomal degradation. When indicated IFNα2A (Tebu-Bio) was used at a final concentration of 1000 U for 24 hrs prior to infection. Anti-Flag, anti-HA, anti-actin and anti-tubulin monoclonal antibodies were purchased from SIGMA while, anti-HIV-1 Vif and anti-A3G Apo-C17 antibodies that recognize A3G and also A3A (recognized for its smaller size) were obtained from the AIDS Reagents and Reference Program of the NIH.
To determine the amount of type I interferons released in the supernatants of infected macrophages, we adopted a well established method that allows the simultaneous detection of multiple α and β IFN subtypes thanks to their ability to functionally inhibit the infection of target A549 cells with a GFP-bearing VSV. VSV is replicative virus that is extremely susceptible to the presence of type I interferons. Briefly, supernatant dilutions harvested from macrophage cultures were incubated with A549 cells prior to challenge with GFP-bearing VSV. The percentage of GFP-infected cells is assessed 24 hrs later by flow cytometry and the amount of type I IFN (U/ml) obtained against a standard curve. Plasmacytoid DCs (pDCs obtained by positive selection, Miltenyi) purified from the blood of healthy donors were used as control.
The plasmids coding HA-tagged A3A and A3G were a kind gift of Dr. Kenzo Tokunaga (National Institute of Infectious Diseases, Tokyo, Japan). Flag-tagged Vpx/Vpr coding plasmids were described before in [67]. The plasmid coding HIV-1 Vif was obtained from the AIDS Reagents and Reference Program of the NIH.
The HIV-1 and SIVMAC251 GFP-coding lentiviral vectors and their production has been described before [67]. Briefly, these single round infectivity vectors were produced by co-transfection of HEK293T cells with 3 plasmids coding: the packaging proteins Gag-Pro-Pol and viral non structural proteins (unless otherwise specified), a mini viral genome bearing a CMV-driven eGFP reporter and the Vesicular Stomatitis Virus G envelope (VSVg) conferring ample cellular tropism to viral vectors. Virions released in the supernatant of transfected cells were then purified through a 25% sucrose cushion, resuspended and their infection titers determined onto HeLa cells or by protein content against standards of known infectivity (exo-RT activity).
Replication competent R5-tropic HIV-1 virus (Yu2) was produced by transient transfection of HEK293T cells then used to infect target cells.
miR30-shRNA hairpins were designed using the Open Biosystem Website facility and cloned in the context of an HIV-1 lentiviral construct allowing their stable expression upon viral transduction (the targeted sequences are listed in Table 1, [67]). In this case, HIV-1 vectors bearing a miR30-shRNA expression cassette were produced as mentioned above and quantified by exo-RT over standards of known infectious titer. For A3A silencing, a mixture of the 3 targeting miR30-shRNAs was used in the transfection during the step of virus production and the overall DNA ratio of packaging to vector constructs was kept constant between specific and control miR30-shRNA coding vectors.
Infections were carried out for 2 hrs on 105 cells prior to extensive cells washing. The percentage of infected cells was monitored 3 to 4 days post infection by flow cytometry. For miR-shRNA lentiviral vector-mediated silencing, infections were carried out on 5.106 monocytes for 24 hrs prior to the addition of fresh media. Cells were then differentiated for 5 days in macrophages or DCs upon incubation with the appropriate cytokine cocktail. Cells were challenged with GFP-coding vectors four days afterwards. A multiplicity of infection (MOI) comprised between 3 and 5 was used for both silencing and challenging vectors.
Replicative infections were carried out as described above at MOI of 0,05 except that a fraction of the supernatant was stored and replaced with fresh media every 2–4 days. Viral spread through the culture was monitored by exo-RT. When used to reveal deamination of produced vDNA, virions were treated twice with 20 U/ml RNase free DNase RQ1 (Promega) for 2 hours at 37°C, and DNA extracted from infected cells was treated overnight at 37°C with DpnI. vDNA obtained after PCR amplification was cloned and sequenced.
M-SCF-differentiated macrophages were transfected with a total of 100 pmol of siRNAs (Darmacon) using Lipofectamine2000, according to the manufacturer's protocol (Invitrogen). After 6 hours, fresh media was added and cells were analyzed 48 hours post transfection by Western blot or challenged.
RNA was extracted on silica columns (NucleoSpin RNA XS; Macherey-Nagel) and reverse transcribed after DNAse treatment, using the SuperScript II Reverse Transcriptase (Invitrogen). Quantitative PCR was performed on a StepOnePlus Real-time PCR system (Applied Biosystems) using the FastStart Universal SYBR Green Master mix (Roche Diagnostics). Primers specific for each member of the APOBEC3 family genes and house-keeping genes (TBP, HPRT1 and RPL13A) had been previously described [44]. The level of expression of these genes does not vary among the different cell types and conditions used here. Results are expressed in mRNA copies per µg of total RNA.
For the analysis of vDNA produced upon infection, cells were lyzed 24 hours post infection and DNA amplified using primers specific for the GFP sequence carried by the challenge virus (see Table 1). Control infections were carried out in the presence of RT inhibitors (Nevirapine and AZT, at 20 µg/ml, obtained through the AIDS Reagents and Reference Program of the NIH). Values were first normalized for the amount of cellular DNA (actin) then subtracted for the values obtained for each sample in control infections performed in the presence of RT inhibitors. Experiments were discarded if the latter represented values higher than 1/10 of the ones obtained in the absence of RT inhibitors.
PMA-differentiated THP-1 cells were lysed in SD buffer 150 mM (50 mM Tris/HCl pH 7.4, 150 mM NaCl, 0,5% Triton-X100) in presence of a cocktail of protease inhibitors (Roche). The lysate was passed through a nichel column with immobilized HIV-1 Vpr protein, then on a second column containing SIVMAC Vpx. Columns were washed thrice with SD buffer 400 mM (50 mM Tris/HCl pH 7.4, 400 mM NaCl, 0,5% Triton-X100) prior to analysis by WB. Transfections of HEK293T cells were carried out in 6 well plates using a 1∶1 ratio of the indicated Vpx/Vpr and APOBEC3 proteins. Thirty-six hours post transfection, cells were lyzed in SD buffer 150 mM, then immunoprecipitated by addition of anti-Flag antibody-coated beads (Sigma) for 3 hours at +4°C. Beads were washed thrice with SD buffer 400 mM prior to WB analysis.
HeLa cells were seeded on coverlips and transfected with 300 ng of DNA using JetPei, according to the manufacturer's protocol (Polyplus). Twenty-four hours post transfection, cells were washed twice in PBS and fixed in 3% Paraformaldehyde in PBS (PFA). Free aldehydes were quenched by the addition of 50 mM NH4Cl in PBS before an additional wash in PBS. Cells were then permeabilized in 0,2% Triton-X100 for 5 min and saturated for 45 min in 0,2% gelatin from cold water skin fish (SIGMA) before immunostaining with DuoLink kit (Eurogentec), following the manufacturer's protocol. Image acquisition was carried out on an Axiovert 100 M Zeiss LSM 510 confocal microscope.
APOBEC3A: Ensembl:ENSG00000128383; APOBEC3B: Ensembl:ENSG00000179750; APOBEC3C: Ensembl:ENSG00000244509; APOBEC3D/E: Ensembl:ENSG00000243811; APOBEC3F: Ensembl:ENSG00000128394; APOBEC3G: Ensembl:ENSG00000239713; APOBEC3H: Ensembl:ENSG00000100298.
|
10.1371/journal.pntd.0005040 | Near-Infrared Spectroscopy, a Rapid Method for Predicting the Age of Male and Female Wild-Type and Wolbachia Infected Aedes aegypti | Estimating the age distribution of mosquito populations is crucial for assessing their capacity to transmit disease and for evaluating the efficacy of available vector control programs. This study reports on the capacity of the near-infrared spectroscopy (NIRS) technique to rapidly predict the ages of the principal dengue and Zika vector, Aedes aegypti. The age of wild-type males and females, and males and females infected with wMel and wMelPop strains of Wolbachia pipientis were characterized using this method. Calibrations were developed using spectra collected from their heads and thoraces using partial least squares (PLS) regression. A highly significant correlation was found between the true and predicted ages of mosquitoes. The coefficients of determination for wild-type females and males across all age groups were R2 = 0.84 and 0.78, respectively. The coefficients of determination for the age of wMel and wMelPop infected females were 0.71 and 0.80, respectively (P< 0.001 in both instances). The age of wild-type female Ae. aegypti could be identified as < or ≥ 8 days old with an accuracy of 91% (N = 501), whereas female Ae. aegypti infected with wMel and wMelPop were differentiated into the two age groups with an accuracy of 83% (N = 284) and 78% (N = 229), respectively. Our results also indicate NIRS can distinguish between young and old male wild-type, wMel and wMelPop infected Ae. aegypti with accuracies of 87% (N = 253), 83% (N = 277) and 78% (N = 234), respectively. We have demonstrated the potential of NIRS as a predictor of the age of female and male wild-type and Wolbachia infected Ae. aegypti mosquitoes under laboratory conditions. After field validation, the tool has the potential to offer a cheap and rapid alternative for surveillance of dengue and Zika vector control programs.
| Aedes aegypti is the principal vector for dengue, chikungunya and Zika viruses. These viruses require a period of development inside the mosquito before they can be transmitted to humans. Depending on environmental factors, dengue and Zika viruses take an average of 8–10 days to replicate inside the mosquito [1,2] whereas chikungunya virus takes 2–7 days to replicate [3]. The age of mosquitoes is therefore a critical determinant of disease transmission. A mosquito control strategy utilizing an endosymbiotic bacterium Wolbachia pipientis, has been proven effective at blocking dengue transmission in Ae. aegypti mosquitoes [4]. For effective virus blocking, mosquitoes infected with the wMel strain of Wolbachia must survive long enough to cover the extrinsic incubation period (EIP) of infected mosquitoes while those infected with wMelPop are expected to have a reduced lifespan to limit the period available for virus replication. These strategies therefore require routine monitoring using age grading techniques. In this study, we investigated the applicability of a rapid and cost effective near-infrared spectroscopy (NIRS) technique as an alternative age grading tool for wild-type and Wolbachia infected Ae. aegypti mosquitoes. We show that NIRS can rapidly predict on average, the age of male and female Ae. aegypti to ±3 days of their true age and can determine which mosquitoes are old enough to be potentially infectious with an accuracy of up to 91%.
| The mosquito Aedes aegypti is the primary vector for dengue, Zika and chikungunya viruses. Up to 100 million dengue cases occur annually [5,6] and an estimated 440,000–1,300,000 Zika cases were reported in early 2016. Notably, 3893 babies born to Zika infected mothers have been diagnosed with microcephaly [7,8]. Both Zika and dengue viruses are transmitted by female Ae. aegypti mosquitoes carrying viruses in their salivary glands. Due to the period required by the virus to replicate inside the mosquito, Ae. aegypti mosquitoes are in most cases only capable of transmitting the Zika or dengue viruses when they are at least 8 days old [1,2,9].
The success of existing arbovirus vector control programs for mosquito-borne viruses recommended by the World Health Organization such as targeted residual spraying and space spraying is dependent on their ability to shorten the lifespan of the mosquito and to limit the period available for virus development. Alternative biological control strategies under trial involve the release of Ae. aegypti mosquitoes transinfected with the wMel strain of Wolbachia pipientis for the suppression of arbovirus transmission [4] or the pathogenic wMelPop strain for arbovirus and/or vector population suppression [10–12]. Although the efficacy of Wolbachia induced arbovirus interference was first demonstrated against dengue and Chikungunya [13], it has now also been demonstrated against Zika virus [14,15]. Unlike the life-shortening strain wMelPop, the success of wMel as a biological control agent is dependent upon the survival of infected mosquitoes [4].
Following the outbreak of Zika virus in South America in 2015, there is renewed interest in defining mosquito survival characteristics as a means of evaluating vector control strategies that affect adult mosquito survival. The efficacy of the WHO recommended control interventions [16], would therefore be defined by their ability to effectively eliminate the old and potentially infectious population. To accurately define these characteristics, evaluations of current interventions would require assessments of mosquito populations on large-scale using rapid and accurate age grading techniques.
Traditionally, entomologists have used techniques based on dissections of the female reproductive system to estimate mosquito age and survival. These include the Detinova technique for differentiating parous and nulliparous mosquitoes [17] and the Polovodova technique for determining the number of times mosquitoes have laid eggs [18]. However, these techniques are time consuming and labour-intensive. Moreover, interpretation of the diagnostic changes to ovarian morphology can be problematic [19]. As a result, estimates of population age structure based on these techniques often lack statistical power as only a small sample size can be dissected within a reasonable timeframe.
Analysis of age-related changes of cuticular hydrocarbons by gas chromatography [20–22] and transcriptional profiling [23] have been evaluated as alternative age grading techniques for Aedes mosquitoes. However, their high costs may restrict their utility and sustainability for large-scale field trials or control programs, particularly in areas where resources are limited. Although age related proteomic changes recently reported for Ae. aegypti [24] may offer cost effective and rapid alternative means of age assessments, these techniques are still early in development.
New approaches are required that can rapidly and cost-effectively assess large numbers of field specimens. In our previous studies, we reported the use of the near-infrared spectroscopy (NIRS) technique for age and species prediction. The tool was first applied to predict the age of female laboratory reared An. gambiae and An. arabiensis and to differentiate these sibling species [25]. It was then used to age grade and differentiate species of semi field reared An. gambiae and An. arabiensis [26], age grade laboratory reared An. arabiensis undergoing various physiological changes [27], preserved specimens [28–30] and to age grade Anopheles gambiae sensu lato mosquitoes reared from wild larvae that were either susceptible or resistant to pyrethroids [31]. More recently NIRS successfully predicted the ages of laboratory Ae. aegypti maintained on a varying larval and adult diets [32]. From laboratory and semi-field studies, we have shown that the accuracy of NIRS for An. gambiae s.s. and An. arabiensis ranges between 79–90% and 80–90% for age grading into <7 d and ≥ 7 d old age groups and for species differentiation, respectively.
Given its rapidity and relative ease of application, NIRS represents a unique opportunity to develop a viable alternative to current ovarian dissections and molecular techniques for age grading. The use of a NIR spectrometer for this purpose is non-destructive, rapid, and requires little training. NIRS facilitates the analysis of hundreds of mosquitoes just in one day because it takes 15 seconds to prepare and collect spectra from a mosquito, without reagents or sample preparation procedures. Moreover, samples can be scanned either fresh or preserved [28–30].
Here, we show that NIRS may be used to predict the age of both male and female wild-type Ae. aegypti mosquitoes up to 30 d old. We also examined the ability of NIRS to predict the age of Ae. aegypti females and males infected with wMel and wMelPop strains of Wolbachia pipientis.
Ethics approvals were obtained for routine blood feeding of mosquito colonies from the QIMR Berghofer Medical Research Institute (QIMR HREC P1162). Written informed consent was provided by all adult volunteers involved in blood feeding, and volunteers were given the right to refuse to participate or withdraw from the experiment at any time.
Colonies of wild-type Ae. aegypti and Ae. aegypti infected with wMel and wMelPop were acquired and maintained at the insectary of QIMR Berghofer Medical Research Institute as previously described [33]. Adults were given a 24 hr window to emerge. Wild-type females and males were collected at 1, 5, 9, 13, 17, 21, 25 and 30 d post emergence. Adult wMel strain mosquitoes were collected at 1, 5, 10, 15, 19 and 20 d post emergence. Adult wMelPop strain mosquitoes were collected at 1, 5, 10, 15 and 19 d post emergence. All mosquitoes were collected either unfed (1 and 5 d old) or blood fed but after oviposition (> 5 days old mosquitoes). Adults were knocked down with carbon dioxide and stored in RNAlater solution. To allow for RNAlater penetration, samples were stored overnight in a 4°C refrigerator and then stored at -20°C for 2 months before scanning [28].
All mosquitoes collected were transferred to Ifakara Health Institute, Tanzania for scanning. Prior to scanning, residual RNAlater was removed from the mosquito specimens by blotting with paper towels. A maximum of 25 mosquitoes at a time were then positioned on a spectralon plate (ASD Inc, Boulder, CO), ventral side up. At least 40 mosquitoes at each age for all species were scanned using a LabSpec 5000 NIR spectrometer (ASD Inc, Boulder, CO), according to previously described protocols [14]. To collect the spectra, the heads and thoraces of mosquitoes were scanned under a 3 mm-bifurcated fibre optic probe containing six collection fibres and 33 illumination fibres. Typical raw spectra collected from the head and thorax of female wMel infected Ae. aegypti at 1, 5 and 19 d age points are shown in Fig 1.
Spectra were analyzed using Grams IQ software (Thermo Galactic, Salem, NH). Due to low light energy at shorter wavelengths and low sensor sensitivity at longer wavelengths, spectra appeared noisy outside the 500–2350 nm region. Therefore results were analyzed using spectra measured within this region. The relationship between spectra and age were analyzed by partial least squares (PLS) regression.
For wild-type Ae. aegypti, mosquitoes were divided into a training set and a validation set. The training set is used for developing a calibration model using cross validation analysis. Cross-validation is a “leave-one-out” self-prediction method where mosquitoes from a set are used to predict the age of mosquitoes from that same set. It is used to select the factors required for the calibration of the predictive model before running the validation set. All Wolbachia infected mosquitoes were analyzed using the cross validation method. The number of factors used in developing models was determined from the cross-validation prediction residual error sum of squares (PRESS) and regression coefficient plots. An example of a coefficient plot used for predicting the age of female wild-type Ae. aegypti is shown in Fig 2. Analysis of variance (ANOVA) was applied to test for statistical differences between mean predicted ages using the Statistical Package for Social Sciences 22 (IBM, Armonk, NY). Tukey post hoc analysis in ANOVA was applied to test for statistical differences between age groups. Actual age and predicted age were coded as independent and dependent variables, respectively. The relationship between true and predicted age was assessed by Spearman’s rank correlation coefficient analysis.
Except for 1 d old mosquitoes, the mean predicted age of female Ae. aegypti mosquitoes was within ± 2d of the actual age across all age groups (Table 1; Fig 3A). An accuracy of 91% (N = 501) was achieved if female mosquitoes excluded from the model were simply grouped into two age categories that separate mosquitoes that are unlikely to be infectious (< 8 d) from those that are old enough to have survived the dengue/Zika incubation period and therefore would be potentially infectious (≥ 8 d). Additionally, Spearman’s correlation analysis indicated a strong positive correlation between the predicted and the actual age for both the training set (R2 = 0.84; P <0.001) and the validation set (R2 = 0.83; P <0.001). Tukey Post hoc analysis indicated that female Ae. aegypti could be differentiated into 5 age groups (1–5, 6–9, 10–13, 14–17 and >17 d old).
The age prediction of male mosquitoes that were excluded from the training set was generally within ± 3 d of the actual age. Moreover, an overall accuracy of 87% (N = 253) into < 8 d and ≥ 8 d old age groups was achieved. Spearman correlation analysis found a highly positive correlation between the predicted age and the actual age for both the training set (R2 = 0.77; P <0.001) and the validation set (R2 = 0.78; P <0.001). Tukey post hoc comparison test between age groups of male Ae. aegypti indicated that mosquitoes could be differentiated into four age groups (<9, 9–13, 14–17 and >17 d old) (Table 1; Fig 3B).
Using the cross validation analysis, females (N = 284) and males (N = 277) infected with the wMel strain were differentiated into < 8 d and ≥ 8 d old age groups with accuracies of 83%. Males (N = 234) and females (N = 229) infected with the wMelPop strain were both differentiated into the two groups with an accuracy of 78%. Spearman correlation analysis found a positive correlation between the predicted age and the actual age of females (R2 = 0.71; P<0.001) and males (R2 = 0.80; P<0.001) infected with wMel and females (R2 = 0.80; P<0.001) and males (R2 = 0.68; P<0.001) infected with wMelPop. All Wolbachia infected females and males were generally differentiated into 4 age groups (<5, 5–9, 10–15 and >15 d old) (Table 2).
Our findings demonstrate the potential of the NIRS as a rapid technique for identifying the age of wild-type and Wolbachia infected Ae. aegypti mosquitoes. Mosquito survival is a fundamental parameter of vectorial capacity. As 1–2 days is required for blood feeding and at least 7 days is required for virus replication [2,9,34], the average infectious age of Ae. aegypti, the principal vector for Zika and dengue viruses, is considered to be at least 8 days. Wolbachia-based strategies utilizing the wMel strain require infected mosquitoes to survive and mate effectively with wild mosquitoes to drive the bacteria through populations. Alternatively, the life shortening properties of the pathogenic wMelPop strain may be harnessed to crash local vector populations [11]. The validation of either strategy will require accurate mosquito age grading techniques that can function against a background of Wolbachia infection.
Recently, Liebman and colleagues used NIRS to predict the age of laboratory reared female Ae. aegypti mosquitoes maintained on varying larval and adult diet up to 16 days post emergence [32]. We report on the ability of NIRS to age grade female and male wild-type Ae. aegypti up to 30 d old as well as females and males Ae. aegypti infected with Wolbachia up to 20 d old. Age predictions were determined on a continuous age scale and into < 8 d or ≥ 8 d old age groups. Although shorter EIPs have been reported for dengue viruses [35], eight days was the favored cut off point because it is widely quoted as the average age at which Ae. aegypti or Ae. albopictus are able to transmit dengue or Zika viruses [2,9,34] and the best accuracy was achieved at this cut off point.
Wavelengths ranging from 500–2350 nm were analyzed. This range comprises carbon-hydrogen (C-H) functional groups at 1220, 1450, 1700 and 1765 nm and protein (N-H) functional groups at 1510, 2055, 2060, 2180 and 2300 nm. Both cuticular hydrocarbons [20,21] and proteins [24] have previously yielded biomarkers suitable for age grading Ae. aegypti mosquito species. Overall, age prediction of wild-type females and Wolbachia infected females across all age categories assessed was within ±3–5 days of actual age. Five day old females infected with Wolbachia were the least accurately predicted. It may be that few developmental changes occur between 5 and10 days. It was recently reported by Hugo and colleagues, who examined changes in protein abundance with age of Ae. aegypti, that age-related changes for the majority of the proteins reported occurred between 1 and 5 d with little or no further protein changes occurring among older age groups [24]. These proteomic and transcriptome approaches may ultimately help identify biomarkers contributing to age related variation in NIRS spectra, provide further validation and improve the prediction accuracy of NIRS. We found strong positive correlation between the actual age and the predicted age for both wild-type and Wolbachia infected mosquitoes.
Age prediction accuracy of male wild-type and male mosquitoes infected with Wolbachia was impressive. NIRS predicted the ages of male wild-type, wMel and wMelPop mosquitoes into < 8 and ≥ 8 d age groups with 87%, 83% and 78% accuracy, respectively. This is the first investigation to report the ability of NIRS to predict the age of male Ae. aegypti mosquitoes. Despite the fact that they are not disease vectors, the survival of male mosquitoes would be a useful indicator of the success or failure of control strategies utilizing Wolbachia [10], genetic modification approaches requiring male competitiveness [36,37], a strategy to induce sex-ratio distortion of mosquito populations towards males [38] and strategies that seek to release sterile males [39]. The efficacy of any sterile male technique, including those that utilise Wolbachia infected males, relies on the ability of the male mosquito to compete and mate with the wild-type population. Nonetheless, to date only one age grading technique has been reported for male mosquitoes. A technique based on the frequency at which spermatocysts are found in male reproductive organs of An. gambiae s.s. and Anopheles culicifacies and the relative size of the sperm reservoir to differentiate mosquitoes into ≤ 4 d old and > 4 d old [40,41]. The difficult dissections and complex quantitative models involved in the application of this age grading technique, as pointed out by the authors, may limit its application on field related studies. Given its rapidity and simple prediction models, NIRS would be a suitable complementary age grading tool for rapidly differentiating young from old male mosquitoes.
NIRS offers a considerably higher throughput than alternative mosquito age grading approaches. We estimate that we can analyze approximately 1000 mosquitoes per day. Comparatively, the analysis of samples by cuticular hydrocarbon or transcriptional profiling age grading techniques currently allows an average of only 20–30 samples to be processed per day (averaged over preparation periods) [42]. With the capacity for larger sample sizes to be processed, NIRS has the potential to reduce sampling error associated with age prediction estimates in comparison to the alternative age grading techniques. The increased capacity and reduced sampling bias has the potential for substantially increasing the accuracy of population survivorship estimates, provided age prediction models are constructed without bias [22].
Accurate determination of mosquito population survivorship is key to understanding transmission risk and to the success of vector control strategies. Until recently, it has been impractical or impossible to accurately assess the age of Ae. aegypti in the field. Traditional dissection methods are technically challenging and only accurately differentiate parous from non-parous mosquitoes. Excitingly, our results that NIRS can accurately age Ae. aegypti mosquitoes over a 3 week lifespan gives us a tool for assessing transmission risk and understanding seasonality. NIRS has considerable advantages over conventional techniques given that results can be gained in real-time, without reagents or sample preparation. To improve the accuracy, alternative statistical methods are being developed. However, as certain age groups may be biochemically indistinguishable from each other and since NIRS relies on assessing alterations in biochemical signatures, it should be acknowledged that shortcomings in age prediction may stem from the underlying biochemistry as opposed to the statistical techniques applied. Although our study is still preliminary, our results demonstrate that upon further calibration, NIRS could be a potentially more accurate tool for predicting the age of Ae. aegypti compared to transcriptional profiling and cuticular hydrocarbon techniques. Follow up studies conducted under semi-field or field environments to improve the current calibration models have been envisaged. These and our recently published findings that NIRS can detect Wolbachia infections in Ae. aegypti [33], suggest a strong potential role for NIRS as a rapid surveillance tool for Ae. aegypti control programs.
|
10.1371/journal.pgen.1006034 | Discovery of Genetic Variation on Chromosome 5q22 Associated with Mortality in Heart Failure | Failure of the human heart to maintain sufficient output of blood for the demands of the body, heart failure, is a common condition with high mortality even with modern therapeutic alternatives. To identify molecular determinants of mortality in patients with new-onset heart failure, we performed a meta-analysis of genome-wide association studies and follow-up genotyping in independent populations. We identified and replicated an association for a genetic variant on chromosome 5q22 with 36% increased risk of death in subjects with heart failure (rs9885413, P = 2.7x10-9). We provide evidence from reporter gene assays, computational predictions and epigenomic marks that this polymorphism increases activity of an enhancer region active in multiple human tissues. The polymorphism was further reproducibly associated with a DNA methylation signature in whole blood (P = 4.5x10-40) that also associated with allergic sensitization and expression in blood of the cytokine TSLP (P = 1.1x10-4). Knockdown of the transcription factor predicted to bind the enhancer region (NHLH1) in a human cell line (HEK293) expressing NHLH1 resulted in lower TSLP expression. In addition, we observed evidence of recent positive selection acting on the risk allele in populations of African descent. Our findings provide novel genetic leads to factors that influence mortality in patients with heart failure.
| In this study, we applied a genome-wide mapping approach to study molecular determinants of mortality in subjects with heart failure. We identified a genetic variant on chromosome 5q22 that was associated with mortality in this group and observed that this variant conferred increased function of an enhancer region active in multiple tissues. We further observed association of the genetic variant with a DNA methylation signature in blood that in turn is associated with allergy and expression of the gene TSLP (Thymic stromal lymphoprotein) in blood. Knockdown of the transcription factor predicted to bind the enhancer region also resulted in lower TSLP expression. The TSLP gene encodes a cytokine that induces release of T-cell attracting chemokines from monocytes, promotes T helper type 2 cell responses, enhances maturation of dendritic cells and activates mast cells. Development of TSLP inhibiting therapeutics are underway and currently in phase III clinical trials for asthma and allergy. These findings provide novel genetic leads to factors that influence mortality in patients with heart failure and in the longer term may result in novel therapies.
| Heart failure (HF) is a common clinical condition in which the heart fails to maintain blood circulation adequate to meet the metabolic demands of the body without increased cardiac filling pressures. HF is the result of chronic ventricular remodelling initiated by myocardial injury, volume/pressure overload, or intrinsic cardiomyopathic processes. Progression of HF is a complex process involving many tissues, driven by activation of neurohormonal pathways, which induce gradual myocardial hypertrophy, ventricular dilation, and deterioration of cardiac function, often resulting in death from low cardiac output, arrhythmia, or thromboembolic complications [1]. Activation of such neurohormonal pathways in the short term increases cardiac output when necessary. However, long-term activation results in accelerated ventricular remodelling and myocyte death. Inhibitors of deleterious neurohormonal pathways, including adrenergic [2–4] and renin-angiotensin-aldosterone (RAAS) [5–8] pathways have been shown to improve ventricular function and survival in patients with HF and are the mainstay of current pharmacological treatment of HF [9–10]. Despite advances in therapy with neurohormonal antagonists, mortality after onset of HF remains high [9–13] and continued progress to identify additional therapeutic targets is needed.
Genome-wide association (GWA) studies have the potential to identify in an agnostic manner genetic variants related to clinical outcomes in humans and has led to the identification of novel pathways [14] and potential treatments [15] for cardiovascular traits. Heritable factors have been shown to be predictive of mortality in certain heart failure patients [16]. We therefore implemented a genome-wide association approach to identify novel molecular determinants of mortality in patients with new-onset HF.
We expanded our previously published GWA study [17] of HF mortality with additional samples and extended follow-up in Stage 1. Stage 1 included 2,828 new-onset HF patients from five community-based cohorts, thus representative of the general population of HF patients, as part of the Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE) consortium [18]: the Atherosclerosis Risk in Communities (ARIC) Study, the Cardiovascular Health Study (CHS), the Framingham Heart Study (FHS), the Health, Aging and Body Composition (Health ABC) Study, and the Rotterdam Study (RS). Cohorts are described in detail in S1 Text. HF was defined using international published criteria as outlined in S1 Table. Subjects in Stage 1 cohorts were of European ancestry, predominantly male, and approximately 20–30% had a history of myocardial infarction at the time of HF diagnosis. Additional characteristics are shown in Table 1. During an average follow-up time of 3.5 years, 1,798 deaths occurred. The sample-size weighted average 1-year mortality rate was 28%. Among deaths, 51% were classified as cardiovascular, 19% were due to neoplasms, 10% were respiratory deaths, and the remaining were due to other miscellaneous causes. Genotyping using high-density Illumina or Affymetrix single nucleotide polymorphism (SNP) arrays, followed by imputation to the HapMap CEU release 22 imputation panel was performed in each cohort. Population stratification was assessed and corrected in each cohort as described in S1 Text. Association with time to death following HF diagnosis was examined in each cohort using Cox proportional hazards models with censoring at loss to follow-up. Mild inflation of test statistics was observed only in the Framingham Heart Study (FHS) as shown in S1 Fig (λGC = 1.07, other cohorts ≤ 1.03), and genomic control was applied in each individual study. In the meta-analysis of all cohorts, there was no evidence of inflated test statistics overall (λGC = 1.00) as shown in S2 Fig, so no further genomic control was needed. Results for all SNPs across the genome are plotted in S3 Fig. Single nucleotide polymorphisms (SNPs) passing a significance threshold specified a priori as P < 5.0x10-7, as used in our previous article [17], were carried forward to a second stage of genotyping in independent cohorts.
Five SNPs on chromosome 5q22 and one SNP on chromosome 3p22 passed the pre-specified P-value threshold. Results for all six SNPs are shown in Table 2 and S3 Table. The five SNPs on chromosome 5q22 were highly correlated (pairwise r2 > 0.9). Two sentinel SNPs, rs9885413 and rs12638540, on chromosomes 5q22 and 3p22, respectively, were next genotyped in 1,870 European-ancestry subjects with new-onset HF from four independent cohorts in Stage 2: Malmö Diet and Cancer, Malmö Preventive Project, Physicians’ Health Study, and the PROSPER trial. Characteristics of populations in Stage 2 are shown in S2 Table. During an average sample-size weighted follow-up of 4.3 years in Stage 2 samples, 889 patients died. We observed evidence of association with mortality for rs9885413 on chromosome 5q22 (P = 0.006) but not for the SNP rs12638540 (P = 0.18) which reached nominal significance in our previous analysis [17]. Results for both SNPs are shown in Table 2. In the combined results from Stages 1 and 2, rs9885413 was associated with a 36% relative increase in mortality per minor allele (P = 2.7x10-9). There was no evidence for effect heterogeneity across cohorts in the two stages (P for heterogeneity = 0.39) as shown in S4 Table. The SNP had a similar minor allele frequency (MAF = 0.07) across cohorts. Information on cause-specific mortality was available from death certificates in a subset of cohorts (S5 Table) and was explored descriptively due well-known problems with substantial misclassification in death certificate data and low power for agnostic GWAS of individual causes. The minor allele frequency was slightly higher for several causes of death associated with heart failure, including renal, pulmonary and endocrine mortality and death from ischemic heart disease.
We next examined whether rs9885413 on chromosome 5q22 that was associated with HF mortality was also associated with differences in myocardial structure and function, which could potentially mediate the association (S6 Table). In 12,612 individuals from the EchoGen Consortium [19], the SNP was not associated with major echocardiographic characteristics. The SNP rs9885413 was not associated with incident HF in 20,926 individuals from the general population in the CHARGE-HF study [20], or with cardiac endocrine function, as determined by plasma levels of atrial and B-type natriuretic peptides (all P > 0.05), in a GWA study of 5,453 individuals from the population-based Malmö Diet and Cancer study [21]. No association was observed with electrocardiographic measures of cardiac conduction (n = 39,222) [22] or repolarization (n = 74,149) [23], which confer risk of ventricular arrhythmia, or with sudden cardiac death in 4,496 sudden death cases and over 25,000 controls from the general population (described in S1 Text).
The lead SNP rs9885413 on chromosome 5q22 that was associated with mortality is located in an intergenic region, 100 kb downstream of the gene SLC25A46, 114 kb upstream of TMEM232, and 230 kb upstream of TSLP as shown in Fig 1. The SNP is not in linkage disequilibrium with any known coding SNP in the 1000 Genomes Project database (no coding SNP with r2 > 0.01 to the sentinel SNP). We therefore sought to evaluate gene regulatory functions of this SNP.
In 129 human tissues from the ROADMAP Epigenomics project [24], we studied whether rs9885413 or strongly correlated SNPs (a total of 9 at r2 > 0.8) are located in regulatory regions, as determined by histone modification patterns. None of the 9 SNPs was located in an active regulatory region in cardiac tissues (S7 Table), but rs9885413 was located in a predicted enhancer in several epithelial or mesenchymal tissues, including keratinocytes, gastrointestinal cell types and adipose cells (Fig 2 and S7 Table). Regulatory motif annotations in HaploReg indicate that the SNP causes a change in a regulatory motif predicted to bind the transcription factor NHLH1 as shown in S8 Table. Interestingly, NHLH1-null mice have been shown to be predisposed to premature, adult-onset unexpected death in the absence of signs of cardiac structural or conduction abnormalities, in particular when mice were exposed to stress [25]. Little is known about the function of NHLH1, but it is widely expressed in human tissues and has been shown to regulate expression of key inflammatory genes [26].
To experimentally test the effect of rs9885413 on enhancer activity, the 100 bp region flanking the SNP (50 bp on either side) was cloned into a reporter vector and transfected into HEK293 cells expressing NHLH1 (S1 Text). Luciferase activity measured after 24 hours was 4-fold higher with a construct corresponding to the risk allele as compared to the wild-type allele (S4 Fig, P < 0.001), indicating that the risk allele of rs9885413 substantially increases enhancer activity.
We next explored the association of rs9885413 with DNA methylation at the locus, providing functional evidence of epigenetic association and regulation of gene expression. DNA methylation was determined by a microarray assaying in total over 480 000 CpG methylation sites in whole blood samples from 2408 participants of the FHS. Of the 84 CpG methylation sites on the microarray within +/- 500 kb of the SNP, two were significantly associated with rs9885413: cg21070081 (beta 0.017 per T allele, P = 9.0x10-69) and cg02061660 (beta -0.015 per T allele, P = 4.5x10-40), thus constituting strong methylation quantitative trait loci (mQTLs) at the locus. Other, correlated SNPs at the locus were more strongly associated with each of these mQTLs as shown in S5 Fig: rs244431 for cg21070081 (P = 6.7x10-369) and rs72774805 for cg02061660 (P = 7.0x10-85). The SNP rs72774805 (perfect proxy SNP rs3844597 used) but not rs244431 was associated with heart failure mortality (P = 3.3x10-3 and 0.08, respectively), indicating that the methylation site cg02061660 is more strongly related to the underlying signal for heart failure mortality. The association of rs9885413 with lower probability of methylation at cg02061660 was replicated in 731 participants from the Rotterdam study (beta -0.029 per T allele, P = 1.7x10-11). Adjustment for cell types from direct measurement instead of estimates from methylation patterns did not abolish the association (beta -0.029 per T, P = 1.2x10-6). Interestingly, differential methylation at this CpG site was also correlated with a SNP at the locus previously associated with allergic sensitization [27] (rs10056340, P = 4.7x10-29 for mQTL), suggesting a link to inflammatory disease. This SNP was also modestly correlated with rs9885413 (r2 = 0.28) and associated with heart failure mortality (P = 0.01). The association of cg02061660 with rs9885413 (P = 0.52) and rs10056430 (P = 0.87) was abolished in analyses conditioning for rs72774805, for which the association was also markedly attenuated (P = 7.0x10-33 and 2.1x10-46, respectively) indicating that these correlated SNPs may reflect the same underlying signal.
We further assessed the association of rs9885413 with gene expression. No gene was significantly associated with rs9885413 in the diverse tissues from the Gene-Tissue Expression (GTEx) project [28] after correction for multiple tests (S1 Text, S9 Table), although conclusions were limited by a small sample size.
We next assessed association of the SNP with gene expression in two large datasets with each of the tissues most relevant for the phenotype under study: heart tissue and whole blood. We observed no convincing evidence of association (S1 Text, S10 Table) with gene expression in 247 left ventricular samples from patients with advanced heart failure (n = 116) undergoing transplantation and from unused donors (n = 131). Finally, we tested the association of rs9885413 with the expression of genes at the locus in whole blood from 5257 FHS participants [29], and with DNA methylation at cg02061660 among 2262 FHS participants. All five genes at the locus (Fig 1) except TMEM232 were expressed in blood. We did not observe association of the SNP rs9885413 with any transcript, but expression of one gene (TSLP) was significantly associated with the methylation status of cg02061660 (P = 1.1x10-4).
The TSLP gene encodes a cytokine released from epithelial cells that induces release of T cell-attracting chemokines from monocytes, promotes T helper type 2 cell responses, enhances maturation of dendritic cells and activates mast cells. It has also been linked to angiogenesis and fibrosis. A monoclonal antibody targeting and inhibiting TSLP is currently in clinical phase III trials for asthma and allergic inflammation after a promising phase II trial [30–32]. In the myocardium, the TSLP gene has very low expression (S10 Table) but expression has been described in mature myocardial fibroblasts, which are abundant in the myocardium but of substantially smaller volume than cardiomyocytes and likely contribute little to the overall myocardial RNA pool [31,32].
To examine whether the transcription factor NHLH1 affects the expression of any of the five genes in the locus (Fig 1), we knocked down NHLH1 in HEK293 cells using siRNAs. A 50% decrease in NHLH1 mRNA levels was seen 48 hours after transfection, confirming efficient knock down (p<0.05, S6A Fig). TSLP was the only gene at the locus affected by NHLH1 knock down, showing a 30% decrease compared to cells transfected with negative control siRNA (p<0.05, S6A Fig). Moreover, we observed a dose-response relation between level of NHLH1 knockdown and expression of TSLP in HEK293 cells (r2 = 0.74, p<0.0001, S6B Fig).
Finally, distribution of the risk allele of rs9885413 in human populations was assessed using data from HapMap phase II. The derived (non-ancestral) T allele (risk allele for mortality) was highly differentiated among human populations (S7 Fig) having risen to an allele frequency of 0.59 in a Nigerian population (HapMap YRI sample) but only 0.06 in a European population (CEU sample). The fixation index (Fst), a measure of population differentiation in allele frequencies, for comparison of YRI and CEU was 0.48 and more extreme Fst was observed in only 2.4% of SNPs in the HapMap phase 2 dataset. Consistent results were observed for another signature of recent positive selection, based on longer runs of haplotype homozygosity in carriers of the derived allele (standardized integrated haplotype score -0.766 in YRI, where negative score values indicate longer haplotypes on the background of the derived allele) [33]. These observations are consistent with positive selection in recent human history, with a selective sweep resulting in high frequency of the derived allele in western African populations. These findings are of particular interest as HF mortality is well known to be higher in populations of African ancestry, although the current study has not tested for the association with HF mortality in such populations [34].
We identified a SNP on chromosome 5q22 associated with increased mortality in subjects with HF. Although previous genome-wide association studies have described hundreds of loci associated with risk of disease onset, few have examined prognosis in subjects with manifest disease. This approach has the potential to generate targets for novel disease-modifying medications. Through a series of analyses in silico and in vitro we show that the SNP is located in an enhancer region, and confers increased activity of this enhancer.
Interestingly, mice deficient in the transcription factor NHLH1 predicted to bind a motif in this enhancer region have been reported to be predisposed to premature, adult-onset unexpected death in the absence of signs of cardiac structural or conduction abnormalities. NHLH1 has also been shown to regulate expression of key inflammatory cytokines such as interleukin-6 and tumor necrosis factor α. The SNP was not associated with any electrocardiographic, endocrine, or echocardiographic marker of increased risk in the general population, suggesting a mechanism specific to heart failure, an extracardiac pathway of importance in cardiac pathophysiology, or interaction with therapy for heart failure which we were unable to further test given the inception cohort design of this study. We also did not observe any robust eQTL associations for the SNP in heart. The SNP was however associated with a DNA methylation signature in whole blood that was also associated with a SNP previously associated with allergy, and with expression of the cytokine TSLP in blood. Knockdown of NHLH1 also resulted in lower expression of TSLP in HEK293 cells. This non-coding SNP may thus exert an influence on TSLP expression via altered NHLH1 enhancer function and DNA methylation at the methylation site cg02061660. Detailed characterization of causal variants and different association signals at the locus would however require finemapping and sequence data.
The TSLP cytokine is released from epithelial cells and fibroblasts and is considered important in initiation of inflammatory responses to tissue damage, particularly in the type 2 T-helper (Th2) pathways. Th2 pathways are central in the response to extracellular parasites but also play a key role in the pathophysiology of allergies and hypersensitivity reactions. A small subset of HF is known to be caused by Th2-mediated inflammation (eosinophilic cardiomyopathy), yet Th2 cells have received limited attention in HF pathophysiology. Recent experimental work implicates an important role of T-helper cells in HF progression for both systolic and diastolic heart failure, but has mainly focused on type 1 T-helper pathways [35,36]. It remains unclear if the mechanism for rs9885413 is through a specific etiology characterized by high mortality such as eosinophilic cardiomyopathy or a pathway involved in outcomes with manifest disease. The lack of association with HF incidence suggests that it may not act through incidence of a specific etiology, although firm conclusions are limited by sample size. We did not observe significant associations of the SNP with gene expression in any tissue. It is possible that adequately powered samples with a specific cell subtype in a specific context is needed to detect such associations, as illustrated by a recent study which only observed certain eQTLs with single-cell but not across averaged cells [37]. Indeed, baseline expression of TSLP was low in our samples, and is induced by tissue injury, microbes, viruses and proinflammatory cytokines [38].
Evidence of recent positive selection in individuals of African descent suggests that the HF risk allele may have been beneficial in some environments in recent human history. Inflammatory pathways are enriched for signals of recent positive selection, reflecting that infectious disease has been an important cause of mortality throughout recent evolution. Genes such as HBB and APOL1 have also been reported to have been subject to recent positive selection in Africa by conferring protection against infectious diseases such as Malaria and Trypanosomiasis (sleeping sickness) [39], and APOL1 alleles have also been linked to cardiovascular disease [40]. As cardiovascular disease and heart failure often presents after reproductive age, increased mortality in such patients would not be expected to exert purifying (negative) selective pressure. Whether SNPs at 5q22 contribute to higher mortality in subjects of African ancestry remains to be shown.
Thus, although additional work is needed to further clarify the tissues and pathways perturbed by this genetic variant and the mechanisms linking it to mortality in HF patients, the current findings implicate rs9885413 as a novel marker of increased risk among patients with HF. Complementary epigenomic evidence demonstrated candidate regions and genes, which may be mediators in cardiac pathophysiology and potential therapeutic targets to improve prognosis in patients with HF.
A genome-wide association (GWA) study was performed in a total of 2,828 subjects of European ancestry with HF from seven samples collected within five large community-based prospective cohort studies including the Atherosclerosis Risk in Communities (ARIC and ARIC2) Study, the Cardiovascular Health Study (CHS), the Framingham Study (FHS), the Health ABC (Health ABC) study and the Rotterdam Study (RS and RS2). Sample characteristics, data collection and clinical definitions have been described previously and are summarized in S1 Text. [41–46] First diagnosis of heart failure (new-onset) was ascertained using a variety of methods based on international published criteria, as detailed in S1 Table. Mortality was ascertained from telephone contacts with relatives and from medical records, death certificates and/or municipal records (S1 Text).
Genotyping was performed using commercially available assays for genome-wide SNP detection. Imputation of non-genotyped SNPs was performed using CEU reference panels of SNP correlations from the HapMap project phase II (S1 Text), to characterize a total of 2.5 million SNPs. Imputation quality was assessed for each SNP from the ratio of observed over expected variance of allele dosage.
All-cause mortality following initial HF diagnosis was examined for association with additive allele dosage of each genotyped or imputed SNP using Cox proportional hazards models, with censoring at the end of or loss to follow-up. Models were adjusted for age at diagnosis, sex, and recruitment site in multicenter cohorts. In the family-based FHS, Cox models were implemented with clustering on pedigree to account for relatedness. Genomic control was applied to results from each cohort. Cohort-specific GWA results were combined using fixed effects meta-analysis with inverse variance weights. SNPs were excluded from cohort-level analyses if exhibiting implausible beta coefficients (< -5 or > 5) and from the meta-analysis for low heterozygosity (sample size-weighted minor allele frequency ≤ 0.03, corresponding to < 100 minor allele carriers with an endpoint).
SNPs passing a P-value threshold defined a priori as P < 5.0x10-7 in the genome-wide stage 1 were carried forward to the second stage with targeted genotyping in 1,870 HF patients from four independent cohorts. For 2.5 million tests, this threshold limits the expected number of genome-wide false positives to approximately 1, assuming statistical independence of tests. The second stage included four independent cohorts; the Malmö Diet and Cancer Study (MDCS), the Malmö Preventive Project (MPP), the Physicians’ Health Study (PHS) and the Prospective Study of Pravastatin in the Elderly at Risk (PROSPER) [47–50]. Heart failure ascertainment and time of death in these cohorts was similar to in stage 1 cohorts, as shown in S1 Table and S1 Text. Genotyping was performed as outlined in S1 Text. Association analyses and meta-analysis of results were performed as in the first stage. Meta-analysis of stage 1 and 2 was performed, and a combined P-value < 5.0x10-8 was considered significant. Heterogeneity was assessed across the combined stage 1 and 2 cohorts using Cochran’s Q test for heterogeneity, computed as the sum of the squared deviations of each study’s effect from the weighted mean over the study variance, and the I2 test, the percentage of total variation across studies that is due to heterogeneity rather than chance (I2 = [Q—df] / Q) [51, 52].
The association of replicated SNPs with measures of cardiac structure and function was evaluated from summary results of the following GWA consortia: EchoGen [19], CHARGE-HF [20], CHARGE-QRS [22], natriuretic peptides in 5453 subjects from the Malmö Diet and Cancer study [21], QT-IGC [23], and the CHARGE Sudden Cardiac Death consortium (manuscript in preparation). Each of these consortia is described in S1 Text.
The correlation of replicated SNPs with known coding SNPs was examined in the databases for the 1000 Genomes Project and phase III of the HapMap project, using SNAP [53]. The location of SNPs in relation to regulatory motifs was explored using histone methylation patterns generated as part of the ROADMAP Epigenomics project [24]. Enhancers were identified in each of the 129 ROADMAP tissues using the ChromHMM algorithm [54] from patterns of monomethylation (H3K4Me1) of the fourth residue (lysine) and acetylation of the 27th residue (H3K27Ac) of histone H3. The location of SNPs in relation to transcription factor binding sites was assessed in silico using HaploReg version 4.1 (http://www.broadinstitute.org/mammals/haploreg/haploreg.php) [55] and the UCSC Genome Browser (http://genome.ucsc.edu). In HaploReg, position weight matrices (PWMs; probabilistic representations of DNA sequence) were computed with p-values based on literature sources and ENCODE ChIP-Seq experiments as previously described [55], and only instances where a motif in the sequence passed a threshold of P < 4−7 were considered. The NHLH1-binding motif was retrieved into HaploReg from the manually curated TRANSFAC database.
Complementary DNA oligonucleotides corresponding to the 100 bp genomic region flanking rs9885413 (50 bp on either side of the SNP) were cloned into the luciferase reporter vector pGL3-Promoter (Promega, Madison, WI) using the MluI and BglII sites. Two different sets of oligos were cloned, one corresponding to the major allele of rs9885413 (pGL3P-G) and one to the minor allele (pGL3P-T). Oligonucleotide sequences were as following: major allele sense: CGCGTCCTGCCTCACATAATCTTTTTGTTTGTCCCCCTGAAATGGATTCTCAGCTGTTGCCCAAACATTTCATCTTAGCGTTCCAGGTTTGAACTCGCCCTCACGA, minor allele sense: CGCGTCCTGCCTCACATAATCTTTTTGTTTGTCCCCCTGAAATGTATTC TCAGCTGTTGCCCAAACATTTCATCTTAGCGTTCCAGGTTTGAACTCGCCCTCACGA, and the corresponding antisense sequences. The reporter vectors were co-transfected with the pRL-null vector at a ratio of 10:1 into HEK293 cells using Lipofectamine LTX (Life Technologies) according to the manufacturer’s instructions. 24 hours post-transfection, luciferase activity was assayed using the Dual-Luciferase Reporter Assay System (Promega) and Glomax 20/20 Luminometer (Promega). The signal from the reporter vector was normalized to the signal from the pRL-null vector.
Samples of left ventricular cardiac tissue from patients undergoing cardiac surgery were genotyped for the SNP rs9885413 and for all five transcripts within +/- 500 kb of the SNP. Samples of cardiac tissue were acquired from patients from the MAGNet consortium (http://www.med.upenn.edu/magnet/). Gene expression levels were determined using the Affymetrix ST1.1 gene expression array (Affymetrix, Santa Clara, CA, USA) in a cohort including 247 heart samples. Genotyping was performed using the Illumina OmniExpress array. Left ventricular free-wall tissue was harvested at time of cardiac surgery from subjects with heart failure undergoing transplantation or from unused transplant donors. In all cases, the heart was perfused with cold cardioplegia prior to cardiectomy to arrest contraction and prevent ischemic damage. Tissue specimens were then obtained and frozen in liquid nitrogen. Genomic DNA from left ventricle was extracted using the Gentra Puregene Tissue Kit (Qiagen) according to manufacturer’s instruction.
Total RNA was extracted from left ventricle using the miRNeasy Kit (Qiagen) including DNAse treatment on column. RNA concentration and quality was determined using the NanoVue Plus spectrophotometer (GE Healthcare) and the Agilent 2100 RNA Nano Chip (Agilent). For all samples, genome-wide SNP genotypes were generated using the Illumina OmniExpress Array. Caucasian Ancestry was verified using multi-dimensional scaling of genotypes. For Gene expression array experiments, the Affymetrix ST1.1 Gene array was used. Data were normalized using the Robust Multi-array Average algorithm and batch effects were adjusted for using ComBat. Transcript expression levels were considered significantly higher than background noise if expression values from robust multiarray analysis in at least 10% of either cases or controls exceeded of the 80% quantile of expression of genes on the Y-chromosome in female hearts (5.24). Associations of expression levels for expressed genes with SNP genotypes were tested using a likelihood ratio test. Specifically, we fit a linear regression model Y = β0 + β1*D + β2*g + β3*(g x D) where Y is the log2 transformed expression level of a given probe, g is the genotype (coded as 0, 1, and 2) of the test SNP, and D is heart failure disease status (D = 1 for heart failure cases and D = 0 for unused donor controls). Association between the probe and test SNP was assessed by testing H0: β2 = β3 = 0 using a likelihood ratio test. Significance of the test statistic was evaluated by comparing with a Chi-squared distribution with two degrees of freedom. All models were additionally adjusted for age, gender, and study site.
The association of the SNP rs9885413 with DNA methylation was examined in 2408 participants from the FHS Offspring cohort. Methylation at cytosine-guanine dinucleotides (CpG) at the 5q22 locus (+/-500 kb from rs9885413) were ascertained from a gene-centric DNA methylation array (Infinium HumaMethylation450 BeadChip, Illumina, San Diego, CA, USA) which allows interrogation of 485,512 methylation sites across the genome. The array has coverage of at least one methylation site near 99% of RefSeq genes and 96% of CpG islands. Briefly, bisulfite-treated genomic DNA (1 μg) from peripheral blood samples underwent whole-genome amplification, array hybridization and scanning according to manufacturer instructions. Genotyping of rs9885413 was performed as described in S1 Text. Association of rs9885413 and the methylation probe cg02061660 with expression of the five genes at the locus (+/-500 kb from rs9885413) was examined from microarray data (Affymetrix Human Exon Array ST 1.0) in 5257 participants from the FHS Offspring cohort and Third Generation cohort. Procedures for RNA extraction, processing and analysis have been described previously (28). Linear mixed effect (LME) models were fit accounting for familial correlation, cell count heterogeneity and technical covariates to account for batch effects using the pedigreemm package in R [56]. Specifically, the mQTL model utilized a two-step approach: first, the DNA methylation beta-value (ratio of methylated probe intensity to total probe intensity) was residualized with adjustment for age, sex, cell count proportions (imputed using the Houseman method for granulocytes, monocytes, B-lymphocytes, CD4+ T lymphocytes, CD8+ T lymphocytes and NK cells) [57], measured technical covariates (row, chip, column), and the family structure covariance matrix. Second, DNA methylation residuals were specified as dependent variable, SNP genotype dosage as independent variable with additional adjustment for 558 SVAs (surrogate variable analysis) [58] and ten principal components from eigenstrat [59] to account for unmeasured batch effects. The eQTL models similarly residualized gene expression with adjustment for age, sex, imputed cell count proportions (imputed in Offspring Cohort participants utilizing gene expression markers of cell count proportions developed from the Third Generation participants with both gene expression and measured complete blood counts), and family structure covariance matrix. The residual of gene expression was specified as dependent variable and SNP dosage as independent variable adjusted for 20 PEER (probabilistic estimation of expression residuals) factors [60] to account for unmeasured technical and batch effects in the gene expression data. The eQTM models specified gene expression residual as dependent variable and DNA methylation residual as independent variable adjusted for 20 methylation SVAs and 20 expression SVAs to account for unmeasured technical and batch effects.
Replication of the association of rs9885413 with cg02061660 including the same covariates in the model as in FHS was attempted in blood samples from 750 randomly selected participants of the Rotterdam study (RS3) not included in the GWA stage, where information from the same DNA methylation array as FHS was available. DNA was extracted, bisulfite-treated using the Zymo EZ-96 DNA-methylation kit (Zymo Research, Irvine, CA, USA) and hybridized to arrays according to manufacturer instructions. During quality control samples showing incomplete bisulfite treatment were excluded (n = 5) as were samples with a low detection rate (<99%) (n = 7), or gender swaps (n = 4). Probes with a detection P-value>0.01 in >1% samples, were filtered out. A total number of 474,528 probes passed the quality control and the filtered β values were normalized with DASEN implemented in the wateRmelon package in R statistical software. Genotyping was performed using the Illumina 610quad array. Cell counts were estimated using the same method as in FHS and also directly measured on a Coulter AcT Diff II Hematology Analyzer (Beckman Coulter, Brea, CA) for granulocytes, monocytes, lymphocytes). Models including both estimated and directly measured cell counts were explored.
HEK293 cells were seeded at 100,000 cells/well in a 6-well plate the day before transfection. Cells were transfected using Lipofectamine and 50 nM of siRNA designed to target human NHLH1 or negative control siRNA (Life Technologies, Carlsbad, CA, USA) according to the manufacturer’s instructions. After 48 hours, cells were harvested and total RNA extracted using the miRNeasy Mini Kit (Qiagen, Hilden, Germany) according to the manufacturer’s instructions. cDNA was synthesized using the RevertAid H- First Strand cDNA Synthesis Kit (Thermo Fischer Scientific, Waltham, MA, USA) using random hexamer primers and qPCR was performed with TaqMan assays for NHLH1, TMEM232, SLC25A4, WDR36, TSLP, CAMK4 and GAPDH on a StepOne Plus Real-Time PCR System (Life Technologies). Gene expression was normalized to GAPDH and expressed relative to cells transfected with negative control siRNA according to the ΔΔCt-method [61].
The frequencies of ancestral and derived alleles of rs9885413 were examined in populations from the International HapMap Project (http://www.hapmap.org/) [62] and the Human Genome Diversity Project (HGDP, http://hagsc.org/hgdp/) [63]. The fixation index (Fst) was estimated as described by Weir and Cockerham [64], based on allele frequencies in HapMap stage II as also previously described [65]. The integrated haplotype score (iHS) was calculated from HapMap stage II data as described by Voight et al (http://haplotter.uchicago.edu/) [33]. Allele frequency distributions in HGDP populations were visualized using the HGDP selection browser (http://hgdp.uchicago.edu/) [66].
Informed consent was obtained from all participants and all contributing studies were approved by the respective ethics committee as described in S1 Text.
|
10.1371/journal.pntd.0004282 | Characterization of the Paracoccidioides Hypoxia Response Reveals New Insights into Pathogenesis Mechanisms of This Important Human Pathogenic Fungus | Hypoxic microenvironments are generated during fungal infection. It has been described that to survive in the human host, fungi must also tolerate and overcome in vivo microenvironmental stress conditions including low oxygen tension; however nothing is known how Paracoccidioides species respond to hypoxia. The genus Paracoccidioides comprises human thermal dimorphic fungi and are causative agents of paracoccidioidomycosis (PCM), an important mycosis in Latin America.
In this work, a detailed hypoxia characterization was performed in Paracoccidioides. Using NanoUPLC-MSE proteomic approach, we obtained a total of 288 proteins differentially regulated in 12 and 24 h of hypoxia, providing a global view of metabolic changes during this stress. In addition, a functional characterization of the homologue to the most important molecule involved in hypoxia responses in other fungi, the SREBP (sterol regulatory element binding protein) was performed. We observed that Paracoccidioides species have a functional homologue of SREBP, named here as SrbA, detected by using a heterologous genetic approach in the srbA null mutant in Aspergillus fumigatus. Paracoccidioides srbA (PbsrbA), in addition to involvement in hypoxia, is probable involved in iron adaptation and azole drug resistance responses.
In this study, the hypoxia was characterized in Paracoccidioides. The first results can be important for a better understanding of the fungal adaptation to the host and improve the arsenal of molecules for the development of alternative treatment options in future, since molecules related to fungal adaptation to low oxygen levels are important to virulence and pathogenesis in human pathogenic fungi.
| The genus Paracoccidioides is composed of species that are causative agents of paracoccidioidomycosis (PCM), a neglected human granulomatous mycosis, endemic in Latin America. To survive in the human host, fungi must tolerate and overcome in vivo micro environmental stress conditions, including low oxygen levels. Paracoccidioides spp. depicts differential responses to several stresses such as iron/zinc deprivation, oxidative and nitrosative stresses and carbon starvation. In addition, Paracoccidioides yeast cells recovered from liver of infected mice demonstrated adaptability to the host conditions. Mechanisms by which fungi sense oxygen levels have been characterized, although this is the first description in Paracoccidioides spp. Little is known about hypoxia in thermally dimorphic fungi and nothing has been studied in Paracoccidioides genus, one of the representatives of this group of pathogens. A detailed characterization of the hypoxia responses was performed using proteomic and heterologous genetics approaches. Paracoccidioides genus have a functional homologue of the key regulator of hypoxia adaptation in fungi, SrbA, a SREBP (sterol regulatory element binding protein) orthologue. The proteome during hypoxia provided a global view of metabolic changes during this stress and species of the Paracoccidioides genus have a functional SrbA. Our study provides a better understanding of the fungal adaptation to the host and it can improve the arsenal of molecules for the development of alternative treatment options to paracoccidioidomycosis, since molecules related to fungal adaptation to low oxygen levels are important to virulence and pathogenesis in human pathogenic fungi.
| The genus Paracoccidioides is a complex of thermodimorphic fungi, and are causative agents of paracoccidioidomycosis (PCM) a deep systemic granulomatous mycosis, endemic in Latin America [1, 2]. Paracoccidioides spp. grows as yeast in host tissue and in vitro at 36°C, and as mycelium under saprobiotic and laboratory conditions (18–23°C). As the dimorphism is dependent on temperature, when the mycelia or conidia are inhaled into the host respiratory tract, the transition to the pathogenic yeast phase occurs [3]. Once in the lungs, epithelial cells and resident macrophages are the first line of defence against Paracoccidioides cells. Inside macrophages, the parasitic yeast form subverts the normally harsh intraphagosomal environment and proliferates [4]. Adhesion to and invasion of epithelial cells and basal lamina proteins may be required for the extra pulmonary haematogenous fungal dissemination to organs and tissues [1, 3, 5].
To survive in the human host, fungi must also tolerate and overcome in vivo micro environmental stress conditions. Conditions such as high temperature, distinct ambient pHs, carbon and metal ions deprivation, and gas tension (high levels of carbon dioxide and low oxygen levels) induce several stress responses in the invading fungus [6–10]. In Paracoccidioides spp., previous analyses have demonstrated differential responses to iron and zinc deprivation, oxidative and nitrosative stress and carbon starvation faced by the fungus during infection [11–15]. In addition, Paracoccidioides spp. yeast cells recovered from liver of infected mice and from infected macrophages alter their metabolism in order to adapt to the host using available nutrition sources [16, 17].
It is well established that oxygen levels vary throughout the mammalian body depending on numerous factors including tissue type and presence or absence of an inflammatory response [18]. Oxygen levels in most mammalian tissues are found to be considerably below atmospheric levels (21%) [19, 20]. Also, oxygen availability at the sites of inflammation is significantly reduced compared to surrounding tissues [21, 22] since, in inflamed tissues, the blood supply is often interrupted because the vessels are congested with phagocytes or the pathogen itself [23, 24]. Thus, it seems highly probable that hypoxic microenvironments are generated during fungal infection [25, 26].
Mechanisms used by fungi to sense oxygen levels have been characterized [27]. An SREBP (sterol regulatory element binding protein) ortholog, previously characterized in higher eukaryotes [28–32], was first identified and characterized in the fission yeast Schizosaccharomyces pombe as an oxygen sensor [33, 34]. Later, it was characterized in the human pathogenic fungi Cryptococcus neoformans and Aspergillus fumigatus [35–37]. In A. fumigatus, the SREBP homologue, SrbA, controls the expression of genes involved in biosynthesis of lipids, ergosterol, and heme [37, 38]. Recently, a new transcriptional regulator of the fungal hypoxia response and virulence that genetically interacts with SrbA, named SrbB, was also characterized in A. fumigatus [39]. In S. pombe and C. neoformans the SREBP homologues also regulate enzymes in the ergosterol biosynthetic pathway under hypoxic conditions [34, 35, 38].
Oxygen levels are low in subsurface layers of organic matter in natural environments that are habitats of environmental pathogens such Paracoccidioides and Aspergillus [40–43]. In this context, studies regarding the responses of Paracoccidioides to hypoxia are of relevance and in this study are described for the first time. Up to now, hypoxia has not been described in the Paracoccidioides genus, representatives of thermally dimorphic fungi, in which responses to hypoxia remain to be investigated. We observed that Paracoccidioides yeast cells respond to hypoxia regulating the expression of proteins from diverse metabolic pathways. We also observe that species of the Paracoccidioides genus have homologues of the key regulator of hypoxia adaptation in fungi, SrbA. Paracoccidioides srbA was characterized using a heterologous genetics approach that confirmed the functional conservation of this protein in the hypoxia response. Paracoccidioides srbA (PbsrbA) is likely involved in hypoxia, iron adaptation and azole drug resistance responses, as observed by functional complementation of the srbA null mutant in A. fumigatus by PbsrbA. The obtained data, may improve the arsenal of molecules for the development of alternative treatment options since molecules related to fungal adaptation to low oxygen levels are important to virulence and pathogenesis in human pathogenic fungi.
Paracoccidioides, Pb01 (ATCC MYA-826), was used in the experiments. The yeast phase was cultivated for 7 days, at 36°C in BHI semisolid medium added of 4% (w/v) glucose. When required, the cells were grown for 72 h at 36°C in liquid BHI, washed with PBS 1X, and incubated at 36°C in McVeigh/Morton (MMcM) medium as previously described [44]. Pb01 yeast cells were subjected to normoxia and hypoxia as previously described [37, 45]. Normoxia was considered general atmospheric levels within the lab (~21% O2). For hypoxia, an incubator (Multi-Gas Incubator MCO-19M-UV, Panasonic Biomedical) was used. The chamber was maintained at 36°C, and kept at 1% oxygen level, utilizing a gas mixture containing 1% O2, 5% CO2 and 94% N2. Paracoccidioides yeast cells viability was determined as previously described: the number of viable cells was determined at times of 0, 6, 12, 18 and 24 h by staining with 0.01% (w/v) trypan blue in PBS1X [12, 46, 47].
All A. fumigatus strains were routinely grown in glucose minimal medium (GMM) with appropriate supplements at 37°C as previously described [45, 48]. To prepare solid media, 1.5% (w/v) agar was added, before autoclaving. For protein extraction and associated mRNA abundance experiments, 0.5% (w/v) yeast extract was added to liquid GMM to increase hypha mass [45]. For hypoxia cultivations, an incubation chamber (Invivo2 400; Ruskinn) was used. The chamber was maintained at 37°C and kept at 1% O2, 5% CO2, and 94% N2, controlled through a gas mixer (Gas Mixer Q; Ruskinn/Baker Company). Normoxia was also considered general atmospheric levels within the lab (~21% O2).
Following Paracoccidioides yeast cells incubation under normoxia and hypoxia, in biological triplicates, cells were centrifuged at 1,500 x g, resuspended in 50 mM ammonium bicarbonate pH 8.5 and disrupted using glass beads and bead beater apparatus (BioSpec, Oklahoma, USA) in 5 cycles of 30 sec, while on ice. The cell lysate was centrifuged at 10,000 x g for 15 min at 4°C and the supernatants for each condition were polled in equimolar amounts and subjected to the nanoscale liquid chromatography coupled with tandem mass spectrometry in 3 technical replicates. The proteins were quantified using the Bradford reagent (Sigma-Aldrich) [49]. Sample aliquots (70 μg) were prepared for NanoUPLC-MSE as previously described [11, 15, 50, 51], with some modifications. Briefly, 50 mM ammonium bicarbonate was added and was followed by addition of 35 μL of RapiGEST (0.2%v/v) (Waters Corp, Milford, MA). The solution was vortexed and then incubated at 80°C for 15 min. The disulphide bonds were reduced by treating proteins with 10mM D-L-dithiothreitol for 30 min at 60°C. The sample was cooled at room temperature and the proteins were alkylated with 200 mM iodoacetamide in a dark room for 30 min. Proteins were digested with trypsin (Promega, Madison,WI, USA, 1:25 w/v) prepared in 50 mM ammonium bicarbonate, at 37°C overnight. Following the digestion, 10 μL of 5% (v/v) trifluoroacetic acid was added to hydrolyse the RapiGEST, followed by incubation at 37°C for 90 min. The sample was centrifuged at 18,000 x g at 6°C for 30 min, and the supernatant was transferred to a Waters Total Recovery vial (Waters Corp). A solution of one pmol.ul-1 MassPREP Digestion Standard [rabbit phosphorylase B (PHB)] (Waters Corp) was used to prepare the final concentration of 150 fmol.ul-1 of the PHB. The buffer solution of 20 mM ammonium formate (AF) was used to increase the pH. The digested peptides were separated further via NanoUPLC-MSE and analysed using a nanoACQUITY system (Waters Corporation, Manchester, UK). Mass spectrometry data obtained from NanoUPLC-MSE were processed and searched against the Paracoccidioides database (http://www.broadinstitute.org/annotation/genome/paracoccidioides_brasiliensis/MultiHome.html) using ProteinLynx Global Server (PLGS) version 2.4 (Waters Corp). Protein identifications and quantitative data packaging were performed using dedicated algorithms [52, 53]. The ion detection, clustering, and log-scale parametric normalizations were performed in PLGS with an ExpressionE license installed (Waters, Manchester, UK). The false positive rate (FPR) of the algorithm for protein identification was set to 4% in at least two out of three technical replicate injections. Using protein identification replication as a filter, the false positive rate was minimized because false positive protein identifications, i.e., chemical noise, have a random nature and do not tend to replicate across injections. For the analysis of the protein identification and quantification level, the observed intensity measurements were normalized to the intensity measurement of the identified peptides of the digested internal standard. Normalization was performed with a protein that showed no significant difference in abundance in all injections [54] to accurately compare the expression protein level to normoxia and hypoxia samples. For 12 and 24 h, the proteins oxidoreductase 2-nitropropane dioxygenase and 40S ribosomal protein S5 were used as normalizing proteins, respectively (PAAG_01321 and PAAG_05484 from Paracoccidioides genome database http://www.broadinstitute.org/annotation/genome/paracoccidioides_brasiliensis/MultiHome.html). Furthermore, only those proteins with a fold change higher than 50% difference were considered to be expressed at significantly induced/ repressed levels.
Paracoccidioides, Pb01 yeast cells, were grown under normoxia and hypoxia for 12 and 24 h, in biological triplicates. Following that, cells were harvested by centrifugation at 2,000 x g for 5 min at 4°C and diluted in PBS buffer at 106 cells/ml. Cells were stained with Rhodamine 123 (1.2 mM) (Sigma Aldrich) according to the manufacturer's protocol and then washed twice with 1X PBS. Stained cells were observed under a fluorescence microscope (AxioScope A1, Carl Zeiss) and analysed with the 546–512 nm filter. Rhodamine fluorescence intensity was measured using the AxioVision Software (Carl Zeiss). The minimum of 100 cells for each microscope slides, in triplicates, for cells submitted to hypoxia and normoxia for 12 and 24 h were used to measure the rhodamine fluorescence intensity. The software provided the fluorescence intensity (in pixels) and the standard deviation of each analysis. Statistical comparisons were performed using the student’s t test and p-values ≤ 0.05 were considered statistically significant.
The amino acid predicted sequences were obtained from GenBank (http://www.ncbi.nlm.nih.gov/) to Paracoccidioides Pb01 (XP_002794199); Pb03 (KGY15961); Pb18 (EEH47197); Aspergillus fumigatus (XP_749262); Schizosaccharomyces pombe (NP_595694); Cryptococcus neoformans (XP_567526) and Homo sapiens (P36956). The SMART tool (http://smart.embl-heidelberg.de) [55, 56] was used to search for conserved domain bHLH (basic helix-loop- helix leucine zipper DNA-binding domain) and Phobius (http://phobius.sbc.su.se/) [57] and SACS MEMSAT2 Prediction software (http://www.sacs.ucsf.edu/cgi-bin/memsat.py) [58] were used to depict transmembrane segments. The amino acid sequences from all proteins were aligned using CLUSTALX2 [59] to show a conserved tyrosine residue (indicated by asterisk) specific to the SREBP family of bHLH transcription factors.
Following Paracoccidioides incubation under hypoxia and normoxia, cells were harvested, and total RNA was extracted using TRIzol (TRI Reagent, Sigma-Aldrich, St. Louis, MO, USA) and mechanical cell rupture (Mini-Beadbeater-Biospec Products Inc., Bartlesville, OK). Total RNA was extracted and treated with DNase (RQ1 RNase-free DNase, Promega). After in vitro reverse transcription (SuperScript III First-Strand Synthesis SuperMix; Invitrogen, Life Technologies), the cDNAs were submitted to a qRT-PCR reaction, which was performed using SYBR Green PCR Master Mix (Applied Biosystems, Foster City, CA) in a StepOnePlus Real-Time PCR System (Applied Biosystems Inc.). The expression values were calculated using the transcript that encoded alpha tubulin (GenBank accession number XP_002796639) as the endogenous control, as previously reported [11] and, when required, the data were presented as relative expression in comparison to the experimental control cells value set at 1. Relative expression levels of genes of interest were calculated using the standard curve method for relative quantification [60]. Briefly, for each of the three replicates of a sample, the average quantity (avg) was calculated of target cDNA interpolated from the standard curve, the standard deviation of the average (stdev), and the coefficient of variation (CV) according to the formula CV = stdev/ avg. Any outlier points (>17% CV) was removed and avg, stdev and CV were recalculated. For each sample, the gene of interest (GOI) was normalized to the reference gene (RG) for the sample according to the following equation: normalized value = avg GOI quantity/ avg RG quantity. The standard deviation (SD) of the normalized value was calculated according to the equation: SD = (normalized value) x square root (CV reference gene + CV gene of interest)2. The resulting values were plotted as a bar graph of normalized value versus sample name or experimental treatment group, with the error bars equal to the SD, of the biological triplicates of independent experiments [60]. Standard curves were generated by diluting the cDNA solution 1:5. Statistical comparisons were performed using the student’s t test and p-values ≤ 0.01 were considered statistically significant.
Regarding to A. fumigatus, wild type, ΔsrbA and reconstituted strains were cultured in liquid GMM under normoxia or hypoxia. Germlings and mycelia were collected with vacuum filtration and lyophilized, prior to homogenization with 0.1-mm glass beads. Total RNA was extracted, treated with DNase, reversed transcripted to cDNA and submitted to a qRT-PCR reaction, identically to which was performed to Paracoccidioides. Oligonucleotides to amplify the srbA gene from A. fumigatus and Paracoccidioides were used in the experiments. The data were normalized using the A. fumigatus tefA reference gene [61]. Primers are depicted in S3 Table.
The A. fumigatus strains CEA10 (wild type) and a srbA null mutant of A. fumigatus were used in the genetic complementation assays. This srbA null mutant was previously generated by replacement of the srbA coding sequence in A. fumigatus strain CEA17 with the auxotrophic marker pyrG from A. parasiticus as previously described [45, 62, 63]. To perform genetic complementation of the respective ΔsrbA, the Paracoccidioides Pb01 srbA sequence was amplified from Paracoccidioides genomic DNA as template and linked together with a fragment of the gpdA (glyceraldehyde phosphate dehydrogenase) gene from Aspergillus nidulans, used as promoter and a functional pyrG gene from A. parasiticus, used to select the transformed strains (gpdA+PbsrbA+pyrG). The fused product was used to perform fungal transformations. Generation of fungal protoplasts and polyethylene glycol-mediated transformation of A. fumigatus were performed as previously described [45, 64]. Reconstituted strains were confirmed by screening using hypoxia chamber, conventional PCRs, Southern blots, qRT-PCRs and immunoblot analyses. All primers used are shown in S3 Table. In order to eliminate the chance of heterokaryons, each transformant was streaked with sterile toothpicks a minimum of twice, to obtain colonies from single conidia. All strains were stored as frozen stocks with 50% (v/v) glycerol at -80°C.
Ten-well 10% Mini-Protean precast gel (Bio-Rad) was used for SDS-PAGE. Denatured protein was loaded (40 μg per well). After gel electrophoresis, protein was transferred to a nitrocellulose membrane (Hybond-C Extra; Amersham Biosciences). PbSrbA was detected on blots using the A. fumigatus SrbA 1–275 recombinant primary N-terminus antibody at a 1/27,000 dilution and an anti-rabbit alkaline phosphatase (AP)-conjugated secondary antibody raised in goat (Abcam) at a 1/5,000 dilution, as previously described [45]. Chemiluminescence was measured following incubation of blots with Tropix CPD Star substrate (Applied Biosystems) with Immun-star enhancer (Bio-Rad) using a FluorChem FC2 imager (Alpha Innotech).
DNA was isolated from overnight liquid cultures of A. fumigatus. The mycelium was separated from the medium by filtration and glass beads were used to disruption. Additional purification steps were used to isolate the genomic DNA and Southern blot was performed using the digoxigenin labelling system (Roche Molecular Biochemicals, Mannheim, Germany) as previously described [45, 65]. Briefly, 30 μg aliquots of genomic DNA were digested with HindIII and EcoRI to detect gpdA and pyrG, respectively. Restriction digests were separated on a 1% agarose gel and blotted onto nylon membranes. The concentration of the probes in hybridization solution was 50 ng/ml, and hybridization was carried out at 50°C. Membranes were washed in a final solution of 0.1 SSC and 0.1% (w/v) sodium dodecyl sulphate, at 68°C.
Production of biomass was performed to wild type, ΔsrbA and reconstituted strain 1 (Rec 1) of A. fumigatus. A total of 108 cells of each strain were grown under iron starvation (−Fe) and iron sufficiency (0.03 mM, +Fe) in liquid medium for 24 h, at 37°C. The cells were harvested by vacuum filtration and then lyophilized. The data represent the mean ± SD of biological triplicates and the values were normalized to the reconstituted strain. Statistical comparisons were performed using the student’s t test and p-values ≤ 0.01 were considered statistically significant.
P. brasiliensis yeast cells were grown in McVeigh/Morton medium (MMcM) [44] and the yeast cells were incubated at 36°C with shaking at 150 rpm. In order to analyse the kinetic of expression of PbsrbA, the cells were cultivated under iron deprivation or supplementation, using the iron chelator bathophenanthroline disulfonate (BPS; 50 μM; Sigma-Aldrich, Germany) or 3.5 μM Fe(NH4)2(SO4)2, respectively. Total RNA was extracted at 30 min, 1, 3 and 24 h and the quantitative real time PCR was performed as cited above.
In order to start the characterization of Paracoccidioides hypoxia response, we utilized a proteomics approach. The NanoUPLC-MSE [50, 51] was previously used to map metabolic changes in Paracoccidioides at a protein level [11, 13, 15, 17] and was also used in this study. After exposing the cells to normoxia (21% pO2) and hypoxia (1% pO2) and using a proteomics approach at time points 12 and 24 h, we observed significant differences in protein expression indicating that the fungus responds to hypoxia.
As described in Lima and co-workers [15], a 1.5-fold change was used as a threshold to determine positively and negatively differentially proteins. In total, 134 and 154 proteins presented different abundances in 12 and 24 h under hypoxia, respectively, compared to normoxia. In 12 h, the same number of proteins (67) were increased and decreased, upon hypoxia, compared to control (normoxia). At 24 h, 102 proteins were increased and 52 were decreased (S1 and S2 Tables). The adaptation mechanism of Paracoccidioides to hypoxia, as represented by biological processes, as deduced from increased and decreased proteins is shown in Fig 1. Proteins associated with several subcategories of metabolism were represented in both analyses as increased and decreased proteins. Some of them were less represented in 24 h of hypoxia such as nitrogen, purine nucleotide/ nucleoside/ nucleobase and phosphate metabolism. Proteins associated with energy depicted an interesting profile of abundance. Those involved with electron transport/ membrane associated energy conservation were enriched for reduced levels in 12 h of hypoxia, and levels were subsequently restored at 24 h of hypoxia (Fig 1, S1 and S2 Tables).
To further assess this observation, we evaluated mitochondrial activity using rhodamine, a permeable lipophilic cationic fluorescent probe that accumulates in mitochondria and is distributed electrophoretically into the mitochondrial matrix in response to mitochondrial electric potential [66, 67]. The rhodamine probe has been used to stain yeast cells [67], including Paracoccidioides [13, 68]. Consistent with the proteomics data that suggested reduced mitochondrial activity, a low level of staining of rhodamine was observed in yeast cells during 12 h of hypoxia. At 24 h, the intensity of detection was restored, which is consistent with proteomics data (Fig 2). Additionally, proteins such as catalase, thioredoxin, chaperones and gamma-glutamyltranspeptidase were up-regulated in Paracoccidioides in hypoxia for 12 h (S1 Table) and could be associated with the altered mitochondrial activity. Our suggestion is that the fungus possibly induces ROS scavenging enzymes to protect the fungus against low oxygen effects that induces a strong reduction in electron-transfer reactions. In C. neoformans, several genes associated with the mitochondrial activity were identified as essential for hypoxic growth [69].
Proteins in the energy subcategories glycolysis/ gluconeogenesis, TCA cycle and GABA shunt were also differentially abundant under hypoxia. The glycolysis/ gluconeogenesis and GABA shunt were increased at 24 h of hypoxia. On the other hand, proteins of the TCA cycle were reduced at both time points (Fig 1, S1 and S2 Tables). The mechanisms of hypoxia adaptation are variable among fungi [18, 70]. At transcript level, for example, genes involved with glycolysis were induced, while those involved with aerobic respiration were repressed in Candida albicans, a facultative anaerobe, submitted to hypoxia [71–73]. However, in the obligate aerobic yeast C. neoformans, a general lack of changes in glycolytic mRNA abundance was observed in response to hypoxia, and genes involved in mitochondrial function have been observed to be critical for the hypoxia response [36, 74]. In the obligate aerobic mold A. nidulans, exposure to hypoxia results in an increase in glycolytic gene transcripts and the GABA shunt, which bypasses two steps of the tricarboxylic acid (TCA) cycle [75]. Transcriptome data from A. nidulans largely correlated with the proteomic profile, in which proteins in core metabolism and utilization of the GABA shunt was identified [76]. Similar results were found in A. fumigatus, upon short-term hypoxia as the GABA shunt was also induced [77]. On the other hand, cultures exposed to long-term hypoxia revealed increased abundance of proteins involved in glycolysis, respiration, pentose phosphate pathway, and amino acid and pyruvate metabolism [78].
Fig 3 depicts probable mechanisms used by Paracoccidioides to overcome hypoxic environments. It does not represent an integral model of how Paracoccidioides adapts to hypoxia, but from our point of view is an important source to start the understanding of how this fungus adapts to low oxygen levels. The abundance of some enzymes involved in acetyl-CoA production are up-regulated in 12 h of hypoxia compared to normoxia. The induction, for example, of the aldehyde dehydrogenase and long-chain specific acyl-CoA dehydrogenase enzymes suggest that the acetyl-CoA is produced via acetaldehyde and beta-oxidation pathway, respectively. Consistent with these data, proteins involved in glycolysis were decreased in abundance (S2 Table). Acetyl-CoA can be used as an alternative carbon source under these conditions (Fig 3). In fact, the expression of proteins related to glycolysis, acetyl-CoA production from pyruvate and citrate, TCA cycle and oxidative phosphorylation were reduced (Fig 3, S2 Table).
At 24 h, the detected up- and down-regulated proteins could show additional changes in Paracoccidioides strategies to adapt to hypoxia. For example, proteins involved in glycolysis are now increased supporting pyruvate production. In addition, the GABA shunt is increased at 24 h of hypoxia (Fig 3, S1 Table). The increased abundance of two enzymes involved with the GABA shunt pathway, NADP specific glutamate dehydrogenase and succinate-semialdehyde dehydrogenase, support this hypothesis in Paracoccidioides. Reports have shown that GABA is generated from 2-oxoglutarate via glutamate through the actions of glutamate dehydrogenase and glutamate decarboxylase, and that GABA transaminase irreversibly transaminates GABA to succinic semialdehyde, which is then oxidized to succinate by succinic semialdehyde dehydrogenase [76, 79, 80]. Transcripts for this pathway are also up-regulated in A. nidulans and A. fumigatus, under hypoxia [75, 77]. The GABA shunt is hypothesized to help organisms to avoid accumulation of high NADH levels in the absence of a terminal electron acceptor such as oxygen, and also contributes to glutamate formation [77]. This pathway is also described as an alternative route to the TCA cycle [75]. Interesting, the TCA pathway was down-regulated, based on protein levels of key enzymes (Fig 3, S2 Table), although the role of the GABA shunt in the fungal hypoxia response remains to be conclusively determined.
Moreover, enzymes involved in beta-oxidation and in production of ergosterol precursor molecules were also up-regulated according to proteomic data, at 24 h (Fig 3, S1 Table). During Paracoccidioides hypoxia adaptation, the detection of the long-chain specific acyl-CoA dehydrogenase, for example, shows that the fungus activates the beta-oxidation resulting in acetyl-CoA, that could be involved in fatty acid and ergosterol production. The enzyme 3-hydroxybutyryl-CoA dehydrogenase yields 3-acetoacetyl-CoA that together to acetyl-CoA supports ergosterol synthesis. Our suggestion makes sense since acetyl-CoA is probably not produced by pyruvate, neither acetate nor citrate, since enzymes related to their metabolism are down regulated in our data (Fig 3, S2 Table). The relative expression level of the transcript encoding Pberg3 was determined by quantitative real time PCR (Fig 4). The gene Pberg3 encodes C-5 sterol desaturase, an enzyme involved in the late steps in sterol biosynthesis [74, 81]. The data provide additional evidence that Paracoccidioides faces hypoxia and regulates ergosterol production, to compensate the effects of low oxygen levels. Several enzymatic steps in ergosterol biosynthesis are catalysed by iron and oxygen-requiring enzymes including that performed by Erg3 [74]. Also, the metabolism of fatty acids and ergosterol are increased in C. albicans, C. neoformans, A. fumigatus and A. nidulans in response to hypoxia and these molecules are required for the stability, fluidity and structure of the fungus plasma membrane [36, 72–74, 76, 77]. On this way, the fungus might be remodelling the fatty acid content of membrane lipids to keep the membrane fluidity in hypoxia. Along with ergosterol’s role as a target to antifungal drugs, the understanding of the mechanisms that regulate ergosterol biosynthesis is of interest to biomedical research [82, 83]. In S. pombe, A. fumigatus and C. neoformans, the SREBP proteins are effectors which sense changes in oxygen levels indirectly through alterations in ergosterol levels [33, 35, 37]. Therefore, we addressed the question whether Paracoccidioides also relied on an SREBP like protein to adapt to hypoxia.
We hypothesized that Paracoccidioides hypoxia response could be, in part, regulated by a homologue of the SREBPs, an ancient family of regulators, associated with the hypoxic response in fungi [27, 33–35, 37, 84]. In silico analysis using Genbank (http://www.ncbi.nlm.nih.gov/) and Paracoccidioides genome databases (http://www.broadinstitute.org/annotation/genome/paracoccidioides_brasiliensis/MultiHome.html) showed that members of the genus Paracoccidioides, including the isolate 01, contain homologues of SREBPs. We named the gene srbA (PbsrbA), and the accession numbers in the Paracoccidioides genome database are PAAG_03792, PADG_03295 and PABG_11212 for Pb01, Pb18 and Pb03 strains, respectively.
The SREBP proteins are basic helix-loop-helix leucine zipper transcription factors with a conserved tyrosine residue, specific to this family. In addition, the SREBP present transmembrane domains, responsible for associating the protein with endoplasmic reticulum (ER). The Paracoccidioides spp. srbA genes contain those domains (Fig 5 and S1 Text) suggesting that they are an integral membrane protein which requires to be processed to release the N-terminus containing the bHLH DNA binding domain.
In mammals, SREBPs are synthesized as inactive precursors on the endoplasmic reticulum (ER) membrane where they bind to the SREBP cleavage activating protein (SCAP) which mediates sterol-dependent regulation of SREBP activity. The SCAP protein interacts with another ER-resident protein, named INSIG, and other proteases that cleave into the first transmembrane segment, to release the N-terminal transcription factor SREBP, which translocates to the nucleus and regulates expression of genes required when cholesterol levels are low [27, 28, 38, 85]. In fungi, some differences are detected in the SREBP processing illustrating that, while many aspects of SREBP regulation are conserved across organisms, others are not [45]. In general, the differences are involved with the SREBPs processing for their activation. In S. pombe and C. neoformans, SREBPs are regulated in part by proteolysis, although in S. pombe, this processing is dependent on a Golgi E3 ligase complex, encoded by dsc (defective for SREBP cleavage) genes and not homologues of human proteases, as found in C. neoformans [86–88]. In A. fumigatus, the processing is similar to that found in S. pombe involving the Dsc complex, required for cleavage of SrbA. The hypoxic adaptation and virulence of A. fumigatus require both, SREBP and its processing mechanism, demonstrating an important mechanism to fungal pathogenesis [37, 45]. Paracoccidioides spp., in contrast to S. pombe and in accordance with A. fumigatus, does not depict in the genome database homologues for SCAP protein. On the other hand, there is an apparent homolog to the INSIG protein (Table 1). Moreover, the Site-1 and Site-2 proteases homologues were not identified in Paracoccidioides spp. genomes, as found in S. pombe and A. fumigatus (Table 1). These findings reinforce the relevance of studying activation of SrbA in Paracoccidioides spp.
To determine if PbsrbA responds to hypoxia, we first examined mRNA levels of the transcript in different oxygen conditions. The fungus significantly increases the levels of PbsrbA after 1 h upon hypoxia exposure in comparison to normoxia (Fig 6). These results suggest that PbsrbA may be involved in the hypoxia response in Paracoccidioides spp. and further analyses were performed to test this hypothesis.
There are a reduced number of works relating the functional analysis of genes in Paracoccidioides in the last six years because the achievement of viable and stable mutants of Paracoccidioides spp. is a hard task [11, 13, 89–94]. This parsimony in functional analysis surely reflects the complexity of those studies in the genus Paracoccidioides, as well as in other pathogenic fungi. Due these limitations in molecular genetic analyses available in Paracoccidioides, we utilized a heterologous genetics approach to test our hypothesis.
In order to test whether Paracoccidioides srbA was able to replace the A. fumigatus SrbA function, we introduced Paracoccidioides srbA (PbsrbA) under control of the gpdA (glyceraldehyde-3-phosphate dehydrogenase) promoter from A. nidulans into a previously characterized srbA null mutant strain A. fumigatus (ΔsrbA) [45]. Ectopic introduction of the Paracoccidioides srbA gene (PbsrbA) into ΔsrbA allowed us to attribute all resulting phenotypes specifically to the absence of srbA in A. fumigatus [37, 45]. Colonies were exposed to low oxygen growth condition (1% pO2) to randomized screening (S1 Fig) and confirmation of the strain genotype was done with Southern blot and PCR analyses (S2 Fig). A total of one and two copies of the PbsrbA and pyrG gene was observed in Rec1 (reconstituted strain 1) and Rec2 (reconstituted strain 2), respectively. The detected high band on pyrG Southern blot results (around 5 kb) is an unspecific cross-reactive detection because the probe is able to recognize the non-functional pyrG used to knockout the srbA gene in A. fumigatus genome [45] (S2A Fig).
We next confirmed the PbsrbA genome integration using conventional PCR, using primers that amplify the PbsrbA sequence including the AngpdA promoter (S2B Fig). In addition, the PbsrbA transcript and protein expression were assessed (S3 Fig). As expected, the PbsrbA transcript was expressed only in the reconstituted strains (Rec1 and Rec2), increasing when the fungus was submitted to hypoxia (S3A Fig). In agreement, the AfsrbA transcript was not detected in the reconstituted strains. The transcript to AfsrbA was also analyzed and the results are consistent with previously published data and reinforce the obtained data with PbsrbA. Using quantitative real time PCR, we observed that the transcript to AfsrbA was expressed only in the wild type strain, increasing when the fungus faced hypoxia (S3A Fig). In addition, at the protein level, the western blotting analysis, using a polyclonal antibody against A. fumigatus SrbA amino acids 1–275, indicates that PbsrbA is expressed in the reconstituted strain (Rec1) (S3B Fig). The A. fumigatus SrbA protein was also detected in the wild-type strain showing the SrbA precursor and N-terminal cleavage protein [45].
In order to analyse the growth of the reconstituted strains exposed to hypoxia, we measured the colony diameter of each strain every 24 h (Fig 7). As previously described, the A. fumigatus srbA null mutant strain does not growth under hypoxia [37]. However, the PbsrbA reconstituted strains were able to restore the null mutant hyphal growth under hypoxia (Fig 7). This result indicates that Paracoccidioides has a functional SrbA protein that can rapidly promote adaptation to hypoxic microenvironments.
Previous studies showed that the A. fumigatus SrbA protein coordinates iron and ergosterol homeostasis to mediate triazole drug and hypoxia responses [37, 95]. The A. fumigatus SREBP is a key positive regulator of iron homeostasis, particularly related to iron acquisition, which is essential for adaptation to hypoxia and low iron microenvironments [95]. Iron homeostasis has been characterized in Paracoccidioides [11, 12, 96–98] and the elucidation of additional molecules involved in this process can be relevant in the understanding of fungus pathogenesis. In this sense, our purpose was firstly attempted to screen PbsrbA reconstituted strain susceptibility to antifungals drugs using ranges of azoles concentrations [37] (Fig 8). The results showed that PbsrbA restores the failed growth of the mutant and suggest its participation in mechanisms of resistance to azoles. Previous studies in S. pombe, C. neoformans, and A. fumigatus confirmed that fungal SREBPs are key regulators of ergosterol biosynthesis [33, 36, 39, 77]. In A. fumigatus, the SrbA protein is involved, even in part, in regulation of the expression of several ergosterol biosynthesis genes [37, 95]. Taken together, the results suggest that PbsrbA can also be involved in these mechanisms because transcript to Pberg3 involved in ergosterol biosynthesis production, is also regulated in Paracoccidioides Pb01 submitted to hypoxia (Fig 4). Possibly, the fungus increases the expression of genes related to ergosterol biosynthesis, in order to compensate the reduction in ergosterol production in low oxygen, as discussed before in Fig 3.
Regarding iron homeostasis, previous studies showed that the initial responses to hypoxia in A. fumigatus involve transcriptional induction of genes involved in iron acquisition. The null mutant strain to srbA (ΔsrbA) has reduced growth under iron starvation in liquid medium because it coordinates responses to iron and oxygen depletion [39, 95]. Here the PbsrbA reconstituted strain 1 (Rec1) was significantly able to restore the defective growth phenotype of the mutant (Fig 9A). In fact, the transcript to PbsrbA is up-regulated in Paracoccidioides sp. grown upon iron deprivation, mainly after 24 h of incubation (Fig 9B). Even in part, this gene could be important in the mechanisms to compensate the effect of iron depletion in Paracoccidioides sp. yeast cells.
Altogether, the results show that the roles of srbA are also conserved in Paracoccidioides especially those related to hypoxia, susceptibility to the azoles and iron deprivation responses. Even partially, the Pb01 SREBP was able to restore the mutant phenotypes similarly to wild type strain. In this way, SREBP is a relevant molecule to compensate the effects of hypoxia in A. fumigatus and in Paracoccidioides.
In conclusion, the hypoxia response of Paracoccidioides spp. was largely unknown. In this study, we used a large-scale proteomic approach and a detailed functional characterization of the homologue to the most important molecule involved in hypoxia responses in other fungi, the SREBP protein. Our results show that Paracoccidioides modulates several metabolic pathways in order to compensate for hypoxia stress and importantly it has a functional SREBP homologue, the SrbA protein, which could be involved in regulation of the majority of the hypoxia responses in this pathogen. Taken into account that hypoxia is an important condition faced by pathogens during infection, this characterization becomes relevant in the context of Paracoccidioides spp. pathogenesis and warrants further investigation.
|
10.1371/journal.ppat.1006930 | Oxidative stress and protein damage responses mediate artemisinin resistance in malaria parasites | Due to their remarkable parasitocidal activity, artemisinins represent the key components of first-line therapies against Plasmodium falciparum malaria. However, the decline in efficacy of artemisinin-based drugs jeopardizes global efforts to control and ultimately eradicate the disease. To better understand the resistance phenotype, artemisinin-resistant parasite lines were derived from two clones of the 3D7 strain of P. falciparum using a selection regimen that mimics how parasites interact with the drug within patients. This long term in vitro selection induced profound stage-specific resistance to artemisinin and its relative compounds. Chemosensitivity and transcriptional profiling of artemisinin-resistant parasites indicate that enhanced adaptive responses against oxidative stress and protein damage are associated with decreased artemisinin susceptibility. This corroborates our previous findings implicating these cellular functions in artemisinin resistance in natural infections. Genomic characterization of the two derived parasite lines revealed a spectrum of sequence and copy number polymorphisms that could play a role in regulating artemisinin response, but did not include mutations in pfk13, the main marker of artemisinin resistance in Southeast Asia. Taken together, here we present a functional in vitro model of artemisinin resistance that is underlined by a new set of genetic polymorphisms as potential genetic markers.
| The emergence of artemisinin resistance within and beyond Southeast Asia is a looming threat that needs to be promptly addressed. With this in mind, we derived several artemisinin-resistant parasite lines in vitro in order to fully characterize the resistance phenotype at the cellular and molecular levels. In addition to reinforcing the role of stress responses in mediating artemisinin resistance, we also identified novel genetic alterations that could also be responsible for causing artemisinin resistance. Collectively, this work provides additional insight in relation to, and beyond the paradigm of pfk13-driven artemisinin resistance and artemisinin response in P. falciparum. Understanding the processes that govern the acquisition of artemisinin resistance could aid in the development of strategies to prevent and contain it.
| Malaria remains the most prevalent and deadly vector-borne disease in the world, with an estimated two hundred million cases and over four hundred thousand deaths recorded in 2015[1]. Currently, the cornerstone of global malaria control programs is artemisinin combination therapy (ACT). ACT combines the highly potent, rapidly acting artemisinin-based compounds with long-lasting partner drugs[2]. Artemisinin-based compounds have an excellent safety profile, exert a very rapid parasitocidal effect, and are active against gametocytes and all stages of the intraerythrocytic developmental cycle (IDC), from the early rings to the mature schizonts[3,4]. In particular, artemisinin compounds are typified by their short plasma elimination half-life, ranging from <1 to 3 hours for the water-soluble artesunate (ATS) and dihydroartemisinin (DHA), and from 3 to 11 hours for the oil-soluble artemether[3]. This is in sharp contrast to other antimalarial drugs that have considerably slower elimination times, persisting over several days to several weeks[5]. Hence, artemisinin-based drugs are the frontline therapies used by most, if not all, malaria control programs around the world.
In spite of its wide use, understanding of the artemisinin mode of action remains limited. Artemisinin belongs to the class of sesquiterpene lactones with an endoperoxide bridge that is essential for its antimalarial activity[6,7]. It is widely accepted that artemisinin-mediated parasite killing requires bioactivation of the peroxide structure that leads to generation of reactive oxygen species (ROS) and subsequent damage of biomolecules such as proteins, lipids and nucleic acids[6,7,8]. Some notable protein targets of artemisinin include: PfTCTP, a translationally controlled tumor protein homolog[9] which is located in both the cytoplasm and the food vacuole[10]; Pfatp6[11], an ER-resident, parasite ortholog of sarco/endoplasmic reticulum membrane calcium ATPase; and Pfpi3k[12], which is thought to be an early ring stage target of dihydroartemisinin. Artemisinin-derived radicals have been also shown to alkylate heme[13], which could lead to the disruption of hemozoin synthesis[14] which is essential to parasite survival. Additionally, this class of drugs have also been found to induce the ROS-mediated depolarization of both the mitochondrial[15,16] and plasma membranes[16], representing a different mechanism of parasite killing. It could very well be that the potency of endoperoxide-based drugs against the asexual blood stage is due to their ability to interact with a wide range of targets, across multiple cellular compartments[17,18]. However, the true impact of these interactions on parasite killing remains to be fully understood and requires further investigation.
Indeed, the use of artemisinin combination therapy has led to major progress in malaria control throughout the world in the last two decades, paving the way for better cure rates and reduced transmissibility in the field. From the late 2000s, however, pockets of decreased drug sensitivity to artemisinin-based drugs have been found in Southeast Asia. First detected in western Cambodia[19,20,21], resistance has been now reported from multiple locations across Asia including Thailand[22,23], Myanmar[23,24], Vietnam[23,25], and even Southern China[26]. It is believed that artemisinin resistance is continuously emerging de novo[27,28,29], but a few fit lineages are now spreading regionally[30]. What was originally characterized by delayed parasite clearance among patients treated with ACTs has now escalated to an alarming surge in treatment failures[31]. Interestingly, standard ex vivo 72-hour drug assays that are typically used to measure drug sensitivity are not able to differentiate between the slow-clearing (artemisinin-resistant) and fast-clearing (artemisinin-sensitive) parasites[32]. Instead, the Southeast Asian field isolates exhibit decreased susceptibility to artemisinins only in the very early ring stage of the IDC[32]. Transcriptional and cellular characterization of the resistant isolates demonstrated a delayed progression of the first half of the IDC, particularly the ring stage that is also the least susceptible to artemisinin[33,34,35]. These parasites are also characterized by upregulation of several cellular stress response pathways related to antioxidant defense and the unfolded protein response (UPR)[34]. Crucially, Ariey et a.l 2014 identified a biomarker of clinical artemisinin resistance that can be found in both in vitro and in vivo P. falciparum isolates[36]. After sequencing over 150 Cambodian isolates, they found several nonsynonymous single nucleotide polymorphisms (SNPs) in pfk13 located at chromosome 13 that was similarly mutated in an artemisinin-resistant parasite derived in vitro by artemisinin exposure for over five years[36,37]. Interestingly, the chromosome 13 region around pfk13 was identified independently as one of the genetic regions with a strong signature of selection among Thai and Cambodian parasites with slow clearance rates[38,39]. Subsequent surveillance of Southeast Asian isolates with different genetic backgrounds further corroborated pfk13 as a strong genetic correlate of delayed parasite clearance[23,27,28]. Finally, functional studies validated that specific amino acid changes within the Pfk13 propeller domain significantly increases the rate of parasite survival after early ring-stage treatment with DHA[40]. Although pfk13 is currently the best-characterized molecular marker, many questions remain about the mechanistic links between the amino acid changes in Pfk13 (a putative factor of intracellular protein-protein interactions) and the parasite’s resilience to artemisinin. It is particularly important to uncover all molecular components of the artemisinin resistance mechanism that can act in either a pfk13-dependent or -independent manner. Here we identified several putative factors that can facilitate artemisinin resistance by deriving and characterizing two artemisinin-resistant parasite lines from the P. falciparum 3D7 strain. Through the genomic, transcriptional and chemosensitivity profiling of these in vitro artemisinin-resistant parasites, our findings corroborate the central role of the parasite’s stress responses in mediating artemisinin resistance in Plasmodium, as well as demonstrate the possibility of a robust resistance phenotype that is potentially clinically relevant and is driven by different genomic alterations beyond pfk13.
The overall goal of this research was to identify and characterize molecular factors that contribute to resistance of the malaria parasite P. falciparum, to artemisinin. For this purpose, we derived two parasite lines from two isogenic clones of the 3D7 strain termed 6A and 11C[41]. This was done by repeated exposures of synchronized parasite cultures to 900 nM of artemisinin for 4 hours at the ring stage (10–14 hours post invasion, HPI) (Fig 1A). At the initial earlier stages of the selection process, these pulse treatments were applied every other round of the IDC (see materials and methods). The main rationale of this selection regimen was to approximate clinical conditions in the peripheral blood of infected patients where artemisinin peaks at ~900nM[42] and decays below clinical levels within 2 to 5 hours[3], and where the P. falciparum populations consists predominantly of ring-stage parasites (~10 HPI)[33,43]. The ring stage that is otherwise the least sensitive to artemisinin[44], is believed to be driving the currently occurring artemisinin resistance phenotypes observed in natural infections[32,45]. Initially, during the first 13 treatment cycles, the rate of parasite survival after each artemisinin exposure fluctuated between 30–90% (Fig 1C). These surviving parasites were typically arrested in the ring/trophozoite stages for up to 24 hours post treatment instead of progressing to the expected schizont stages (Fig 1A and 1C, S1 Fig). However, from 18 cycles onwards, 70–100% of parasites were consistently surviving the treatment, progressing normally through the IDC (Fig 1C). We observed a marked decrease in artemisinin susceptibility in both clones as early as 6 rounds of treatment (26 days) for 6A-R, and 8 rounds of treatment for 11C-R (33 days). At that stage the 6A-R and 11C-R exhibited a 3- and 17-fold increase of artemisinin resistance, respectively, as measured by a survival assay establishing the 50% inhibition concentration for parasites exposed to the drug for 4 hours at 10 HPI (IC5010hpi/4hr) (Fig 1B, S1 Table). This drug pulse assay was designed to match the window of the drug selection, resembling the previously utilized shorter exposure drug assays that were shown to capture the stage-dependent artemisinin activities[44,46]. Using this assay, we were observed marked differences in the dynamics of the progression of artemisinin resistance between the two clones throughout the drug selection regimen (Fig 1B, S1 Table). 6A-R showed a gradual increase of IC5010hpi/4hr, starting at 55.49 nM at 6 cycles of artemisinin exposures, progressing to 3,880 nM after 11 months and peaking at 33,726 nM after approximately 1.5 years of continuous treatments. On the other hand, 11C-R exhibited a rapid increase of resistance between 6 and 37 cycles (first 5 months of drug selection) to IC5010hpi/4hr = 3,052 nM. Subsequently, this level of resistance plateaued for the next 19 months of continuous cultivation under drug selection (Fig 1B). Hence, compared to their corresponding parental lines, the resulting artemisinin resistant lines 6A-R and 11C-R exhibited up to a 398- and 69-fold increase in IC5010hpi/4hr, respectively. The drug resistance phenotypes of both lines remained fully intact in parasites that were cryopreserved and reintroduced to culture. Moreover, a significantly elevated IC5010hpi/4hr was maintained after three months of cultivation in the complete absence of drug pressure. In summary, here we derived two artemisinin resistant lines of P. falciparum that could be actively maintained in an in vitro culture and thus serve as a tool for mechanistic studies of artemisinin resistance. The differences in the resistance levels and selection dynamics suggest that the two resistant parasite lines employ (to at least some degree) distinct molecular factors to withstand artemisinin.
Interestingly, the derived resistance phenotype(s) of both 6A-R and 11C-R are predominant in the rings (10 HPI) and do not affect the later stages of IDC development (Fig 2A, S2A Fig, S2 Table). However, for 11C-R, the window of resistance extends until the early trophozoite stage (~20 HPI), where a moderate level of resistance can still be observed. The robustness of the ring-specific artemisinin resistance is likely the main reason for the observed resistance demonstrated by both parasite lines in the standard 72-hour drug assay that measures parasite survival after artemisinin exposure across all stages of the IDC (Fig 2B, S2B Fig, S2 Table). Crucially, both parasite lines also showed decreased drug susceptibility in the ring survival assay (RSA)[32,45] carried out with parasites at 0–3 HPI (Fig 2C). Both 6A-R and 11C-R passed the 1% RSA survival cutoff employed in the field to denote resistance[45,47,48]. This contrasts the current phenotype observed in natural infections that exhibit high levels of ring-stage specific resistance (in the RSA), but show no changes in the standard 72-hour drug assay[32]. Both 6A-R and 11C-R also have significantly elevated IC5010hpi/4hr to other semisynthetic artemisinin derivatives. 6A-R exhibited 5- and 8-fold higher IC5010hpi/4hr to dihydroartemisinin (DHA) and artesunate (ATS), and 11C-R showed 2- and 3-fold higher IC5010hpi/4hr to DHA and ATS, respectively. Both, 6A-R and 11C-R, however, showed no changes in sensitivities to other antimalarial drugs including two quinolines (quinine and chloroquine) and pyrimethamine (Fig 2D, S3A and S3B Fig, S3 Table). Taken together these results suggest that the derived resistance phenotypes are specific to artemisinin and its endoperoxide-carrying derivatives, and can give rise to full resistance phenotypes of the P. falciparum parasites. Given its relevance in the RSA, these mechanisms may correspond to the current artemisinin resistance in natural infections albeit being independent from pfK13 polymorphisms[32,36] (see below). Interestingly, both parasite lines exhibited increased susceptibility to mefloquine, whose mode of action is presumably related to other quinolines[49]. In future studies, it will be interesting to investigate the relationship between the altered sensitivities of P. falciparum to artemisinins and mefloquine. However, here it is important to note that the connection between these two chemosensitivity phenotypes is not absolute as observed in another artemisinin-resistant line derived from a polyclonal population of the T996 P. falciparum strain (S3D Fig).
Stress responses to an oxidative damage and the unfolded protein response (UPR) have been implicated in the mechanisms of artemisinin resistance of P. falciparum in in vitro cultures [50,51] and natural infections[34]. To investigate the role of these two biological processes in the derived resistant parasite lines, we challenged our in vitro-derived resistant parasites with H2O2, dithiothreitol (DTT) and epoxomicin (EPX). While H2O2 causes oxidative damage, DTT and EPX are inducers of ER stress, causing an accumulation of damaged/misfolded proteins inside the cell. Intriguingly, 6A-R, but not 11C-R, exhibited a significant resistance to all three inhibitors (Fig 2E, S3C Fig, S3 Table). This is consistent with our previous suggestion of inherent mechanistic differences in the artemisinin resistance mechanisms between 6A-R and 11C-R and shows that oxidative damage repair and unfolded protein responses play a central role in artemisinin resistance as observed in vivo [33,34,51].
To assess whether the in vitro-derived artemisinin resistant phenotypes reflect the similar physiological state observed in vivo, we characterized the transcriptomes of 6A-R and 11C-R. First we reconstruct the IDC transcriptomes of both resistant clones grown under normal conditions (S4A Fig). The “best fit” parasite aging analysis[33] showed that starting from the mid ring stage (time point 2), both lines progressed identically and completed their IDC in approximately 48 hours. However, both resistant parasite lines appeared to accelerate their early ring stage progression being older (10 HPI) than their sensitive counterparts (4 HPI) at the first sampling interval (S4A Fig). This observation is consistent with the ring-specific resistance in both clones and their resistance in the RSA that appear to be involved in the pfk13-dependent artemisinin resistance observed in vivo. Examining the transcriptomes of 6A-R and 11C-R between 10–20 HPI, we detected broad alterations in mRNA levels of >300 P. falciparum genes (corrected p-value < 0.05, FDR < 0.25), as well as changes in key processes that might be linked to modulating artemisinin response in the parasite (Fig 3A). In 6A-R, pathways related to the redox stress responses and protein turnover were predominant amongst the upregulated genes. Notably, we observed an upregulation of genes that may be related to the parasite’s thioredoxin-based redox system such PF3D7_1457200 (thioredoxin 1), PF3D7_1438900 (thioredoxin peroxidase 1), and PF3D7_1352500 (thioredoxin-related protein). We also observed an enrichment of targets of glutathionylation, as well as targets of the thioredoxin enzyme superfamily. The upregulated protein turnover-associated genes included heat shock and chaperone proteins, and a number of enzymes involved in proteolysis. We also observed an upregulation of genes involved in translational elongation, electron transport, and protein transport, particularly vesicular trafficking between the ER and Golgi complex. On the other hand, the significantly downregulated genes were enriched for pathways related to host-parasite interactions, control of gene expression, and translational initiation (Fig 3A, Table A in S4 Table and S1A File). Interestingly, gene sets involved in cell cycle regulation were also found to be differentially expressed between 6A-R and 6A—which could be related to the slight shift in temporal progression during the early stages of parasite development. In the case of 11C-R, we likewise observed a significant upregulation of genes involved in oxidative stress defense, although to a lesser extent compared to 6A-R. These include genes that encode S-glutathionylated proteins, PF3D7_0306300 (glutaredoxin 1) and PF3D7_0709200 (glutaredoxin-like protein). Pathways involved in protein damage repair, including chaperones and components of proteasome-mediated degradation are also overexpressed. In addition, 11C-R exhibited an upregulation of processes related to early translation events, and transcriptional and post-transcriptional mechanisms of gene regulation such as chromatin modification, stress helicase activity, and the formation of P-bodies. Induction of P-bodies has been observed under stress or conditions that repress translation initiation[52,53,54], and their role in drug resistance may not be ruled out. As for the significantly downregulated functionalities in 11C-R, we identified factors of host-parasite interactions, components of the transcriptional machinery, cellular transport, hemoglobin digestion, several translational elongation factors and ATP synthesis (Fig 3A, Table B in S4 Table and S1B File). Evaluating the transcriptional correspondence of 6A-R and 11C-R with slow clearing isolates in Southeast Asia from the TRAC I[34], we found a great degree of overlap between significantly upregulated pathways in the in vitro and in vivo datasets (Fig 3B). Strikingly, several of these pathways have also been associated with longer parasite clearance half-lives in the field such as coping mechanisms against ER stress (ER trafficking, proteasome-mediated degradation, translation) and oxidative stress (targets of glutathionylation), as well as mRNA processing[34]. Not only does this observation reinforce the involvement of these cellular processes in modulating artemisinin resistance in Plasmodium, it also demonstrates that 6A-R and 11C-R are each able to recapitulate key aspects of in vivo artemisinin resistance at the transcriptional level.
Next we analyzed global transcriptional responses of 6A-R and 11C-R to artemisinin drug exposure that is identical to the selection conditions (synchronized parasites were treated with 900nM artemisinin from 10 to 14 HPI) (S4B Fig). Here we observed many similarities between how 6A-R and 11C-R respond to a ring-stage artemisinin challenge in relation to their sensitive counterparts. Notably, both lines exhibit a downregulation of processes pertaining to pathogenesis, transcriptional control, translation, cellular transport and cell cycle regulation (S4B Fig, Tables A and B in S5 Table). That both in vitro-derived lines demonstrate a marked dysregulation of genes involved in cell cycle regulation could be related to their ability to overcome the drug induced quiescence caused by artemisinin. Interestingly, GSEA identified transport across the ER-Golgi and digestive vacuole (DV) membranes as significantly upregulated in 11C-R (S4B Fig, Table B in S5 Table and S2B File). Likewise, 6A-R also displayed an upregulation in transmembrane transport components—a number of which have been linked to drug resistance in Plasmodium. A notable example is the DV-resident chloroquine resistance transporter, pfcrt, which is significantly upregulated in both resistant lines and has been associated with chloroquine resistance [55,56,57,58]. Pfcrt also plays a role in glutathione transport and antioxidant defense within the DV[59]. Pfexp1, a glutathione transferase located on the parasitophorous vacuole and is associated with artesunate sensitivity and metabolism[60] is also found to be upregulated in 6A-R (S2A File). On the other hand, we also observed transcriptional features that are distinct to only one parasite line, such as the downregulation of autophagy-related pathways in 6A-R vs. 6A, and the observed downregulation of heat shock proteins in 11C-R compared the 11C. It is probable that the differences we observed in global transcriptional profiles between 6A-R and 11C-R could account for some of the phenotypic variations between these two parasite lines such as resistance to H2O2/DTT/EPX. In the future it will be interesting to study these variations as they could represent genuine differences in drug resistance phenotypes in vivo.
The whole genome sequencing of the two resistant parasite lines identified several intragenic SNPs compared to their parental lines. These included 3 and 5 missense mutations in 6A-R and 11C-R, respectively; one nonsense mutation in each parasite line and an additional intronic SNP in 6A-R (Table 1). As a result, there are mutation alleles for five genes in 6A-R and six genes in 11C-R. Cross-referencing our SNP data with the Pf3k[61] and MalariaGEN[62] databases, the nonsynonymous mutations detected in PF3D7_1427100, PF3D7_0810600 and PF3D7_1115700 were found to also occur in natural infections of African origin. Crucially, there was no overlap between the mutated genes in the two parasite lines, both of which also carried the wild-type allele of the K13 gene (validated by PCR-based genotyping of pfk13[36,63]). No polymorphisms were also detected in previously identified drug resistance markers, such as pfcrt[56,64], pfmrp1[65,66], pfmdr1[67], pfnhe-1[68], pfdhps[69], pfdhfr[70,71], pfatp6[72], pfubp1[73], pfap2mu[74] PF3D7_101700[39], and PF3D7_1343400[39]. Moreover, none of the SNP-containing genes in 6A-R and 11C-R match the previously reported putative targets and interacting partners of artemisinin, such as pfatp6[11], pfpi3k[12], pftctcp[9] and other proteins [17,18]. The only exception is the nonsense mutation in pffp2a (PF3D7_1115700) that encodes falcipain 2a, the main factor of hemoglobin digestion, whose nonsense polymorphism was previously linked with artemisinin resistance in vitro[36,37]. On the other hand, both 6A-R and 11C-R harbor mutations in genes that might play a role in gene expression regulation such as AP2-like transcription factors, a PHD finger protein and an RNA helicase. Such genes could be implicated in the regulation of the Plasmodium IDC transcriptional cascade and subsequently contribute to the resistance phenotypes of both parasite lines.
Next, we characterized the genome-wide patterns of copy number variations (CNVs) using microarray-based comparative genomic hybridization (CGH) as previously described[41]. In both 6A-R and 11C-R, we detected two gDNA segments whose amplifications could be directly related to their artemisinin resistance status (Fig 4, S6 Table). Namely, there is a segment on chromosome 14 spanning 40 genes (PF3D7_1454000-PF3D7_1458000) amplified in 6A-R, and a segment on chromosome 12 spanning 9 genes (PF3D7_1228000—PF3D7_1228800) amplified in 11C-R. Moreover, both 6A-R and 11C-R also carry a common amplification on chromosome 10 spanning 17 genes (PF3D7_1028700—PF3D7_1030300). This selfsame amplification has also been identified previously in an artemisinin sensitive P. falciparum 3D7[75] strain. The three CNVs on chromosomes 10, 12 and 14 were detected during the later stages of drug selection and subsequent culturing, and were consistently detected over a period of five months (89 generations) (S5 Fig). Comparing our CNVs with a dataset collated from 122 clinical isolates from Africa, South East Asia and South America[76], we found that none of the isolates contained the chromosome 10, chromosome 12 and chromosome 14 amplification clusters in their entirety. However, one isolate collected from Peru harbored a copy gain for the putative gamma-adaptin encoding PF3D7_1455500.
Given the scope of the detected transcriptional changes in 6A-R and 11C-R, we wished to investigate the possibility that CNV-driven variations in expression can mediate artemisinin resistance. Evaluating the individual expression levels of each gene in the three CNV clusters identified, we found that not all transcripts appear to be significantly overexpressed across the IDC between resistant parasites and their sensitive counterparts (Fig 5A and S6 Table). However, comparing the collective expression among the amplified genes on chromosomes 10, 12 and 14, we were able to detect a significant enrichment of upregulation in the genes located in these regions (Fig 5A). This observation is particularly striking in the case of the chromosome 14 amplification in 6A-R, where 30 out of the 40 genes were significantly overexpressed across the IDC (S6 Table). Here we focused on three genes on the 6A-R chromosome 14 amplification that are likely to be involved in adaptive responses against cellular damage within the parasite. These include 6-phosphogluconate dehydrogenase (PF3D7_1454700, pf6pgd) and thioredoxin 1 (PF3D7_1457200, pftrx1)—both of which are involved in antioxidant defense[77,78,79], and an ER-resident signal peptide peptidase (PF3D7_1457000, pfspp)[80] that is a component of ER associated degradation (ERAD)[81]. All three candidate genes were found to be significantly overexpressed in 6A-R compared throughout the IDC (Fig 5A and S6 Table). In order to assess their potential to confer artemisinin resistance, we generated transgenic P. falciparum lines in which each candidate gene was overexpressed episomally. Briefly, each gene was fused with the HA-antibody epitope at the C-terminus and cloned into the pBcamR_3xHA transfection vector (see materials and methods) that allows adjustable expression via increased copy number driven by blasticidin (BSD). Quantitative RT-PCR demonstrated increased transcription of the transgenic contracts by 7-fold for pf6pgd and 2-3-fold for pftrx1 and pfspp (Fig 5B). Western blot analysis confirmed the production of the HA-tagged transgene protein products at their expected molecular weights in the transgenic cell lines grown in the presence of 2.5 ug/mL BSD (Fig 5C). Crucially, overexpression of pftrx1, and pfspp resulted in a subtle but significant decrease in artemisinin sensitivity, with IC5010hpi/4hr 1.7-fold, and 2.9-fold higher than the “empty vector” control, respectively (Fig 5D, S6 Fig and S7 Table). On the other hand, no significant difference in artemisinin sensitivity could be observed in the parasites overexpressing pf6pgd. These results indicate that the specific upregulation, possibly as a result of gene amplification, of pftrx1 and pfspp contributed to the decreased sensitivity of 6A-R to artemisinin.
It has been previously shown that resistance can be induced in culture-adapted P. falciparum parasites through long-term exposures to artemisinin and/or its derivatives[37,50,82,83]. That includes the studies that discovered the current principal marker of artemisinin resistance in Southeast Asia, pfk13 [36,37]. The identification of the pfk13 gene highlights the value of such in vitro models to systematically investigate the mechanisms that drive artemisinin response and resistance in the clinical setting. Here, we developed two artemisinin resistant cell lines from isogenic clones of the 3D7 P. falciparum strain. For this study, we chose two isogenic clones of the 3D7 reference strain that has been previously extensively characterized, and thus will lead to efficient identification of all derived genetic variation. The 3D7 strains also represent a fully artemisin-susceptible background which provides a “naïve” baseline genome that potentially allows for the identification of causative factors of artemisinin resistance that are independent of any potential genetic background with a propensity for drug resistance[27,84]. This yielded a resistance phenotype(s) that is (are) distinct from the previous reports. Essentially all previously derived P. falciparum parasites involved an artemisinin-induced growth arrest and recovery as a major component of the resistance phenotype[37],[50],[85],[83]. In contrast, 6A-R and 11C-R are both characterized by an increased survival in the presence of artemisinin with no detectable levels of growth retardation or arrest. This marked difference is likely due to the pulse-based regimen that contrasts the previous studies in which the parasite lines were treated for considerably longer time periods, ranging from 24–48 hour drug exposure intervals[36,37] to continuous drug pressure[50,82,83]. Moreover, 6A-R and 11C-R displayed significant decreases of artemisinin sensitivity (IC5010hpi/4hr) within as early as 1.5 months of selection. This is also in stark contrast with previous reports by Witkowski et al. that showed that the chemosensitivity of the P. falciparum F32 strain remained unaltered for up to 3 years and/or 100 cycles of drug pressure when the parasites were treated with artemisinin for 24 hours at a time[37]. Similarly, Cui et al. were unable to raise drug resistance in the 3D7 strain at all and could only generate resistant parasites using other culture adapted P. falciparum strains including 7G8, Dd2, HB3 and D10 after at least one to two months of continuous exposure to DHA[50]. This collectively indicates that artemisinin resistance of P. falciparum could be derived by multiple ways, each of which may induce a distinct mechanism.
Unsurprisingly, the artemisinin resistance in both parasite lines extends to its cognate drugs ATS and DHA. But while 6A-R and 11C-R showed up to almost 400- and 70-fold increases in IC5010hpi/4hr values for artemisinin (Fig 1B, S1 Table), respectively, both lines exhibited increases in IC5010hpi/4hr for ATS and DHA by less than 10-fold (Fig 2D, S3 Table). This is likely a reflection of key differences in pharmacodynamic profiles between artemisinin and its two synthetic derivatives. Compared to the plant-derived artemisinin, both ATS and DHA are more potent antimalarials with DHA being the primary cytopathic metabolite responsible for the parasite killing[6,86]. In contrast, artemisinin is not metabolized to DHA but instead acts as the primary antimalarial agent itself and is subsequently transformed into inactive deoxyartemisinin and dihydrodeoxyartemisinin[4,87,88]. The variance in the resistance level of the two derived clones could be attributed to differences in the overall levels of the cytopathic activities, the mode of activation, and/or the protein targets that each compound is specifically engaging. Moreover, while 6A-R and 11C-R did not exhibit cross-resistance to other types of antimalarials, both clones are more susceptible to mefloquine (Fig 2D). Interestingly, DHA-resistant parasites previously derived by Cui et al. from a Dd2 parent, displayed decreased sensitivity to other artemisinin-based drugs albeit to a lesser extent compared to DHA, but also to quinine, chloroquine and mefloquine[50]. Furthermore, parasites derived using long term exposure to artelinic acid from the D6 and W2 backgrounds showed cross-resistance to mefloquine but increased susceptibility to chloroquine[82,83]. These findings allude to the possibility that resistance to artemisinin-based drugs could also affect the clinical efficacy of its partner drugs used in the currently deployed ACTs. These results highlight the importance of testing for cross-resistance as an integral part of drug development, and also demonstrate a key use for in vitro drug resistance models that can be utilized as a platform with which to perform such extensive and rigorous studies.
The P. falciparum parasites causing the current state of slow clearing infections in the Southeast Asian patients show are marked by higher RSA values[45] but show no differential sensitivities in standard in vitro drug assays[20,32,89]. These in vivo parasites are characterized by transcriptional induction of oxidative and (other types of) protein damage responses, and at the same time, a deceleration of the early stages the IDC [33,34]. Both of these transcriptional phenotypes are strongly linked with mutations in the pfk13 gene as the main marker of artemisinin resistance [27,28,36]. Here we observed several main similarities between in vivo artemisinin resistance and the in vitro-derived phenotypes of 6A-R and 11C-R. First, like the slow-clearing isolates in Southeast Asia, the resistance of 6A-R and 11C-R is tied to the earlier stages of the IDC, and fades as the parasites progress into the later stages. This is reflected by an elevation in the RSA index (>1%) for both 6A-R and 11C-R that is comparable to the in vivo isolates (Fig 2C). Second, both resistant lines demonstrated a steady state upregulation of genes and pathways that are involved in antioxidant defense, as well as the UPR (Fig 3A and 3B)[34]. Crucially, the induced artemisinin resistance in the 6A-R clone also gave rise to cross-resistance against oxidative agents (e.g. H2O2), protein-folding disruptors (e.g. DTT), and stressors of protein processing in the ER (EPX) (Fig 2E). This indicates that that its derived resilience to artemisinin is tied to an increased capacity to mediate oxidative stress and protein damage. These findings suggest that one possible mechanism for artemisinin resistance is an enhanced ability of the P. falciparum parasites to cope with the oxidative stress and protein damage presumably caused by artemisinin directly. On the other hand, the in vitro-derived lines were unable to recapitulate certain features of the resistant isolates. Neither 6A-R nor 11C-R experienced a dramatic shift in the temporal progression of the IDC, nor did they develop artemisinin-resistance associated genotypes that have been previously observed in vivo—most notably, mutations in pfk13. In spite of these genotypic and phenotypic discrepancies, these derived parasite lines nonetheless provide a unique opportunity for future analyses of artemisinin resistance in the context of multiple genetic backgrounds[27,28]. The apparent lack of pfk13 polymorphisms in 6A-R and 11C-R suggests that these parasites may serve as a model to study the relevant mechanisms driving the PfK13-independent artemisinin resistance phenotype newly emerging in Southeast Asia[48] and Africa[90].
The prerequisite of a genetic background and the possibility of “PfK13-independence” suggest that other genetic polymorphisms will contribute to the overall phenotype of artemisinin resistance that is currently in existence or will emerge in the future. Genomic profiling of 6A-R and 11C-R revealed unique sets of SNPs and CNVs that could represent such polymorphisms. Surprisingly, these polymorphisms did not involve genes with associations to any drug sensitivity phenotypes of malaria parasites reported in the past. The two exceptions include pffp2a and pfprp22. Pffp2a is a cysteine protease involved in hemoglobin digestion that is believed to mediate the activation of artemisinin presumably via the release of haemoglobin-derived iron[91]. Indeed Pffp2a can modulate artemisinin sensitivity in the ring stages[92], and a nonsense mutation in pffp2a has been previously found in an in vitro-derived artemisinin-resistant parasite line[36,37]. Hence the presence of the nonsense mutation the pffp2a likely contributes to artemisinin resistance in 6A-R. The amplification of pfprp22 in both 6A-R and 11C-R on the common segment of chromosome 10 coincides with its duplication in another resistant parasite line derived from a D6 strain using artelinic acid[83,93]. However, this amplification at chromosome 10 had already been reported in naturally occurring infections including artemisinin sensitive parasites[75]. Hence its role in artemisinin resistance remains unclear. Here, we were able to substantiate the potential of the chromosome 14 CNV to influence the parasite’s sensitivity to artemisinin by the specific overexpression of two key genes in this region: pftrx1 and pfspp. Both PfTrx1 and PfSpp play a role in the parasite’s antioxidant defense system and/or protein damage stress response. Thioredoxin 1 is a key enzyme in the Plasmodium NADPH-dependent thioredoxin system which is involved in the detoxification of reactive oxygen metabolites, redox regulation of transcription factors, and control of protein folding[77,78,94], while Signal Peptide Peptidase is a transmembrane protease component of the ER-associated degradation pathway, which is utilized by the parasite to cope with damaged or misfolded proteins[81]. Hence, the upregulation of pftrx1 and pfspp likely supports the increased capacity of the UPR which, in eukaryotic cells, subsequently employs trafficking across cellular compartments, enzymatic processing of proteins to mediate their folding and degradation, and attenuation of translation to mitigate the ER workload[95,96,97]. Consistent with this model, 6A-R and 11C-R both demonstrated differential expression of genes related to translational control and the regulation of gene expression, including early translational events (initiation), binding and processing of messenger RNA, as well as transcriptional regulation via transcription factors and chromatin modification (see Fig 3A, Tables A and B in S4 Table). It has been previously shown that that Plasmodium is also able to cope with cellular stresses via translational repression involving the eif2α-mediated attenuation of global protein synthesis[97,98,99], and the association of mRNA with RNA-binding proteins that facilitate their stability (stress granules) or degradation (P-bodies)[98,99,100]. Transcriptional changes in many of these pathways were also observed in the in vivo isolates[34]. Taken together, this data represents a spectrum of SNPs and CNVs that may represent multiple, alternative genetic events that are yet to be observed or validated in the field but could emerge and spread in the (near) future. These could either deepen the existing pfk13-dependent artemisinin resistance phenotypes, or could give rise to new mechanisms compounding alternative genetic backgrounds of P. falciparum populations (e.g. Indian or African)[27,48,90].
Two clonal parasite lines, named 6A and 11C, were previously derived from the P. falciparum 3D7 strain using limiting dilution[41] and subsequently used for in vitro drug selection. Continuous cultivation of parasites was performed as previously described[101]. Cultures were maintained in purified human red blood cells at 1–2% hematocrit, in RPMI 1640 medium (Gibco) supplemented with 0.25% Albumax I (Gibco), 2 g/L Sodium bicarbonate (Sigma), 0.1 mM hypoxanthine (Sigma), and 50 μg/L gentamicin (Gibco). Parasite cultures were kept at 37 o C with 5% CO2, 3% O2, and 92% N2 and treated twice with 5% (v/v) sorbitol (Sigma) every cycle to maintain stage synchronicity. Culture medium was replenished every 12–24 hours, and freshly washed uninfected red blood cells (RBC) was added to the culture as needed. Monitoring of parasitemia and parasite morphology was performed using microscopic evaluation of thin blood smears that were first fixed with methanol (Merck), and then stained with Giemsa (Sigma).
Ethical approval for the use of blood in this study was granted by the Institutional Review Board of the Nanyang Technological University. All of the blood utilized for the in vitro cultivation of parasites was derived from healthy adult volunteers, and extracted by trained personnel at the National University Hospital Blood Donation Center, Singapore. All donors provided their written informed consent.
6A and 11C parents were each divided into two parasite lines: one selection line (6A-R, 11C-R), which would be subjected to artemisinin selection, and one control line (6A and 11C), which would undergo mock treatment with dimethyl sulfoxide (DMSO) (Sigma). All parasite lines were synchronized at 4 HPI and diluted to a parasitemia (percentage of parasitized erythrocytes) of 2–5% prior to drug treatment. Each selection line was then pulse treated with a 900 nM artemisinin (Sigma) diluted in DMSO for four hours from 10–14 HPI; Control lines were also pulse treated in parallel with pure DMSO, for four hours at 10–14 HPI. During treatment, all parasites were kept at 2% hematocrit with 1 mL of parasitized blood. After treatment, the media containing artemisinin and DMSO were removed, and the parasite pellets washed twice with fresh media. Parasites were then resuspended in fresh media. Blood smears fixed with methanol and then stained with Giemsa were prepared for each parasite line 20–24 hours after washing to obtain post-treatment parasitemia as well as observe any morphological effects of drug treatment. During the initial phase of drug selection, artemisinin-treated parasites were allowed to recover to a viable parasitemia of at least 2% before artemisinin treatment. Once the parasite lines were able to consistently survive artemisinin pressure, they were maintained as synchronized cultures and subjected to pulse treatment with 900nM artemsinin from 10–14 HPI every other asexual cycle. Cultures were not kept away from artemisinin/DMSO treatment for more than three consecutive generations. Both sets of parasite lines were subjected to the same number of artemisinin and DMSO treatments throughout the course of drug selection, and at the same generations.
|
10.1371/journal.ppat.1001216 | Legionella Metaeffector Exploits Host Proteasome to Temporally Regulate Cognate Effector | Pathogen-associated secretion systems translocate numerous effector proteins into eukaryotic host cells to coordinate cellular processes important for infection. Spatiotemporal regulation is therefore important for modulating distinct activities of effectors at different stages of infection. Here we provide the first evidence of “metaeffector,” a designation for an effector protein that regulates the function of another effector within the host cell. Legionella LubX protein functions as an E3 ubiquitin ligase that hijacks the host proteasome to specifically target the bacterial effector protein SidH for degradation. Delayed delivery of LubX to the host cytoplasm leads to the shutdown of SidH within the host cells at later stages of infection. This demonstrates a sophisticated level of coevolution between eukaryotic cells and L. pneumophila involving an effector that functions as a key regulator to temporally coordinate the function of a cognate effector protein.
| Many bacterial pathogens encode a large array of “effector proteins” that are essential for successful infection. By definition, effector proteins are synthesized in bacteria and transported from bacteria into host cells. Within host cells, effector proteins directly interact with host factors in order to modulate their functions. Effector expression, translocation or activity within host cells must be precisely regulated over infection stages. Here we demonstrate the first example of an effector protein which targets and regulates another effector within host cells: Legionella effector protein LubX targets another effector protein SidH to proteasome-mediated protein degradation in the host cells. Expression and delivery of these effector proteins are differentially regulated, which results in LubX-dependent SidH shutdown at late stages of infection. We propose the designation “metaeffector” for this class of bacterial effector protein: an effector that targets and regulates another effector within host cells. Future studies may reveal that metaeffectors which play critical roles in coordinating the functional expression of other effectors spatiotemporally are prevalent among bacterial pathogens.
| Many bacterial pathogens encode a large array of “effector proteins,” that manipulate host cellular processes during infection. Effector proteins are translocated from bacteria directly into the cytosol of host cells. This process is mediated by dedicated bacterial protein delivery systems, including the type III and the type IV secretion systems. In some cases, effector proteins delivered into host cells by a bacterium have opposing functions on a single host protein. For example, Legionella pneumophila DrrA (SidM) and LepB are effector proteins with opposing effects on the host Rab1 GTPase, with DrrA functioning as a guanine nucleotide exchange factor (GEF) and guanine nucleotide dissociation inhibitor-displacement factor (GDF), and LepB having GTPase-activating protein (GAP) activity[1], [2], [3], [4]. Similarly, the Salmonella enterica serovar typhimurium effectors SopE and SptP have GEF and GAP activities for the Rho family of GTPases[5], [6], respectively. Although the GEF activity of SopE is dominant in the host cell immediately after infection, degradation of SopE by the host proteasome alters the balance of these effectors, resulting in the GAP activity of SptP to be dominant later in infection [7]. Although differential regulation of gene transcription and post-translational modifications of effectors have also been shown to regulate their activities in host cells[8], [9], details on how these processes are controlled remain largely unknown; other effector-regulating mechanisms probably also exist.
L. pneumophila is a gram-negative bacterium ubiquitously found in freshwater environments [10]. When phagocytosed by eukaryotic cells, L. pneumophila remodels the Legionella-containing phagosome to form a compartment that allows its intracellular replication [11], [12], [13]. As a result, L. pneumophila is able to replicate in a wide variety of phagocytic cells, from amoebae to macrophages; human infections can result in a severe pneumonia called Legionnaires' disease [14]. The Dot/Icm type IV secretion system is an essential virulence determinant that translocates L. pneumophila effector proteins into host cells during infection [15], [16]. These effector proteins control host cell functions to initiate trafficking of the L. pneumophila vacuole, promote host cell survival, modulate innate immune responses, and promote bacterial egress [17]. Although over 100 L. pneumophila effector proteins have been identified, the biochemical and cellular functions of most effector proteins remain unknown [17], [18].
Ubiquitin is a small, well-conserved peptide of 76 amino acids, present in all eukaryotes [19]. Ubiquitination of substrate proteins involves a cascade of reactions. At the last step of the cascade, an E3 ligase recognizes a substrate protein and transfers ubiquitins to the substrate from an E2 conjugating enzyme [20]. Ubiquitinated proteins are subjected to further cellular processes, most notably proteasomal degradation [21]. E3 ligases can be divided into several major families; HECT-type, RING-type, U-box-type and NEL-type [22], [23], [24]. The each family has distinct structural feature, while RING and U-box domains are closely related. Many but not all RING-type E3 ligases work as multi-subunit complexes called SCF complexes containing Skp1, Cullin and F-box proteins. U-box-type E3 ligases contain single U-box domain that serves as an E2-binding site, while L. pneumophila effector protein LubX carries two U-box domains, one of which functions as a substrate-binding site (Figure 1A, see below). NEL-type is the most recent addition to E3 ubiquitin ligase families [24]. NEL family is composed of IpaH/SspH family proteins from bacterial pathogens Shigella flexneri and Salmonella enterica [25], [26], [27] and more than 30 homologous proteins found in bacterial pathogens[24]. Most notably NEL-type E3 ligases seem to be prevalent among bacterial pathogens, whereas no homologous protein has found in eukaryotic cells.
It is well documented that bacterial pathogens exploit the host ubiquitin-proteasome pathway by delivering effectors that function as E3 ubiquitin ligases or as deubiquitinating enzymes [24], [28], [29]. L. pneumophila encode one U-box protein (LubX) [30] and several F-box-containing proteins including AnkB/LegAU13/Lpg2144/Lpp2082 [31], [32], [33], [34], [35]. We previously reported that LubX functions as a U-box-type E3 ubiquitin ligase in vitro and in host cells [30]. LubX mediates polyubiquitination of a host kinase Clk1, but its consequence remains unknown. Interestingly, the expression and translocation of LubX is induced upon infection and the levels of LubX within host cells come to maximum at later stages of infection, compared to other L. pneumophila effectors so far characterized. The gene encoding LubX is in close proximity to genes that encode several other type IV effectors, including VipD [36] and SidH [37] (Figure 1B). Surprisingly, analyses of these effector proteins led us to identify SidH as a target of LubX. LubX acts as a negative temporal regulator of SidH within host cells. This is the first example of the bacterial effector that targets and regulates a cognate effector within the host cells, and we propose the designation “metaeffector” for this class of bacterial effectors.
Translocation of putative effector proteins encoded in vicinity of the lubX gene was assessed by measuring cAMP production in the host cytosol generated by an effector containing an amino-terminal fusion to an adenylate cyclase (Cya) domain that is only active in the cytosol of eukaryotic cells[38], [39]. These measurements appear to indicate that translocation of the Cya-SidH fusion protein by wild-type L. pneumophila was significantly less than that by an isogenic lubX mutant producing the same fusion protein at eight hours post infection (Figure 1C—the asterisk [*], Lp01 vs. ΔlubX). The difference was the most potent at late stages of infection, while we did not see significant difference at one hour post infection (Figure S1). Translocation of other Cya-tagged effectors such as RalF [40], however, was not affected by the lubX mutation (Figure 1C) suggesting that the lubX mutation does not have a general effect on type IV secretion. Because LubX has ubiquitin ligase activity, we investigated whether LubX affected SidH translocation through a process requiring host proteasome activity, by treating cells with the proteasome inhibitor MG132. Remarkably, wild type L. pneumophila and the lubX mutant appeared to translocate Cya-SidH equally in cells treated with the proteasome inhibitor (Figure 1C vs. 1D), indicating that inhibition of the host proteasome mimics a bacterial mutant deficient in LubX. These results suggested that LubX-mediated proteasomal degradation of a factor within the host cytosol was required for the reduced levels of Cya-SidH activity. It should be noted though that the Cya fusion assay is not an ideal system to examine the dynamics of intracellular effector levels in infection context, partly because the Cya fusions are under control of a non-authentic constitutive promoter and expressed in trans.
Because LubX is able to directly target proteins for degradation, we next examined whether LubX-mediated degradation of SidH in the host cytosol was due to a direct interaction between LubX and SidH. Purified proteins were used to test for direct interactions between SidH and LubX in vitro (Figure 2A). His-SidH became bound to the purified GST-LubXΔC, a deletion derivative lacking the C-terminal domain of LubX, but not with GST alone, suggesting interaction between SidH and LubX. The SidH interaction was detected using a fusion protein containing the LubX U-box2 region (GST-U-box2), but not using a fusion protein containing the LubX U-box1 region (GST-U-box1). Another Legionella effector RalF became bound neither to GST-LubXΔC nor to GST-U-box2 (Figure 2B), suggesting that LubX U-box2 is a specific protein binding domain. Collectively the U-box2 region of LubX binds specifically and directly to the effector protein SidH.
An in vitro ubiquitination assay was used to determine if LubX binding to SidH could target SidH for ubiquitination. Purified components used were ubiquitin, E1, UbcH5c (E2), LubXΔC or its inactive derivative LubXΔCI39A (E3), His-SidH or another effector protein RalF, and ATP, and the reactions were conducted as described previously[30]. In a functional LubX dependent manner, His-SidH shifted to a very high molecular weight species (Figure 3A, IB: αSidH). Another effector protein RalF was not affected (Figure 3A, IB: αRalF), suggesting the specificity of the reaction. To examine whether the retarded His-SidH species contain ubiquitin, His-SidH was isolated from the reaction mixtures by pull-down using nickel resin and analyzed by western immunibloting using anti-polyubiquitin antibodies (Figure 3A, PD: His-SidH IB: αUbiquitin). The results indicated that the retarded His-SidH species were polyubiquitinated. These results clearly demonstrate that His-SidH is polyubiquitinated by LubXΔC in vitro.
To determine whether SidH is polyubiquitinated by LubX in the cytosol of infected host cells, nucleotide sequences encoding an amino-terminal triple-FLAG (3× FLAG) epitope tag were appended to the sidH gene on the L. pneumophila chromosome. The strain encoding the 3× FLAG-SidH protein expressed SidH at similar levels and the regulation of LubX expression was not affected (Figure S2). Chinese hamster ovary (CHO)-FcγRII cells infected for eight hours with L. pneumophila strains producing 3× FLAG-SidH were extracted with a buffer containing 1% digitonin. L. pneumophila proteins recovered in the extracts contain proteins translocated into the host cells, but not proteins in the bacterial cells[30], [41]. The 3× FLAG-SidH protein was not detected in cells infected with wild-type L. pneumophila, while it was readily detectable in cells infected with the lubX mutant (Figure 3B: DMSO, wild type vs ΔlubX). Importantly, when host cells were treated with MG132 to inhibit proteasome-mediated degradation, polyubiquitinated 3× FLAG-SidH derivatives were detected from cells infected with L. pneumophila producing a functional LubX protein, but only unmodified SidH was detected from cells infected with the isogenic lubX mutant (Figure 3B: MG132). The defect in SidH degradation in cells infected with the lubX mutant was rescued by producing LubX in trans (Figure 3C). Thus LubX is essential for polyubiquitination of SidH in host cells, and polyubiquitinated SidH is degraded by the host proteasome.
LubX expression by L. pneumophila is induced intracellularly. L. pneumophila grown extracellularly on laboratory media does not produce LubX, and the levels of LubX increase gradually upon host cell infection, peaking at 10 hours post infection [30] (and Figure S2). In contrast, expression of SidH is induced at the stationary phase of growth in laboratory media, which means SidH levels in the bacterial cell are high when infection is initiated (Figure S2). Accordingly, we hypothesized that SidH levels are regulated intracellularly by the temporal expression and intracellular activities of LubX during infection. SidH levels were measured over time in CHO-FcγRII cells infected with L. pneumophila producing 3xFLAG-SidH to test this hypothesis. Full-length 3× FLAG-SidH was detected within the host cells shortly after infection (Figure 4A, 15 m). At one hour post infection, polyubiquitinated 3× FLAG-SidH was detected. Intracellular levels of 3xFLAG-SidH declined over time, and by eight hours post infection, 3× FLAG-SidH was no longer detected. By contrast, LubX levels within the host cells increased over time (Figure 4A), consistent with LubX mediating the intracellular ubiquitination and degradation of SidH. In cells infected with the L. pneumophila lubXI39A mutant, similar levels of 3× FLAG-SidH were detected in host cells at all time points and the protein was not polyubiquitinated (Figure 4B: lubXI39A), indicating that LubX E3 ubiquitin ligase activity is required for the temporal degradation of SidH. When cells were treated with MG132, we observed the increased levels of polyubiquitinated 3× FLAG-SidH derivatives over time by a process requiring LubX activity (Figure 4C: MG132). Lastly, pretreatment of L. pneumophila with the irreversible bacterial translation inhibitor gentamicin abrogated the shutdown of SidH, indicating that intracellular production of LubX was necessary for SidH degradation (Figure 4D: Gm and Figure S3). These results clearly indicate that SidH transiently accumulates within host cells at an early stage of infection, and that the eventual disappearance of SidH in host cells results from LubX-mediated proteasomal degradation (Figure 4E).
The temporal regulation model predicts that the persistence of intracellular SidH led by lubX disruption adversely affects L. pneumophila fitness in hosts. To address the prediction, we utilized the Drosophila infection model [42]. It has been shown that L. pneumophila infect and grow to high levels within Drosophila cells, and their replication depends on the Dot/Icm type IV secretion system [43]. Correspondingly, although most flies infected with wild-type Legionella died within twelve days, ∼80% of flies infected with Legionella defective in the Dot/Icm type IV secretion system (ΔdotA) survived (Figure 5A). Numbers of the surviving flies infected with the sidH mutant were similar to or slightly higher than those infected with the wild-type strain (Figure 5A). The lubX mutant consistently showed hyper-lethality to flies compared with the sidH mutant or the lubX sidH double mutant (Figure 5A, P<0.01 for both comparisons). Viable bacterial counts in survived flies infected with the lubX mutant were consistently lower than those in flies infected with the wild-type strain (Figure 5B). This apparent defect of replication of the lubX mutant in flies was rescued by further introduction of the sidH mutant (Figure 5B). Thus, the loss of lubX in sidH+ L. pneumophila gives a disadvantage in multiplication within the model host. It should be noted that the hyper-lethality of the lubX mutant did not stem from increased number of viable bacteria (Figure 5B). The lubX mutant might be more toxic to a specific type of fly cells (e.g. phagocyte) important for survival.
The temporal regulation of SidH mediated by L. pneumophila LubX E3 ligase and the host proteasome system illustrates a novel mechanism by which bacterial effectors are regulated. In the previously reported temporal regulation of the induced membrane ruffling by Salmonella effectors SopE and SptP, it was shown that these two effectors have distinct susceptibility towards ubiquitin-mediated proteasomal degradation [7]. Importantly, the determinants of the susceptibility are encoded in the effectors themselves. By contrast, the L. pneumophila SidH is an intrinsically stable protein within host cells; the effector protein LubX controls SidH instability in the host cytosol by directly targeting this protein for host ubiquitination. Thus LubX represents a bona fide metaeffector—a designation for an effector that regulates the function of another effector within the host cell. Discovery of an effector having a regulatory role on another effector provides unique insight into the sophisticated mechanisms that underlie the ability of L. pneumophila to coordinate the function of such a large array of effector proteins with diverse activities.
As shown previously [30] and partly in Figure 4 as well, the intracellular level of LubX increases over time and reaches to the maximum at 10–12 hours post infection, which is far late stages compared to the critical time window in which SidH is polyubiquitinated and targeted to proteasomal degradation. It appears that relatively small amount of LubX detected within first couple of hours post infection is sufficient for the polyubiquitination and degradation of SidH (Figure 4). This raised the question on the role of LubX at late stages of infection. We previously reported that host kinase Clk1 is a substrate of LubX [30]. Together with the current findings, one of reasonable explanations would be that LubX has multiple targets within host cells including bacterial SidH and host Clk1.
Accumulating lines of evidence suggest that effector proteins functioning as an E3 ubiquitin ligase are prevalent among many plant and animal bacterial pathogens. Some effector proteins which have been reported to possess E3 ligase activity do not show similarity to eukaryotic E3 ligases at sequence level. AvrPtoB of a plant pathogen Pseudomonas syringae shows structural similarity to U-box E3 ligase, whereas they do not share sequence level homology [44]. This suggests that there could be more bacterial E3 ligases than predicted by simple homology search. Structural analyses demonstrated that IpaH/SspH family proteins Shigella flexneri IpaH1.4, IpaH3 and Salmonella enterica SspH2 show no similarity to eukaryotic E3 ligases and represents a new class of E3 ligase [25], [26], [27]. This class of E3 ligase, called NEL, seems to comprise a large family of bacterial E3 ligases, because more than 30 homologous proteins are found in a subset of bacterial pathogens [24]. Furthermore, there are more than 30 bacterial F-box proteins which are expected to be a key component of SCF E3 ligase complexes [28]. However, the targets of these bacterial E3 ligases remain largely unknown, and all previous studies have aimed to identify host target proteins. SidH is the first example of bacterial targets of the bacterial E3 ligase effector. Future studies without prejudice will reveal that some of bacterial effectors implicated in manipulation of the host ubiquitin system possess regulatory functions as metaeffector that coordinate the expression of effector functions spatiotemporally.
All Legionella strains used in this study were derivatives of L. pneumophila strain Lp01 [45] and were grown on charcoal-yeast extract (CYE) plates or in ACES-buffered yeast extract (AYE) broth as described previously[46]. The strains defective in lubX and dotA genes have been previously described [30], [47]. Legionella strains defective in sidH and lubX-sidH genes as well as Legionella strains carrying 3xFLAG-sidH or the lubX I39A mutation were constructed by allelic exchange[47]. E. coli strain BL21 (strain B lon ompT) was used as an expression host for protein purification. Plasmids used in this study and details of plasmid construction are provided in the Tables S1 and S2. Chinese hamster ovary (CHO)-FcγRII cells were cultured at 37°C in 5% CO2 in α-MEM, supplemented with heat-inactivated 10% FBS, as described[48].
Custom made anti-serum against SidH peptide CQNIKGPEPVATPMETPE (SidH 2196–2212) was purchased from MBL. Antibodies were purified from the anti-serum by affinity chromatography using peptide-conjugated SulfoLink resins (Pierce). The affinity-purified rabbit polyclonal antibodies against RalF, LubX and GroEL were described previously [30], [40]. Mouse monoclonal antibody against polyubiquitin (clone FK1) was purchased from BIOMOL. Monoclonal antibody against the FLAG tag (M2) was purchased from Sigma-Aldrich.
Translocation of Cya-fused proteins into CHO-FcγRII cells after infection with Legionella was assayed as described previously [30], [38] with minor modifications. Briefly, CHO-FcγRII cells were replated in 24-well plates, and challenged by Legionella strains expressing Cya fusions at a multiplicity of infection (moi) of 30 in the presence of opsonizing antibody (1∶3000 dilution). Eight hours later, infected cells were lysed in 500 µl of lysis reagent 1B provided from a cAMP Biotrak EIA System (GE Healthcare, RPN2251); cAMP levels were determined according to manufacturer's instructions.
Purification of LubXΔC, RalF and GST fusion proteins has been described previously [30], [40]. For purification of His-SidH, BL21 cell pellets from a 2-liter culture expressing His-SidH were suspended with 80 ml PBS containing Complete Protease Inhibitor Cocktail (Roche) and 20 mg lysozyme (Wako Chemical). After incubation with stirring for 30 minutes at 4°C, the lysozyme-treated cells were lysed by sonication. After centrifugation (16,000× g for 20 minutes) to remove unsolubilized materials, the supernatant fraction was mixed with ammonium sulfate (final 40% saturation) and incubated for 30 minutes at 4°C. After centrifugation to remove precipitates, the supernatant fraction was mixed with ammonium sulfate (final 60% saturation) and incubated for 30 minutes further at 4°C. After centrifugation, the precipitates were dissolved in 20 ml PBS containing Complete Protease Inhibitor Cocktail (Roche). This solution was dialyzed against PBS to remove residual ammonium sulfate. After centrifugation to remove insoluble materials, the supernatant fraction was mixed with 6 ml (bed volume) HIS-Select Resin (Sigma-Aldrich) and incubated for 30 minutes at 4°C. The resins were washed 5 times with PBS containing 5 mM imidazole, and the bound proteins were eluted with 6 ml PBS containing 100 mM imidazole. The elution step was repeated once more, and the resulting two eluate fractions were pooled. The pooled fraction was mixed with half its volume of 20 mM Tris HCl, pH 7.5 to reduce salt concentration (final 0.1 M NaCl). The resulting solution was applied to a MonoQ 5/50GL chromatography column (GE healthcare). After elution by NaCl gradient (0.1 M to 0.5 M in 20 mM Tris HCl, pH. 7.5), the peak fractions were further subjected to a Superose 6 10/300GL column (GE healthcare) equilibrated with 20 mM Tris HCl pH 7.5, 150 mM NaCl. The peak fractions were pooled and concentrated using a Microcon device (Millipore).
For a GST pull-down using purified proteins, GST, GST-LubXΔC, GST-U-box1 or GST-U-box 2 (10 µg) were mixed with 5 µg of His-SidH or RalF in 500 µl PBS containing 1 mM EDTA, 1 mM DTT, and 1% (w/v) Triton X-100. The resulting solutions were mixed with 25 µl of a 50% suspension of Glutathione-Sepharose and incubated for three hours with gentle rotation at 4°C. Unbound proteins were removed by centrifugation, and resins were washed four times with the same buffer, and once with a buffer omitting TritonX-100. GST and interacting proteins were eluted with 50 µl of SDS sample buffer containing reducing agent.
The in vitro ubiquitin polymerization assay was performed essentially as described [30] with a couple of modifications; 300 nM E3 enzyme (LubX derivatives) were employed; where indicated, 240 nM of purified His-SidH or RalF was included. Where indicated, His-SidH derivatives were pulled down with His-Select resin (Sigma-Aldrich) and washed in the presence of 2.5 mM imidazole to suppress nonspecific interaction between reaction component proteins and the resin. Pulled-down materials were eluted with 250 mM imidazole. Samples were subjected to 10% SDS-PAGE and analyzed by immunoblotting using antibodies against SidH, RalF or polyubiquitin.
CHO-FcγRII cells were replated in a 6-well culture dish, and challenged by Legionella strains at a moi of 30 in the presence of opsonizing antibody (1∶3000 dilution). One hour after infection, the cells were washed three times with PBS (pre-warmed to 37°C) to remove non-internalized bacteria, then further incubated in cell culture medium. When indicated, 10 µM of MG132 (Calbiochem) or equivalent amount of the solvent DMSO was added to the cell medium 30 min prior to infection as well as to the replacing medium. At the indicated time points, the cells were washed three times with cold PBS and lysed in 150 µl of PBS containing 1% (w/v) of digitonin (Calbiochem), 10 mM of N-ethylmaleimide (Sigma) to prevent deubiquitination, and protease inhibitor cocktail (1∶100 dilution, Sigma-Aldrich). The cells were scraped off, collected into microfuge tubes and centrifuged at 16,000× g for 10 min at 4°C to separate the digitonin-soluble fraction containing translocated proteins from the digitonin-insoluble fraction containing internalized bacteria. The digitonin-soluble fractions were filtrated through a 0.45 µm filter unit (Millex-HV, Millipore). Immunoprecipitates with anti-FLAG or anti-LubX antibodies, extracted using nProteinA Separose (GE Healthcare), were analyzed by SDS-PAGE followed by immunoblotting using anti-FLAG or anti-LubX antibodies, respectively.
Five to seven days olds yw male Drosophila melanogaster flies were used for infection experiments. Before injection, the bacteria-containing medium was adjusted to 0.1 OD using Gene Quant pro (Amersham) with distilled water. Flies were anesthetized with CO2 and injected with each strain of bacteria in 65 nl of water (approximately 500 colony-forming units). Injection was carried out by using an individually calibrated pulled glass needle attached to IM-300 microinjector (Narishige). Flies were always injected in the abdomen, close to the junction with thorax and just ventral to the junction between the ventral and dorsal cuticles. After injection, flies were transferred to fresh vials once a week. For colony-forming assay, 10 days after bacterial injection flies were homogenized in 10 mM MgSO4 solution and the diluted series of the homogenized samples were plated on CYE media containing 100 µg/ml streptomycin.
|
10.1371/journal.pcbi.1006234 | Subgraphs of functional brain networks identify dynamical constraints of cognitive control | Brain anatomy and physiology support the human ability to navigate a complex space of perceptions and actions. To maneuver across an ever-changing landscape of mental states, the brain invokes cognitive control—a set of dynamic processes that engage and disengage different groups of brain regions to modulate attention, switch between tasks, and inhibit prepotent responses. Current theory posits that correlated and anticorrelated brain activity may signify cooperative and competitive interactions between brain areas that subserve adaptive behavior. In this study, we use a quantitative approach to identify distinct topological motifs of functional interactions and examine how their expression relates to cognitive control processes and behavior. In particular, we acquire fMRI BOLD signal in twenty-eight healthy subjects as they perform two cognitive control tasks—a Stroop interference task and a local-global perception switching task using Navon figures—each with low and high cognitive control demand conditions. Based on these data, we construct dynamic functional brain networks and use a parts-based, network decomposition technique called non-negative matrix factorization to identify putative cognitive control subgraphs whose temporal expression captures distributed network structures involved in different phases of cooperative and competitive control processes. Our results demonstrate that temporal expression of the subgraphs fluctuate alongside changes in cognitive demand and are associated with individual differences in task performance. These findings offer insight into how coordinated changes in the cooperative and competitive roles of cognitive systems map trajectories between cognitively demanding brain states.
| Brain networks support the human ability to navigate a complex space of perceptions and actions through cognitive control. Here we ask, “How do brain networks coordinate task-relevant information as individuals adapt to cognitive demands imposed by a task?” We study the fMRI BOLD signal of twenty-eight healthy subjects as they perform two cognitive control tasks—a Stroop interference task and a local-global perception switching task using Navon figures—with low and high cognitive load conditions. We construct functional networks and use a machine learning technique called non-negative matrix factorization to identify topological motifs whose expression fluctuates across different phases of cognitive control processes. We find that motifs stratify the brain network into a hierarchy of distributed functional processes that adapt to changes in cognitive demand and predict individual differences in task performance. These data offer insight into how network interactions linking cognitive systems coordinate transitions between cognitively demanding brain states.
| In human cognition, internally-generated cognitive control processes modulate attention, facilitate task switching, and inhibit prepotent behavior [1]. One avenue by which the brain may rapidly traverse a cognitive state-space is through its functional interactions—coherent fluctuations in brain activity shaped by the structural connectome [2]. The brain’s distributed functional interactions form a functional network whose architecture is temporally dynamic [3], conferring adaptivity in the face of environmental pressures or task demands [4] such as those elicited during learning [5] and other tasks demanding executive cognition [6]. Cognitive control processes have been widely reputed to recruit several cognitive systems that include executive, attention, and salience systems that span prefrontal cortices, striatum, parietal regions, and cerebellum [7–12]. The notion that cognitive control involves a heterogenous collection of brain systems is supported by several univariate studies demonstrating concurrent activation of functionally-specialized brain areas across different cognitive control tasks [13]. If patterns of measured brain activity signal involvement of different brain regions across a diverse set of cognitive control tasks, then how do functional brain networks encode and coordinate this task-relevant information to adapt to fluctuations in cognitive demand (Fig 1A)?
One mechanistic theory, known as the “adaptive coding model of cognitive control” [17], posits that brain regions that activate during higher cognitive functions can alter their dynamical properties based on the current goals of the neural system. More recent studies have challenged this hypothesis by presenting data that suggests that changes in the cognitive demands of a task lead to recruitment of mechanistically-specialized brain regions based on an anatomically-defined gradient [18, 19]. To reconcile these opposing theories of the neuronal basis of cognitive control, [20] applied multivoxel pattern analysis—a machine learning technique for identifying consistent patterns of voxel-wise activation—to the fMRI of subjects as they performed simple and cognitively demanding tasks. The authors found a consistent pattern of activation in frontoparietal brain areas that was specific to highly demanding conditions across multiple cognitive tasks. Their findings support the hypothesis that a consistent group of brain regions activate in response to increases in cognitive demand. However, parallel lines of investigation on the underpinnings of cognitive control in functional brain networks suggest that the integrated cognitive control network dissociates into several, segregated sub-networks that are responsible for different aspects of cognitive control processes [13]. To address these conflicting reports, a data-driven approach that can disentangle parts of functional brain networks that encode cognitive states associated with control tasks—and track their expression alongside changes in cognitive demand—is required. Such a capability would improve our understanding of which components of functional brain networks are important for different facets of cognitive control, and how these components encode shifts between cognitively demanding states.
In the present work, we identify components of functional brain networks associated with transitions between cognitively demanding states by using an unsupervised machine learning technique known as non-negative matrix factorization (NMF) [21]. NMF decomposes functional brain networks into: (i) additive subgraphs that represent clusters of graph edges that track with one another over time, and (ii) time-varying coefficients that quantify the degree to which a subgraph is expressed at a given point in time [16, 22, 23]. This computational tool allows us to track how groups of functionally interacting brain areas are dynamically expressed during experimentally modulated changes in cognitive demand (Fig 1B; a discussion regarding the differences between NMF and components analysis can be found in textitMaterials and methods). In particular, we ask participants to engage in the following two cognitive control tasks: a response inhibition Stroop task (Fig 1C; [24]) and a local-global perception switching task based on classical Navon figures (Fig 1D; [25]). Our methodological approach enables us to address a critical question in cognitive control: “How do brain networks coordinate task-relevant information as individuals adapt to the cognitive demands imposed by a task?”
To address this question using NMF, we draw upon recent studies that suggest that task-driven reconfiguration of functional brain networks integrates otherwise functionally-specialized and segregated information [26, 27]. One compelling current theory proposes that transitions between cognitively demanding brain states are facilitated by dynamic changes in the patterns of correlated and anticorrelated brain activity such that anticorrelated fluctuations in brain activity represent segregated brain functions, and correlated fluctuations in brain activity represent integrated brain functions [28, 29]. Correlated and anticorrelated dynamics may explain how task-relevant information is shared between different regions of the network during cognitively demanding tasks. In this study we construct functional brain networks by applying the Pearson correlation function to block level fMRI collected during cognitive control tasks. By accounting for correlated and anticorrelated functional interactions in the NMF framework, we can determine the likelihood that the functional interactions within a subgraph are collectively correlated or anticorrelated at a particular point in time—providing a perspective on integrated and segregated information processing among composite sets of brain regions.
Based on prior studies demonstrating that behavioral tasks can be used to dissociate intrinsic and task-specific architectures of functional brain networks [30], we first hypothesize that NMF will identify functional subgraphs whose expression is either generalized across the Stroop and Navon tasks or specific to distinct cognitive conditions within and between tasks. In particular, we expect task-general subgraphs to reflect interactions relevant for task saliency and cognitive control processes common to both tasks. We also expect task-specific subgraphs to reflect interactions relevant for stimulus processing and attentional mechanisms necessary for either response inhibition in the Stroop task or task-switching in the Navon task. Building upon recent evidence that functional interactions dynamically reorganize between integrated and segregated network states [26], we next hypothesize that functional subgraphs will shift their roles between correlated and anticorrelated modes of interaction in response to experimentally driven changes in cognitive demand. Lastly, we hypothesize that changes in subgraph expression during experimental modulation of cognitive demand will reflect inter-individual differences in behavioral performance on the task. Specifically, based on previous theories regarding the behavioral influence of correlated and anticorrelated functional interactions in cognitive control [29], we expect that components of the frontoparietal and default mode systems will most prominently participate in subgraphs associated with individual differences in performance.
To uncover the topological organization and putative roles of correlated and anticorrelated functional interactions in cognitive control, we first acquire fMRI data as 30 healthy adult human subjects perform Stroop and Navon cognitive control tasks. Two subjects are excluded on the basis of poor performance and technical problems on the day of scanning, leaving 28 subjects for further analysis. In particular, we measure fMRI BOLD signals from 262 functional brain areas (Fig 2A)—including cortex, subcortex, and cerebellum [31, 32]—during three separate conditions of both the Stroop and Navon tasks: fixation, low cognitive demand, and high cognitive demand conditions (Fig 2B). Briefly, the low cognitive demand condition is designed to elicit a neural response associated with performing each task with low cognitive control demands and the high cognitive demand condition is designed to elicit a neural response associated with either task shifting or inhibition cost (see Material and methods for more details). We then construct dynamic functional brain networks for each subject where network nodes represent brain regions and network edges represent the Pearson correlation coefficient between regional BOLD time series (Fig 2C). Specifically, we compute a 262 × 262 adjacency matrix for each of 6 task blocks (corresponding to 30 seconds of BOLD activity, and comprising several trials) in each of the 3 task conditions (fixation, low demand, high demand) for each of 2 tasks (Stroop and Navon). This process results in 36 block-level adjacency matrices per subject. Importantly, positive Pearson correlations underlie integrated and coherent activation between brain regions or correlated functional interactions, and negative Pearson correlations underlie segregated and discordant activation between brain regions or anticorrelated functional interactions [28]. To separate positively-weighted network edges (correlated interactions) from negatively-weighted network edges (anticorrelated interactions), we duplicate the adjacency matrix of each block and separately threshold edge weights either greater than zero or less than zero (see Materials and methods for details). Lastly, we aggregate all functional brain networks into a network configuration matrix (Fig 2D) with size 2016 × 34191. The first dimension of size 2016 corresponds to all combinations of two tasks, three task conditions, six repeated blocks, twenty-eight subjects, and two edge types (correlated or anticorrelated); the second dimension of size 34191 corresponds to all unique, pairwise edges between the 262 brain regions.
We first assess the extent to which task-specific differences in functional network topology are explained by first-order, global network statistics, by comparing the distribution of mean edge strengths across different dimensions of the network configuration matrix (S1 Fig). We find no significant difference in mean edge strength across subjects between blocks during the Stroop task and blocks during the Navon task (paired t-test, t27 = −1.5, p = 0.14; S1B Fig). We also find no significant difference in mean edge strength across subjects between blocks during the low cognitive demand condition and blocks during the high cognitive demand condition (paired t-test, t27 = 0.35, p = 0.73; S1D Fig). We find a significant decrease in mean edge strength between blocks during the fixation period and blocks during the cognitive control task period (paired t-test, t27 = 4.7, p6.3 × 10−5; S1C Fig), suggesting that engaging in cognitive control tasks is associated with a reduction of correlated and anticorrelated BOLD dynamics. Critically, this result suggests that attention-related brain states thought to be associated with the fixation period may be subserved by stronger and more well-defined functional relationships between brain regions than more complex, task-driven brain states. Overall, our findings suggest that differences in functional network topology during cognitive control tasks and control conditions are not driven by first-order differences in mean edge strength of the network. Rather, we expect that differences in the topological organization of correlated and anticorrelated BOLD dynamics may be complex and heterogenously distributed across the functional network.
To disentangle patterns of correlated and anticorrelated BOLD dynamics related to cognitive control processes, we extract functional subgraphs and their dynamic expression from functional brain networks. Specifically, we apply an unsupervised machine learning algorithm called non-negative matrix factorization (NMF) to the network configuration matrix. This technique enables us to pursue a parts-based decomposition of network edges into additive functional subgraphs with accompanying expression coefficients that measure the degree to which the subgraph is expressed in a particular task block, task condition, subject, and edge type (Fig 2D) [22, 23]. Each subgraph composes a 262 × 262 adjacency matrix and each subgraph’s expression coefficients compose a vector of length 2016. Thus, subgraphs detail topological components of the functional brain network and temporal coefficients quantify their expression during different phases of the cognitive control tasks. Moreover, each subgraph is associated with a positive expression component associated with correlated BOLD dynamics and a negative expression component associated with anticorrelated BOLD dynamics (Fig 2E). A critical step in using NMF is optimizing model parameters (number of subgraphs m, sparsity of subgraph edge weights β, and regularization of temporal expression coefficients α) to ensure generalizability of component subgraphs without overfitting the model on observed data. By designing a four-fold, leave-seven-subjects-out cross-validation scheme, we minimize the average cross-validation error on held-out subjects and find the optimal number of subgraphs to be twelve, the subgraph sparsity to be 0.29, and the regularization of the temporal expression coefficients to be 0.56 (S2 Fig; see Materials and methods for more details). For a quality check on the effect of motion confounds on subgraph expression, we refer the reader to S3 Fig. For a test-retest reliability assessment of subgraph reproducibility we refer the reader to S4 Fig.
We next rank the twelve subgraphs (A-L) in decreasing order of their relative positive or negative expression across all conditions in the cognitive control tasks. Specifically, we compute the difference between the positive expression coefficient corresponding to correlated dynamics and the negative expression coefficient corresponding to anticorrelated dynamics for each task block and average the difference across blocks of each subject (Fig 2F). Intuitively, subgraphs whose mean relative expression values are positive are more likely to capture correlated BOLD dynamics and subgraphs whose mean relative expression values are negative are more likely to capture anticorrelated BOLD dynamics. We refer to specific subgraphs according to their assigned letter for the remainder of the study.
We next ask whether the functional subgraphs expressed during the cognitive control tasks reflect functional interactions within and across known cognitive systems. To study the relationship between the functional subgraph architecture and known cognitive brain systems, we assign each of the 262 brain regions into one of nine cognitive systems [33]: dorsal attention, default mode, frontoparietal, limbic, somatosensory, subcortical, ventral attention, visual, and cerebellum. Thus, we re-organize the rows and columns of each subgraph’s 262 × 262 adjacency matrix such that nodes assigned to the same brain system are contiguously ordered, and we visualize the resulting adjacency matrices as circular, ring graphs (Fig 3; for matrix representation see S5 Fig). To quantitatively confirm that each subgraph captures functional interactions that are indeed distributed within and between cognitive systems, we compare the average subgraph edge weight between pairs of nodes of the same or different cognitive systems to a null distribution of the average subgraph edge weight—constructed by permuting subgraph edge weights between nodes and recomputing the average subgraph edge weight for each pair of cognitive systems for 10000 permutations. We find that functional subgraphs cluster interactions between brain regions of the same cognitive system and between brain regions of different cognitive systems (p < 0.05; Bonferroni corrected for multiple comparisons; see S5 Fig)—implicating a distributed functional architecture underlying the cognitive control tasks. In other words, the functional subgraphs recovered by NMF span several cognitive brain systems defined a priori [33].
Based on the distribution of subgraph edges within and between known cognitive systems, we examine how subgraph topology might underlie different information processes during cognitive control. Interconnected complex systems that underlie distributed information processes—such as those involved in cognitive control—can exhibit core-periphery structure in which a strongly interconnected core of nodes is connected to other nodes in the network periphery, which tend to solely connect with core nodes and remain otherwise isolated from other network regions [34, 35]. Intuitively, the putative function of the network core is to integrate information from different, specialized systems located in the network periphery [36, 37]. The core-periphery model has also been extended to accommodate dynamic functional networks in which the network core exhibits less flexible functional connectivity and the network periphery exhibits more flexible functional connectivity [36]. A critical assumption of recent applications of the core-periphery model is that there is a single set of core regions and a single set of periphery regions. It is plausible that brain networks consist of multiple core structures [37–39] that are activated based on ongoing cognitive processes reflected by network subgraphs. To identify core-periphery organization in a subgraph, we compute a core-periphery index (see Methods) that quantifies the difference between mean edge strength within a cognitive system (network core) and mean edge strength between a cognitive system and all other systems, averaged for each cognitive system. Intuitively, the core-periphery index ranges between −1—stronger edges in the periphery than in the core—and + 1—stronger edges in the core than in the periphery; index values closer to 0 imply equally strong edges in the core and in the periphery characteristic of traditional core-periphery structure. We use a surrogate subgraph model (10000 rewiring permutations) to statistically test whether each cognitive system of each subgraph exhibits core and periphery architecture (details regarding specific cognitive systems significantly involved in each subgraph may be found in S6 Fig. We find that different functional subgraphs exhibit varying degrees of core-periphery organization (Fig 3). For example, subgraph A expresses significant core connectivity in seven of nine cognitive systems and significant periphery connectivity in one of nine cognitive systems (S6 Fig). In contrast, subgraph K expresses significant core connectivity in zero of nine cognitive systems and significant periphery connectivity in three of nine cognitive systems (S6 Fig). Intuitively, subgraph A reflects core organization where systems exhibit more centralized topology and subgraph K reflects periphery organization where systems exhibit more decentralized topology. We find that cognitive systems in the remaining subgraphs tend to exhibit both internally centralized connectivity as well as decentralized connectivity to other systems. The differentiation of subgraphs into constituent core-periphery architectures suggests that subgraphs may reflect different modes of integrated and segregated network processes in which task-relevant information may be organized within core cognitive systems and shared with cognitive systems in the network periphery.
Logically, we next ask the question “How does the core-periphery organization of a functional subgraph relate to its dynamical expression during cognitive control?” To answer this question, we examine the relationship between the core-periphery index of a subgraph and its mean relative expression. We hypothesize that functional subgraphs with stronger edges adjoining brain regions in the network core (core-periphery index closer to +1) are expressed more positively and functional subgraphs with stronger edges adjoining brain regions between the network core and network periphery (core-periphery index closer to −1) are expressed more negatively. Intuitively, subgraphs with stronger edges within the core and weaker edges between the core and periphery will be associated with more correlated BOLD dynamics underlying states of integrated cognitive processes, and subgraphs with stronger edges between the core and periphery and weaker edges within the core will be associated with more anticorrelated BOLD dynamics underlying states of segregated cognitive processes. Using the Spearman’s ρ, we find a significant positive correlation between core-periphery index and relative subgraph expression (ρ = 0.76, p = 0.004; Fig 3). This result supports the hypothesis that subgraphs with greater sensitivity to topology within the network core tend be positively expressed and subgraphs with greater sensitivity to topology between the network core and network periphery tend to be negatively expressed. Importantly, we observe that functional subgraphs with more evenly balanced core-periphery topology (core-periphery index close to 0) also tend to be more positively expressed. Collectively, our findings demonstrate that subgraphs with strong core topology or balanced core-periphery topology are associated with network states in which brain regions exhibit correlated dynamics and that subgraphs with strong periphery topology are associated with network states in which brain regions exhibit anticorrelated dynamics. By examining the relationship between subgraph topology and subgraph expression, we may now begin to bridge theoretical interpretations of subgraph architecture with experimentally driven and empirically observed changes in cognitive brain state.
Based on the set of twelve functional subgraphs and their time-varying expression, we next ask “Are functional subgraphs differentially recruited during separate cognitive control tasks?” We hypothesize that a functional subgraph is either sensitive to cognitive control processes specific to each task or to cognitive control processes that are shared between the two tasks. To motivate our hypothesis, we examine relative differences in the distributions of mean strength of each edge between all task blocks of the Stroop task and all task blocks of the Navon task (S7 Fig), before extracting functional subgraphs using NMF. Specifically, we compare the strength of an edge during the Stroop task to its strength during the Navon task by computing the mean difference of Fisher’s r-to-Z transformed correlations across subjects, separately for positive correlations and negative correlations. We observe stronger positive correlations within and between the dorsal attention, visual, and cerebellar systems during the Navon task than during the Stroop task, and we observe stronger negative correlations between the default mode system and dorsal attention, visual, and cerebellar systems during the Navon task than during the Stroop task.
We use these results to inform our expectation regarding cognitive systems that might be involved in task-specific functional subgraphs. We examine the relationship between the mean relative expression of a subgraph during the Stroop task and the mean relative expression of a subgraph during the Navon task (Fig 4). We find that the expression of a subgraph during the Stroop task is significantly associated with its expression during the Navon task (Spearman’s ρ, ρ = 0.97, p = 1.3 × 10−7). This result suggests that subgraphs are similarly ranked based on their respective expression values between the two tasks. Critically, this result implies that subgraph expression may follow a consistent hierarchy of expression during two different cognitive control tasks.
While the relative relationships between subgraph expression are preserved between the Stroop task and the Navon task, we also identify differences in the magnitude of subgraph expression between the tasks. Specifically, we compare the distribution of relative subgraph expression between the Stroop task and the Navon task for each subgraph. Using paired t-tests and FDR correction for multiple comparisons, we find greater positive expression during the Navon task than during the Stroop task for subgraph B (t27 = 4.4, p = 1.4 × 10−4) and subgraph D (t27 = 2.9, p = 7.0 × 10−3), and we find greater negative expression during the Navon task than during the Stroop task for subgraph K (t27 = 5.1, p = 1.4 × 10−5). These findings suggest that (i) the Navon task exhibits greater correlated BOLD dynamics within and between dorsal attention, visual, and cerebellar systems (subgraph B) and between the default mode system and other broadly distributed cognitive systems (subgraph D) than the Stroop task, and (ii) the Navon task exhibits greater anticorrelated BOLD dynamics between the default mode system and dorsal attention, visual, and cerebellar systems (subgraph K) than the Stroop task. Critically, subgraph D and subgraph K both capture functional relationships between the default mode system and other cognitive systems. However, they exhibit different types of interactions—correlated versus anticorrelated—and involve different sub-regions of the default mode system that engage or disengage with other cognitive systems. This heterogeneity may underlie a multi-faceted functional role for cognitive systems involved in both positively expressed and negatively expressed subgraphs as regions of information integration and information segregation during these tasks.
To summarize, our results imply that functional subgraphs follow a general hierarchy of expression during two cognitive tasks that invoke different control processes—pre-potent response inhibition during the Stroop task and perceptual, rule-based task switching during the Navon task. While this hierarchy may establish a task-general functional network organization related to complex cognitive processes, specific processes associated with different forms of cognitive control may be represented through small deviations in subgraph expression that significantly differ between tasks. Accordingly, nine of the twelve subgraphs were not significantly more expressed in any particular task than expected by chance and may implicate functional network components that are expressed during processes that are agnostic to task-specific mechanics, such as arousal.
We next ask “How do functional subgraphs adapt to experimentally imposed changes in cognitive demand during the different cognitive control tasks?” We hypothesize that a functional subgraph is either sensitive to cognitive control processes specific to the experimentally imposed changes in cognitive demand or to the stimulus and task mechanics that are shared between low and high cognitive demand conditions of each task. To motivate our hypothesis, we examine the relative differences in the distributions of mean strength of each edge between the low demand conditions and high demand conditions of the Stroop task and the Navon task (S8 Fig), before extracting functional subgraphs using NMF. Specifically, we compare the strength of an edge during the low demand condition of a cognitive task to its strength during the high demand condition of the task by computing the mean difference of Fisher’s r-to-Z transformed correlations over subjects, separately for positive and negative correlations. For the Stroop task, we observe: (i) stronger positive correlations within the dorsal attention system and between the dorsal attention, cerebellar, default mode, and frontoparietal systems, and (ii) stronger negative correlations within the cortical limbic system and between the cortical limbic system and other broadly distributed cognitive systems during the high demand condition compared to the low demand condition. For the Navon task, we observe: (i) stronger positive correlations within and between the dorsal attention, visual, and cerebellar systems, (ii) stronger positive correlations within the frontoparietal system, and between the frontoparietal system and other broadly distributed cognitive systems, and (iii) stronger negative correlations within somatosensory and ventral attention systems, and between somatosensory and ventral attention systems and other broadly distributed cognitive systems during the high demand condition compared to the low demand condition.
We use these results to inform our expectation regarding cognitive systems that might be involved in functional subgraphs that adapt to changes in cognitive demand. We first examine the relationship between the mean relative expression of a subgraph during the low cognitive demand condition of a task and during the high cognitive demand condition of a task (Fig 5). We find that the expression of a subgraph during the low cognitive demand condition is significantly associated with its expression during the high cognitive demand condition for the Stroop task (Spearman’s ρ, ρ = 0.99, p = 4.1 × 10−9) and for the Navon task (Spearman’s ρ, ρ = 0.99, p = 4.1 × 10−9), suggesting that subgraphs follow a similar ranked order in their relative expression before and after the increase in cognitive demand. Critically, this result implies that subgraphs follow a consistent hierarchy of expression during the low demand and high demand conditions of each task.
While the relative relationships between subgraph expression are preserved between cognitive demand conditions, we also identify differences in the magnitude of subgraph expression between demand conditions. Specifically, we compare the distribution of relative subgraph expression between the low cognitive demand condition and the high cognitive demand condition of each task for each subgraph using paired t-tests and FDR correction for multiple comparisons. For the Stroop task, we find greater positive expression during the high demand condition than the low demand condition for subgraph B (t27 = 3.3, p = 2.7 × 10−3) and subgraph E (t27 = 3.2, p = 3.6 × 10−3) and greater negative expression during the high demand condition than the low demand condition for subgraph L (t27 = 2.5, p = 0.01). These findings suggest that the Stroop task (i) exhibits greater correlated BOLD dynamics during the high demand condition than during the low demand condition within and between dorsal attention, visual and cerebellar systems (subgraph B), and within and between default mode and frontoparietal systems (subgraph E), and (ii) exhibits greater anticorrelated BOLD dynamics during the high demand condition than during the low demand condition within the limbic and subcortical systems, and between the limic and subcortical systems and other broadly distributed cognitive systems (subgraph L). For the Navon task, we find greater positive expression during the high demand condition than during the low demand condition for subgraph G (t27 = 2.9, p = 8.2 × 10−3), and we find greater negative expression during the high demand condition than during the low demand condition for subgraph F (t27 = 2.7, p = 0.01). These findings suggest that the Navon task (i) exhibits greater correlated BOLD dynamics during the high demand condition than during the low demand condition between frontoparietal and default mode systems and other broadly distributed cognitive systems (subgraph G), and (ii) exhibits greater anticorrelated BOLD dynamics during the high demand condition than during the low demand condition within somatosensory and ventral attention systems, and between somatosensory and ventral attention systems and other broadly distributed cognitive systems (subgraph F).
Overall, we find that functional subgraphs follow a general hierarchy of expression that remains consistent between low cognitive demand conditions and high cognitive demand conditions, and adaptively shift their expression alongside experimentally invoked changes in cognitive demand. Critically, our findings imply that subgraphs may maintain a robust network representation of each cognitive control task between different states of cognitive demand and may adaptively encode different cognitive control processes via shifts in positive or negative expression such that the overall hierarchical representation of the task remains undisturbed. These shifts in subgraph expression are evidently coordinated through changes in correlated and anticorrelated BOLD dynamics involving select subgraphs. Accordingly, these results suggest that the functional brain network may utilize task-specific control strategies by coordinating antagonistic changes in the co-activation between different cognitive systems during pre-potent response inhibition (Stroop task) and during perceptual, rule-based task switching (Navon task).
We next examine how the recruitment of functional subgraphs relates to the change in inter-individual performance as participants invoke cognitive control mechanisms. Our approach is based upon prior studies that posit a functional role of antagonistic dynamics between correlated and anti-correlated brain activity in cognitive control processes [28, 29]. We use the subgraph characterization of the functional network to directly examine how behavioral performance is related to the extent that distinct networks exhibit more correlated or anti-correlated dynamics. To evaluate the change in an individual’s task performance—also known as performance cost—we separately compute mean change in an individual’s reaction time between consecutive blocks of the low demand condition and the high demand condition. Intuitively, a lower reaction time cost indicates better performance and a higher reaction time cost indicates worse performance. Using the reaction time cost as a behavioral marker for inter-individual differences in cognitive control processes, we study the functional role of subgraphs during the following two phases of the cognitive control tasks: (i) task activation associated with the low cognitive demand condition, and (ii) task control associated with the high cognitive demand condition. To quantify the association between subgraph expression and performance, we first compute each individual’s relative subgraph expression as the difference between the likelihood that a subgraph is positively expressed (i.e., functional dynamics are correlated) and the likelihood that a subgraph is negatively expressed (i.e., functional dynamics are anti-correlated). We next use Spearman’s ρ to assess the relationship between relative subgraph expression during the low or high demand condition, and the reaction time cost across individuals on each task (Fig 6A–6D; left). Indeed, if the correlation between relative expression and reaction time cost is positive, then individuals who express more correlated dynamics (and less anti-correlated dynamics) within a subgraph exhibit poorer performance and individuals who express more anti-correlated dynamics (and less correlated dynamics) within a subgraph exhibit better performance. This analysis approach enables us to understand the extent to which behavior is explained by both the degree to which regions in a subgraph engage with one another via correlated dynamics and the degree to which regions in a subgraph disengage from one another via anti-correlated dynamics.
We find a diverse set of subgraphs whose relative expression during low demand, task activation conditions or high demand, task control conditions correlate with reaction time cost. For the Stroop task, we find that a lower reaction time cost (better performance) is associated with: (i) greater negative expression of subgraph A (ρ = 0.44, p = 0.01; uncorrected for multiple comparisons) during the low demand condition, and (ii) greater positive expression of subgraph B (ρ = −0.39, p = 0.04; uncorrected for multiple comparisons) during the high demand condition. These results suggest that a smaller change in the reaction time between the low demand condition and the high demand condition of the Stroop task is associated with: (i) greater anticorrelated dynamics within dorsal attention, default mode, frontoparietal, somatosensory, ventral attention, and visual systems during task activation, and (ii) greater correlated dynamics within and between dorsal attention, visual, and cerebellar systems during task control. For the Navon task, we find that a lower reaction time cost (better performance) is associated with: (i) greater negative expression of subgraph E (ρ = 0.55, p = 2.2 × 10−3; uncorrected for multiple comparisons) during the low demand condition, and (ii) greater negative expression of subgraph E (ρ = 0.47, p = 0.01; uncorrected for multiple comparisons) and subgraph J (ρ = 0.44, p = 0.02; uncorrected for multiple comparisons) during the high demand condition. These results suggest that a smaller change in the reaction time between the low demand condition and the high demand condition of the Navon task is associated with: (i) greater anticorrelated dynamics within default mode and frontoparietal systems, and between default mode and frontoparietal systems and other broadly distributed cognitive systems during task activation, and (ii) greater anticorrelated dynamics within default mode, frontoparietal, and visual systems, and between default mode, frontoparietal, and visual systems and other broadly distributed cognitive systems during task control. In sum, we find that changes in the correlated and anticorrelated BOLD dynamics within and between distributed cognitive systems is associated with cognitive processes during task activation and task control that explain inter-individual differences in performance during cognitive control tasks. Based on these data and our previous result that subgraphs maintain a consistent hierarchical organization in terms of their ranked expression between cognitive demand conditions, our findings suggest that individual variability in behavior during cognitive control may be marked by subtle individual differences in subgraph expression amid a hierarchical order that is defined at the population level.
Lastly, we ask whether there are individual brain regions that are more likely to participate in subgraphs associated with task performance. By quantifying the extent to which brain regions participate in subgraphs, we aim to link our analysis with classical univariate approaches for examining functional brain activation during cognitive control tasks. We hypothesize that brain regions commonly associated with executive and higher cognitive functions, such as frontoparietal, default mode, attention, and salience systems are more likely to participate in subgraphs that are associated with task performance. To test this hypothesis, we computed the performance participation score—a nodal measure linking the participation of a node in a subgraph with the relationship between the subgraph and behavioral performance. Specifically, we first compute node participation in a subgraph as the sum of the subgraph edge weights from a node to all other nodes—yielding one node participation score for each of the 262 brain regions in each of the twelve subgraphs [16]. We next compute the sum of a node’s participation scores, weighted by the Spearman’s ρ value between relative subgraph expression and performance cost: that is, nodes of the same subgraph were weighted by the same ρ value. Intuitively, a node with positive participation score tends to become disengaged in the brain network, via anti-correlated dynamics, during better task performance and a node with negative participation score tends to become engaged in the brain network, via correlated dynamics, during better task performance. To determine whether a brain region exhibits a greater participation score than expected by chance, we construct null distributions of regional participation scores by uniformly permuting the edges of each subgraph 10000 times and recomputing the regional participation score for each permutation. We retain regional participation scores that exceeded the 95% confidence interval of the null distribution after using Bonferroni correction for multiple comparisons testing.
Using this approach, we find a broad range of brain regions that are significantly involved in correlated and anti-correlated brain activity during improved task performance (Fig 6A–6D; right). For the Stroop task we observe that individuals exhibit lower reaction time cost when (i) during the low demand condition, regions within frontoparietal, default mode, subcortical, and visual systems are more disengaged from each other and regions within the cortical limbic system are more engaged with each other (Fig 6A; right), and (ii) during the high demand condition, regions within default mode and limbic systems are more disengaged from each other and regions within visual and somatosensory systems are more engaged with each other (Fig 6B; right). For the Navon task we observe that individuals exhibit lower reaction time cost when (i) during the low demand condition, regions within frontoparietal, default mode, and visual systems are more disengaged from each other and regions within the cortical limbic system are more engaged with each other (Fig 6C; right), and (ii) during the high demand condition, regions within frontoparietal, default mode, and visual systems are more disengaged with each other and regions within subcortical and limbic systems are more engaged with each other (Fig 6D; right).
Together, these results demonstrate that brain regions classically considered as key components of executive and higher cognitive functions, such as regions in frontoparietal and default mode systems, tend to be more influential in subgraphs that are associated with task performance. Notably, our approach characterizes the functional role that these brain areas play during task activation and task control based on their participation in correlated and anticorrelated BOLD dynamics. During different phases of cognitive control, regions involved in correlated dynamics may serve as integrators of task-relevant information while regions involved in anticorrelated dynamics may serve as segregators of task-relevant information.
In this work, we ask “What functional constraints shape internally-guided transitions in brain state during cognitive control?” To answer this question, we apply a powerful machine-learning approach referred to as non-negative matrix factorization, to dynamic functional brain networks measured during two cognitive control tasks—yielding subgraphs or clusters of temporally co-varying functional interactions between brain regions. We study the expression of these functional subgraphs during correlated and anticorrelated BOLD dynamics as subjects transition between different levels of task-induced cognitive demand. We show that the subgraphs differentiate clusters of functional interactions that are specific to the mechanics of the cognitive control tasks from those that are generalized to the network processes common to the cognitive control tasks. Specifically, we demonstrate for the first time clear evidence that functional subgraphs adaptively alter their expression depending on the type of cognitive control task and the amount of cognitive demand imposed on the system. Our results significantly extend our understanding of how objectively-defined clusters of functional interactions, beyond individual region-region co-activation, relate to transitions between cognitive states.
Our non-negative matrix factorization (NMF) approach enables us to objectively account for: (i) the dissociability of brain networks into composite subgraphs that are associated with specific cognitive control functions, and (ii) the flexible and adaptive expression of these putative cognitive sub-networks during fluctuations in cognitive demand. Intuitively, these subgraphs represent clusters of functional interactions whose weights tend to fluctuate together across tasks and across conditions. Unlike other graph partitioning techniques, such as community detection, that pursue a hard partitioning of network nodes into discrete clusters, NMF enables a soft partitioning of the high dimensional set of network edges into subgraphs that allow an edge to participate in multiple network sub-units [16, 22]. This capability is advantageous for examining how pairs of brain areas functionally interact within different topological contexts. Mathematically, NMF recovers a non-orthogonal spanning set of graph edges whose linear combination—weighted by dynamic expression coefficients—can reconstruct the original space of observed network topologies across the experimental task conditions. In other words, subgraphs represent a set of functional relationships for the cognitive control data from which they were recovered and subgraph expression coefficients represent the encoding of those relationships for the different task conditions (we refer the reader to [40] for a discussion on neural coding theory).
Thus from the perspective of network-based encoding of cognitive control tasks, indeed, we find that subgraphs are comprised of functional interactions that are either sensitive to the specific needs of a particular task or generalized to needs common across tasks. These data support the theory that there exist separate task-specific and task-general network architectures [30]. We examine the particular cognitive systems involved in task-specific and task-general subgraphs and find a dual-role for correlated and anticorrelated interactions between traditional cognitive control systems and the default mode system: these systems are positively expressed during cognitive control involving Stroop-based response inhibition and negatively expressed during cognitive control involving Navon-based task switching. Our finding of anticorrelated interactions between cognitive control and default mode systems is well supported by the popular theory that the task-negative, default mode system deactivates as task-positive, executive areas activate [41–43]. On the other hand, our finding of correlated interactions between these systems challenges the notion that these systems must decouple during cognitive control. Prior studies have in fact demonstrated that individuals that exhibit greater integration between the default mode network and executive areas tend to display better behavioral performance during cognitive control tasks that involve switching between task-rules [28, 44]. Based on these results, we posit that differences in the nature of functional interactions between these systems might be explained by task-specific requirements for cognitive control. Importantly, NMF demonstrates the ability to tease apart functional interactions underlying intrinsic differences in cognitive control processes by recovering task-specific subgraphs.
There is a longstanding question in network neuroscience regarding the putative roles of task-specific functional architectures and their relationship to intrinsic functional networks that generalize across tasks [30]. The canonical model posits that task-general processes shape intrinsic functional networks and task-specific processes update subsets of these intrinsic functional connections [30]. A critical assumption has been that networks related to task-specific processes also facilitate behavioral performance of the task. In this study, we present data that support the canonical model yet challenge the assumption that task processes and behavioral metrics of performance on the task stem from the same network structures. First, we find that a robust hierarchy of subgraphs persist between different forms of cognitive control processes (Fig 4A) and different levels of cognitive demand (Fig 5A and 5B). Indeed, changes in subgraph expression within the bounds of this hierarchy accompany specific task states, however we also find that the subgraphs that best predict individual differences in behavior are not necessarily those that are modulated by different task conditions. In other words, functional architectures most strongly associated with behavior may represent task-general cognitive functions that are distinct from networks that are differentially expressed between cognitive conditions, consistently across individuals. For example, the Stroop task is designed to recruit general processes related to stimulus perception and color-word discrimination as well as cognitive control processes such as inhibition of the prepotent response to an incongruent stimulus and Navon task is designed to recruit general processes such as perceptual decision making or specific processes such as decision making based on rules that periodically switch. A task-general subgraph that is modulated by lower level perceptual or cognitive processes during low and high task conditions may still be modulated differently across individuals and reflect differences in behavior. Conversely, based on recent work demonstrating that different components of functional brain networks may be highly similar or highly dissimilar across individuals [45], a task-general subgraph that does not strongly vary with individual differences in behavior might reflect intermediate task processes that are common to the low and high cognitive demand conditions. Indeed, a future study that uses NMF in conjunction with faster imaging modalities may amenably tease apart subgraphs involved with different temporal phases of cognitive control processes.
A growing body of literature in network neuroscience has shown that the brain possesses an ability to maintain a homeostasis of its own internal dynamics through antagonistic, push-pull interactions in various areas of healthy cognition [28, 29, 46] and disease [47]. Simply, push-pull control strategies may prevent imbalances of activity in complex, interconnected systems like the brain [48, 49]. A push-pull mechanism would be a critical component of cognitive control in which brain networks must perform two antagonistic functions: (i) segregated information processing in functionally-specific domains, and (ii) integrated information processing to adapt to environmentally-driven changes in cognitive demand [26]. Our results pertaining to the adaptive shifting in subgraph expression during changes in cognitive demand may be associated with a putative push-pull control mechanism in which: (i) subgraphs first establish a consistent hierarchy of expression that enforces a baseline level of expression that remains consistent relative to other subgraphs during cognitive control processes, and (ii) subgraphs then shift their expression above or below their baseline—via changes in correlated or anticorrelated BOLD dynamics—depending on cognitive demand. We posit that a push-pull mechanism might internally regulate the direction of change in subgraph expression, collectively across the network: an excessive increase or decrease in subgraph expression might disrupt the hierarchical order of subgraph expression and lead to brain states in which information is overly integrated or overly segregated across the network. In our analysis, we observe that the shift between cognitively demanding brain states involves a change in the interacting roles between brain areas distributed across several cognitive systems: including frontoparietal, default mode, attentional, and cerebellar regions. Recent studies focusing on functional interactions between cerebellum and traditional cognitive control regions [12] have suggested that the cerebellum may subserve cognitive processes related to error correction [50, 51]. Our results add new insight to this discussion by demonstrating in two different cognitive control tasks that frontoparietal, cerebellar, and sensory systems are involved in subgraphs that significantly change in expression with increasing cognitive demand.
We also consider the possibility that regulatory mechanisms involved in cognitive control might also explain differences in individual performance on cognitively demanding tasks. We found that subgraphs may be heterogeneously associated with individual cognitive performance: greater correlated BOLD dynamics and greater anticorrelated BOLD dynamics between regions of subgraphs are associated with improved task performance. These data suggest that cognitive control is associated with enhanced integration and segregation of task-relevant information between different composite sets of brain regions. In addition, we use functional subgraphs to uncover the relationship between functional interactions and sub-processes of cognitive control that differentially contribute to the performance cost associated with an increase in cognitive demand. Namely, we find subgraphs whose expression during task activation is associated with lower performance cost accompanying an increase in cognitive demand, and we find subgraphs whose expression during task control is associated with lower performance cost accompanying an increase in cognitive demand. We speculate that the rich distribution of performance modes exhibited by functional subgraphs implicates a network homeostasis on cognitive control processes [46].
Critically, we contextualize the relationship between network reorganization during task states and its relationship with task performance via the following sequence of events. First, global network correlations decrease between the fixation period and the task. As the network becomes less correlated, select subgraphs become increasingly specialized and exhibit heightened levels of expression relative to non-task related subgraphs. These task-related subgraphs remain highly expressed across individuals and inter-individual differences in expression scale with task performance. Brain regions with greatest levels of participation within task-related subgraphs are putative mediators of the relationship between subgraph expression and performance.
In sum, we demonstrate that functional brain networks capably adapt their topological architecture in response to task-driven modulation in cognitive demand. Critically, we observe that cognitive control may not necessarily activate discrete cognitive brain systems, but rather recruit several interconnected systems, in concert, between changes in cognitively demanding brain states. When individuals under- or over-express functional interactions between these cognitive systems they tend to respond more slowly during difficult cognitive tasks, implicating specific brain sub-networks in facilitating or impeding an individual’s ability to transition between states.
While we narratively describe cognitive control to be recruited continuously in response to task demands, it is also important to acknowledge that cognitive control functions can be considered to be distinct processes [52] with partially dissociable substrates [53]. Given these broader debates about shared and unique CC mechanisms, we should continue to consider the spatiotemporal signatures of brain activity that distinguish between accounts of CC. Future studies could use NMF-based subgraph analysis to dissect networks involved in tasks where demand is parametrically varied and test whether a continuous or discrete representation of specific CC functions better describes observed network dynamics.
Lastly, we focus on the mechanistic role that functional brain networks play in regulating internal dynamics during cognitive control. Our novel approach and findings open new doors for querying how such regulatory mechanisms could be modulated to influence behavior. For instance, can we perturb specific network components to improve the likelihood that an individual is able to access shorter trajectories to switch between low demanding states and high demanding states? By marrying machine-learning approaches that objectively tease apart concurrent network processes attributed to different facets of cognition with burgeoning neurotechnologies such as neurofeedback [54], neurostimulation [55], or pharmacological intervention [56–58] that can exogenously control network dynamics, we can explore how disrupting network components that exhibit task-based adaptation causally influence behavior. The prospect of such scientific inquiry is equally exciting in diseases such as schizophrenia in which patients experience more probable transitions to more disruptive cognitive states.
All participants completed a Stroop task with color-word pairings that were eligible and ineligible to elicit interference effects [24], and a local-global perception task based on classical Navon figures [25]. For the Stroop task, trials were comprised of words presented one at a time at the center of the screen printed in one of four colors—red, green, yellow, or blue -– on a gray background. For all trials, subjects responded using their right hand with a four-button response box. All subjects were trained on the task outside the scanner until proficient at reporting responses using a fixed mapping between the color and button presses (i.e., index finger = “red”, middle finger = “green”, ring finger = “yellow”, pinky finger = “blue”). Trials were presented in randomly intermixed blocks containing trials that were either eligible or ineligible to produce color-word interference effects. In the scanner, blocks were administered with 20 trials apiece separated by 20 s fixation periods with a black crosshair at the center of the screen. Each trial was presented for a fixed duration of 1900 ms separated by an interstimulus interval of 100 ms during which a gray screen was presented. In the trials ineligible for interference, the words were selected to not conflict with printed colors (“far,” “horse,” “deal,” and “plenty”). In the trials eligible for interference (i.e., those designed to elicit the classic Stroop effect [24]), the words were selected to introduce conflict (i.e., printed words were “red,” “green,” “yellow,” and “blue” and always printed in an incongruent color). In our analysis, we refer to blocks that are eligible (ineligible) to produce color-word interference effects as high demand (low demand) conditions (Fig 1B).
For the Navon task, local-global stimuli were comprised of four shapes—a circle, X, triangle, or square—that were used to build the global and local aspects of the stimuli. On all trials, the local feature did not match the global feature, ensuring that subjects could not use information about one scale to infer information about another. Stimuli were presented on a black background in a block design with three blocks. In the first block type, subjects viewed white local-global stimuli. In the second block type, subjects viewed green local-global stimuli. In the third block type, stimuli switched between white and green across trials uniformly at random with the constraint that 70% of trials included a switch in each block. In all blocks, subjects were instructed to report only the local features of the stimuli if the stimulus was white, and to report only the global feature of the stimuli if the stimulus was green. Blocks were administered in a random order. Subjects responded using their right hand with a four-button response box. All subjects were trained on the task outside the scanner until proficient at reporting responses using a fixed mapping between the shape and the button presses (i.e., index finger = “circle”, middle finger = “X”, ring finger = “triangle”, and pinky finger = “square”). In the scanner, blocks were administered with 20 trials apiece separated by 20 s fixation periods with a white crosshair at the center of the screen. Each trial was presented for a fixed duration of 1900 ms separated by an interstimulus interval of 100 ms during which a black screen was presented. In our analysis, we refer to blocks that switch between local-global perception as the high demand condition and blocks that do not switch as the low demand condition (Fig 1C).
We acquired T1-weighted anatomical scans on a Siemens 3.0T Tim Trio for all subjects. Anatomical scans were segmented using FreeSurfer [59] and parcellated using the connectome mapping toolkit [31] into N = 234 cortical and subcortical brain regions. We also included a cerebellar parcellation (N = 28 brain regions [32]) by using FSL to nonlinearly register the individual’s T1 to MNI space. Then, we used the inverse warp parameters to warp the cerebellum atlas to the individual T1. Finally, we merged the cerebellar label image with the dilated cortical and subcortical parcellation image resulting in N = 262 brain regions.
Functional magnetic resonance imaging data was acquired on a 3.0T Siemens Tim Trio whole-body scanner with a whole-head elliptical coil by means of a single-shot gradient-echo T2* (TR = 1500 ms; TE = 30 ms; flip angle = 60 degrees; FOV = 19.2 cm, resolution 3mm x 3mm x 3mm). Preprocessing was performed using FEAT v. 6.0 (fMRI Expert Analysis Tool) a component of the FSL software package [60]. To prepare the functional images for analyses, we completed the following steps: skull-stripping with BET to remove non-brain material, motion correction with MCFLIRT (FMRIB’s Linear Image Registration Tool; [60]), slice timing correction (interleaved), spatial smoothing with a 6-mm 3D Gaussian kernel, and high pass temporal filtering to reduce low frequency artifacts. We also performed EPI unwarping with fieldmaps to improve subject registration to standard space. Native image transformation to a standard template was completed using FSL’s affine registration tool, FLIRT [60]. Subject-specific functional images were co-registered to their corresponding high-resolution anatomical images via a Boundary Based Registration technique (BBR [61]) and were then registered to the standard MNI-152 structural template via a 12-parameter linear transformation. Finally, each participant’s individual anatomical image was segmented into grey matter, white matter, and CSF using the binary segmentation function of FAST v. 4.0 (FMRIB’s Automated Segmentation Tool [62]). The white matter and CSF masks for each participant were then transformed to native functional space and the average timeseries were extracted. Based on the commonly accepted notion that smoothing reduces scan-related, spatially-distributed Gaussian noise across voxels and enhances BOLD signal-to-noise ratio, we conducted smoothing by applying a kernel with full-width half-maximum of 6 mm to voxels prior to ROI time series extraction. An important consideration of smoothing is that voxels at the edge of an ROI may contain overlapping information from adjacent ROIs. However, our analysis occurs at the level of the aggregate BOLD activity across many voxels in an ROI, and thus voxel-level precision was not a goal in this study. The white matter and CSF signals were used as confound regressors on the time series along with 18 translation and rotation parameters as estimated by MCFLIRT [63]. To preserve natural anti-correlation in the BOLD signal, we did not regress the global signal [64].
We refer the reader to [65] for additional methodological details regarding data acquisition and pre-processing.
We constructed functional brain networks to study the functional interactions between brain regions during the Stroop and Navon cognitive control tasks. To measure functional interactions, we first separately divided the BOLD signal into six low demand blocks, six high demand blocks, and twelve fixation blocks (before each cognitive demand block) for each behavioral task of each subject. Each block contained 20 samples or 30 seconds of signals (Fig 2B). We next computed a Pearson correlation coefficient between each pair of BOLD signals from the N brain regions (graph nodes) in each of the K experimental blocks. We then aggregated correlations (graph edges) into an N × N × K adjacency matrix A for each subject. We note that due to confounding delays in hemodynamic response, it is possible that fixation blocks contain both task-related and task-unrelated activity. To mitigate this concern, we take two steps. First, we align each block with the peak hemodynamic response by shifting analysis windows by 4 TRs, which corresponds to the canonical hemodynamic lag of 6 seconds. Second, we compute NMF-based subgraphs (see next section) using fixation blocks to increase the length of the physiologic signal, but we restrict our analysis of the subgraphs specifically to task blocks.
To analyze positively correlated (correlated) and negatively correlated (anticorrelated) functional interactions, we separated positively-weighted edges from negatively-weighted edges for each block k in A using a threshold of zero. This procedure resulted in a thresholded adjacency matrix A* of size N × N × 2 × K where each block k is associated with one N × N matrix with positive edge weights and another N × N matrix with negative edge weights (Fig 2C). We retain all correlation values after the thresholding procedure such that both positive adjacency matrices and negative adjacency matrices are both fully-weighted graphs.
An alternate representation of the adjacency matrix A* is a two-dimensional network configuration matrix A ^ *, which tabulates all N × N pairwise edge weights across K blocks, and across positive and negative edge types (Fig 2D). Due to symmetry of A k *, we unravel the upper triangle of A k *, resulting in the weights of N(N − 1)/2 connections. Thus, A ^ * has dimensions N(N − 1)/2 × 2*K.
To identify network subgraphs—sets of network edges whose strengths co-vary over experimental task conditions—we applied an unsupervised machine learning algorithm called non-negative matrix factorization (NMF) [21] to the network configuration matrix. This technique enabled us to pursue a parts-based decomposition of the network configuration matrix into subgraphs with expression coefficients that vary with time (Fig 2E and 2F). Briefly, NMF holds two distinct advantages to principal components analysis (PCA) and independent components analysis (ICA) for studying components of interconnected network structures. First, PCA/ICA quantify subgraphs that are statistically orthogonal/independent from each other, while NMF quantifies subgraphs that are statistically redundant such that they can flexibly co-occur with other subgraphs during different brain states. The unique property of NMF to characterize overlapping network structures is conceptually valuable for the analysis of brain graphs, which assume that each node encompasses statistical relationships with all other nodes in the network—this assumption is violated by PCA/ICA. Second, PCA/ICA arbitrarily assign positive and negative weights to subgraphs, while NMF enforces non-negative weights to subgraphs. The non-negative property of NMF uniquely quantifies subgraphs that are additive parts of the network and interpretable on the basis of their positive contribution to the functional network at each point in time—this interpretation is obfuscated by PCA/ICA. For further, in-depth discussion regarding network subgraphs, we refer the reader to [16]. For recent applications of NMF to the study of functional brain networks, please see [22, 23, 66, 67].
To apply NMF to functional networks, we first computed the magnitude of the network configuration matrix A ^ * such that all entries of the matrix were non-negative. We next applied two normalization procedures to account for differences in the magnitude of edge weights between positive correlations and negative correlations and between study participants. First, based on the finding that mean negative correlations are significantly lower in magnitude than mean positive correlations across subjects (paired t-test; t27 = 20.0, p = 9.7 × 10−18; S1E Fig), we sought to normalize the distribution of positive edge weights and negative edge weights for each observed graph (each row of the configuration matrix). Therefore, we divided the edge weights in each row of the configuration matrix by their sum such that the weight edge density for each observed graph was equal to one. Second, based on the finding that the distribution of edge weights differs between subjects (one-way ANOVA; F = 2.5, p = 3.5 × 10−5; S1A Fig), we sought to standardize the vector of weights associated with each edge (each column of the configuration matrix), separately, for each subject. Therefore, we scaled the weights of each edge by their Euclidean length (L2-norm), separately, for each subject [68]. We also note that BOLD autocorrelation was not removed from the measured edge weights. As NMF is a linear operation and based on a recent study showing that the edge weights before removing BOLD autocorrelation are linearly correlated with edge weights after removing BOLD autocorrelation [69], we did not expect this procedure to influence NMF analysis.
We next formulated the matrix factorization problem A ^ * ≈ W H s.t.W > = 0 , H > = 0 as the decomposition of the network configuration matrix A ^ * into two non-negative matrices W—the weighted subgraph matrix consisting of recurring patterns of functional interactions, or network edges—and H—the dynamic expression matrix consisting of coefficients reflecting the weight of a subgraph during different task conditions of each subject [16]. To quantify W and H, we optimized the following cost function:
min W , H1 2 ‖ A ^ − W H ‖ F 2 + α ‖ W ‖ F 2 + β ∑ t = 1 T ‖ H ( : , t ) ‖ 1 2 , (1)
where m ∈ [2, min(N(N − 1)/2, T) − 1] is the number of subgraphs to decompose, β is a penalty weight to impose sparse temporal expression coefficients, and α is a regularization of the interaction strengths for subgraphs [70]. To solve the NMF equation, we used an alternating non-negative least squares with block-pivoting method with 100 iterations for fast and efficient factorization of large matrices [71]. We initialized W and H with non-negative weights drawn from a uniform random distribution on the interval [0, 1].
To select the parameters m, β, and α, we pursued a random sampling scheme—shown to be effective in optimizing high-dimensional parameter spaces [16, 72]—in which we re-ran the NMF algorithm for 1000 parameter sets in which m is drawn from U ( 3 , 50 ), β is drawn from U ( 0 . 01 , 5 ), and α is drawn from U ( 0 . 01 , 5 ) (S2 Fig). We evaluated subgraph learning performance based on a four-fold cross-validation scheme in which the twenty eight subjects are uniformly partitioned into folds of seven subjects and, iteratively, three folds are used to identify subgraphs and the held-out fold is used to compute the cross-validation error (‖ A ^ − W H ‖ F 2). The optimal parameter set should yield subgraphs that minimize the cross-validation error and reliably span the space of observed network topologies [16]. Based on these criteria, we identified an optimum parameter set ( m ¯ , β ¯ , α ¯ ) that exhibited a low residual error in the bottom 5th percentile of our random sampling scheme (S2G and S2I Fig).
Due to the non-deterministic nature of this approach, we integrated subgraph estimates over multiple runs of the algorithm using consensus clustering—a general method of testing robustness and stability of clusters over many runs of one or more non-deterministic clustering algorithms [73]. Our adapted consensus clustering procedure entailed the following steps: (i) run the NMF algorithm R times per network configuration matrix, (ii) concatenate subgraph matrix W across R runs into an aggregate matrix with dimensions E × ( R * m ¯ ), and (iii) apply NMF to the aggregate matrix to determine a final set of subgraphs Wconsensus and expression coefficients Hconsensus (we refer the reader to [16] for more details). In this study, we set R = 1000.
To investigate putative core-periphery organization in each functional subgraph, we quantify the core-periphery index as a measure of the balance between mean edge strength within each cognitive system and mean edge strength of each cognitive system to all other cognitive systems. Specifically, we define the core-periphery index for a symmetric, subgraph adjacency matrix W* with dimensions N × N using the following equations:
core s = 1 | s | ∑ i j [ W i j * ] δ ( s i , s j ) (2) periphery s = 1 | s | * ( N − | s | ) ∑ i j [ W i j * ] ( 1 − δ ( s i , s j ) ) (3) core-periphery = 1 9 ∑ s = 1 9 c o r e s − p e r i p h e r y s c o r e s + p e r i p h e r y s (4)
where N is the number of network regions, s is one of nine cognitive systems, |s| is the number of nodes in cognitive system s, si, sj refer to the cognitive system assignments of nodes i and j, and δ(si, sj) = 1 if si = sj and δ(si, sj) = 0 if si ≠ sj. Intuitively, the core-periphery index is bounded between −1 and +1, where positive values indicate greater subgraph edge strength within a cognitive system, indicating that the subgraph reflects functional interactions within a network core, negative values indicate greater subgraph edge strength between cognitive systems, indicating that the subgraph reflects functional interactions within a network periphery, values approaching zero imply that a subgraph reflects balanced functional interactions between the network core and the network periphery. To examine the specific cognitive systems that participate in core-periphery organization of each subgraph, we first generate 10000 surrogates of each subgraph by randomly permuting subgraph edges to disrupt system-level architecture. We compute the core score and the periphery score for each cognitive system s of each of the surrogates, separately for each subgraph. Using Bonferroni correction for multiple comparisons testing, we identify subgraph-specific cognitive systems that exhibit significantly greater core and periphery scores than expected by the surrogate model S6 Fig.
It is important to consider the reproducibility of subgraphs measured using NMF given different data splits. To quantify the reproducibility of functional subgraphs, we measured the extent to which the pattern of subgraph edge weights measured in one dataset predicts the pattern of subgraph edge weights measured in a second dataset. Specifically, we first divided the whole cognitive control dataset into two datasets such that the first dataset contains the first three experimental blocks across subjects and the second dataset contains the second three experimental blocks across subjects. We next applied NMF using the optimal parameter set to the two datasets (A ^ 1 corresponds to the network configuration matrix of the first dataset and A ^ 2 corresponds to the network configuration matrix of the second dataset), resulting in two subgraph matrices (W1 and W2). Note that the subgraphs along the columns of W1 may not necessarily be ordered similarly as the subgraphs along the columns of W2 due to the stochastic nature of the NMF algorithm. To reorder subgraphs from the second dataset such that they correspond to the same order as subgraphs from the first dataset, we sought a mapping Xi,j of subgraph W 1 i to subgraph W 2 j, where X is a Boolean matrix that prescribes whether the ith subgraph from the first dataset is uniquely assigned to the jth subgraph from the second dataset. The cost Ci,j associated with assigning W 1 i to W 2 j is equal to ‖ W 1 i − W 2 j ‖. To determine a unique X, we minimized the cost function ∑i∑j Ci, j Xi, j using the well-known Hungarian algorithm [74]. After calculating an optimal assignment between subgraphs of the two datasets, we measured the similarity in the pattern of edge weights between assigned subgraph pairs (i, j) by computing the Pearson correlation coefficient. We compared the true Pearson correlation coefficient of every subgraph pair to a null distribution in which we re-computed the Pearson correlation coefficient between every possible, non-assigned subgraph pair. This approach enabled us to assess the reproducibility of each individual subgraph based on the magnitude of the Pearson correlation similarity measure relative to that expected by chance.
|
10.1371/journal.pntd.0006817 | Clinical manifestations of dengue in relation to dengue serotype and genotype in Malaysia: A retrospective observational study | Malaysia experienced an unprecedented dengue outbreak from the year 2014 to 2016 that resulted in an enormous increase in the number of cases and mortality as compared to previous years. The causes that attribute to a dengue outbreak can be multifactorial. Viral factors, such as dengue serotype and genotype, are the components of interest in this study. Although only a small number of studies investigated the association between the serotype of dengue virus and clinical manifestations, none of these studies included analyses on dengue genotypes. The present study aims to investigate dengue serotype and genotype-specific clinical characteristics among dengue fever and severe dengue cases from two Malaysian tertiary hospitals between 2014 and mid-2017.
A total of 120 retrospective dengue serum specimens were subjected to serotyping and genotyping by Taqman Real-Time RT-PCR, sequencing and phylogenetic analysis. Subsequently, the dengue serotype and genotype data were statistically analyzed for 101 of 120 corresponding patients’ clinical manifestations to generate a descriptive relation between the genetic components and clinical outcomes of dengue infected patients. During the study period, predominant dengue serotype and genotype were found to be DENV 1 genotype I. Additionally, non-severe clinical manifestations were commonly observed in patients infected with DENV 1 and DENV 3. Meanwhile, patients with DENV 2 infection showed significant warning signs and developed severe dengue (p = 0.007). Cases infected with DENV 2 were also commonly presented with persistent vomiting (p = 0.010), epigastric pain (p = 0.018), plasma leakage (p = 0.004) and shock (p = 0.038). Moreover, myalgia and arthralgia were highly prevalent among DENV 3 infection (p = 0.015; p = 0.014). The comparison of genotype-specific clinical manifestations showed that DENV 2 Cosmopolitan was significantly common among severe dengue patients. An association was also found between genotype I of DENV 3 and myalgia. In a similar vein, genotype III of DENV 3 was significantly common among patients with arthralgia.
The current data contended that different dengue serotype and genotype had caused distinct clinical characteristics in infected patients.
| The study highlights interesting relationship between viral factors and clinical manifestation of dengue disease during an outbreak. The viral factors which include serotype and genotype of dengue virus were studied to discover if the clinical manifestation in patients were serotype and genotype-specific. As most clinical symptoms of severe dengue infection only manifest at a much later stage of dengue infection, therefore, information on serotype or genotype-specific dengue manifestations may serve as early surrogate markers to predict disease progression. We found that specific clinical manifestations were over-represented by a specific DENV serotype and genotype. Severe dengue was significantly present in DENV 2 Cosmopolitan-infected group while non-severe dengue was prominent among DENV 1 genotype I-infected patients. DENV 3-infected patients commonly manifested musculoskeletal symptoms. This study was undertaken between 2014 and mid-2017, which was considered to be crucial, given that Malaysia experienced an unprecedented outbreak of dengue during this period. Consequently, the occurrence of serotype shift and possible re-emergence of DENV 3 genotype I were reported.
| Since the 1950’s, dengue has become a serious health problem in the South-East Asia region. In 1902, Malaysia experienced its first case of dengue [1]. Since then, Malaysia has increasingly become popular for perpetual dengue endemic issues, resulting from the continuous rise in reported dengue infection cases. The country experienced major outbreaks in 1974, 1978, 1982, 1990. Notably, Malaysia recorded the highest number of dengue cases between 2014 and 2017. In 2014, a total of 108,698 cases were reported in Malaysia which was equivalent to an incidence rate (IR) of 361.1 cases in 100,000 populations with 215 mortalities. This alarming figure had outnumbered the previous recorded dengue cases in 2013 by 150.8%. Meanwhile, there was a noticeable increase in 2015 with 120,836 cases (IR = 396.4), along with 336 mortalities. In 2016, the total number of dengue cases declined to around 101,357, although, the mortality rate was 10% higher as compared to that in 2014. [2, 3]. Finally in 2017, the dengue situation in Malaysia came under control as the total number of cases continued to drop. However, the number was still higher as compared to that in 2013. Subsequently, this prompted the Malaysian government to implement a program to eradicate Aedes aegypti mosquito breeding sites. Recently, a National Dengue Plan (2015–2020) was implemented by the Malaysian government to intensify the readiness and response capacity in detecting dengue cases and outbreaks, requiring immediate action and attention.
Several risk factors at various intensities have contributed towards the severity of dengue infection during the course of an outbreak. Viral factors are often considered as one of the risk factors and components of interest in many studies in this field. A shift in the distribution of dengue serotypes and genotypes may contribute to the accelerating number of dengue cases due to an antibody-dependent enhancement (ADE) effect [4]. Interestingly, new genotype clades were discovered in some countries such as India and Sri Lanka during dengue outbreaks [5, 6]. While many studies investigated the association between certain serotype of dengue virus and disease severity, only a few studies provided comparative details of the clinical manifestations among dengue serotypes and genotypes [7, 8, 9]. This comparison is particularly important to further aid in the early prediction of a patient’s condition based on the clinical characteristics and information on the serotype and genotype of the dengue virus infecting the patient.
With advanced laboratory tests, the serotyping of the dengue virus from an infected patient can be performed even on the first day of fever. Most clinical symptoms of severe dengue infection only manifest at a later stage of dengue infection. Therefore, information on serotype or genotype-specific dengue manifestations may serve as early surrogate markers to predict disease progression. Furthermore, specific clinical manifestations may be over-represented in patients infected with certain DENV serotype and genotype. In consideration of the above discussion, this study aims to investigate the clinical manifestations of dengue patients in relation to dengue serotype and genotype during a dengue outbreak period in Malaysia.
This study has obtained an ethical approval from the Medical Research & Ethics Committee, Ministry of Health Malaysia (Reference number: NMRR-15-923-25233). All patients’ data were totally anonymous and requested from clinicians involved in this study.
This is a retrospective observational study performed with a total of 120 dengue serum specimens obtained from two tertiary hospitals and a research institute in Malaysia. The serum specimens were acquired from patients who were primarily admitted for dengue fever and confirmed for dengue infection at Hospital Serdang (n = 94) and Hospital Ampang (n = 24) from October 2014—May 2017. Two more sera from DENV 4-infected patients were obtained from the Virology Unit, at the Institute for Medical Research (IMR), in Kuala Lumpur. The inclusion criteria at the time of sample collection included samples that were positive for dengue NS1 antigen whereas exclusion criteria included suspicion for dengue but were negative for NS1 and identified as other febrile illnesses. The dengue confirmation criteria incorporated the results from the NS1 antigen rapid test, with or without Dengue IgM/IgG rapid combo tests performed by the hospitals. These tests were typically performed as soon as dengue infection was suspected in a patient. If dengue IgM was negative before day seven of the onset of fever, a repeat sample was taken at the recovery phase. The IgM and IgG rapid tests results were used to classify the cases as primary and secondary infection. Dengue cases that were positive for IgM but negative for IgG were regarded as primary infection whereas cases that were positive for either IgG only or both IgM and IgG were classified as secondary infection. The commercially available diagnostic kits incorporate an immunochromatography-based technique manufactured by Pan Bio (Brisbane, Australia). The NS1 antigen rapid test was interpreted by a single band targeting the dengue NS1 antigen. The IgM/IgG Dengue Duo Cassette highlights the presence of anti-dengue IgM and IgG antibodies in their specific bands. The results of the tests were displayed as reactive or non-reactive without titration. The reactive IgG result was semi-quantitative, showing the presence of antibodies in the serum which was equivalent to HAI titer of 1:2,560, indicating a secondary dengue infection.
Dengue fever was diagnosed and defined according to the World Health Organization (WHO) 2009 dengue classification and severity level. The classification was used to indicate dengue, dengue with warning sign and severe dengue. Severe dengue includes severe plasma leakage, severe haemorrhage and severe organ dysfunction. To make the clinical diagnosis and determine the severity of the dengue infection, a medical officer performed physical examination on the patient, after which the findings were keyed in the e-file and request was made for necessary laboratory tests to be performed. Following this, a team of experienced clinicians including consultants or specialists further verified the accuracy of the clinical diagnosis during daily clinical ward round based on patients’ progress in their symptoms, clinical findings as well as the latest laboratory test report. The implication of the observation was to determine whether the patient requires intensive unit (ICU) care or normal ward management.
Dengue serotyping and phylogenetic analysis were performed for all 120 specimens. Socio-demographic, clinical profiles and laboratory data were obtained from the patient's record from the respective hospitals. The information was later analyzed with the dengue serotyping and genotyping results performed in this study. Notably, in the sta-tistical analysis, some samples were excluded due to incomplete clinical information, the presence of co-infection with leptospirosis and small sample size for a particular serotype. The rationale for excluding the samples of patients co-infected with leptospirosis was due to the overlapping clinical features of the patients infected with leptospirosis and dengue. These samples, however, were not tested for other closely related co-infections such as Zika or Chikungunya due to budgetary constraints. Moreover, the information on the history and clinical presentation of the patients were unlikely to be of these diseases.
QIAamp Viral RNA Mini Kit (Qiagen) was used to perform the extraction of dengue viral ribonucleic acid (RNA) from the serum specimens according to the manufacturer’s instructions. The eluted 50 μl viral RNA was used as a template in the PCR assays.
Dengue virus serotyping was carried out in a fourplex Taqman Real-Time RT-PCR detection platform as described by Johnson et al. (2005) [10]. PCR reactions were prepared in a cocktail of 12.5 μl of 2X RT (reverse transcriptase)-PCR Mix (i-Script One Step RT-PCR kit, Biorad, USA), 0.5 μl of each primers (DENV 1 and DENV 3 primers: 50 μM; DENV 2 and DENV 4 primers: 25 μM), 0.45 μl of each probes (10μM), 0.5 μl of RT Enzyme Mix, and 1.2 μl of nuclease-free water. Positive controls for each serotype comprised of RNA from previously confirmed dengue patients, and obtained from the Virology Unit, IMR. The negative control consisted of reactions without an RNA template, which was substituted with 5μl of nuclease-free water. Taqman Real-Time RT-PCR amplification was performed on the CFX 96 (Biorad, USA) platform at 50°C for 10min, 95°C for 5min, followed by 45 cycles of 95°C for 15 sec and 60°C for 30 sec. The aforementioned PCR mix and cycling conditions were optimized by the Virology Unit. io: dx.doi.org/10.17504/protocols.io.rabd2an.[PROTOCOL DOI]
The partial E gene of dengue virus was amplified before sequencing by using four sets of serotype-specific oligonucleotides [11]. All amplification reactions were carried out in a 96-well conventional Thermal Cycler (Bio Rad, USA). The PCR was undertaken at 50°C for 30 min, 94°C for 2 min and 45 cycles of (94°C for 15 sec, 50°C for 30 sec and 68°C for 1 min) followed by an extension reaction at 68°C for 5 min. A 25 μl aliquot of each PCR reaction was analyzed on 1.5% pre-stained agarose by gel electrophoresis and viewed under UV illumination. The corresponding amplicons were extracted from the agarose gel and purified by a Gel Extraction Kit (Qiagen, USA) according to the manufacturer’s instruction. The final elution contained 30 μl of purified PCR amplicons whereby 5 μl of these were reanalyzed on 1.5% agarose gel to substantiate the accuracy of purification step. The purified PCR amplicons were outsourced for Sanger Sequencing (1st Base, IDT, Singapore). io:dx.doi.org/10.17504/protocols.io.racd2aw.[PROTOCOL DOI]
The sense and antisense sequences obtained by sequencing were aligned to produce a consensus partial E gene sequence by using CLUSTAL Omega software (https://www.ebi.ac.uk/Tools/msa/clustalo/). Reference sequences of E gene for each serotype were extracted from the GenBank database from various geographical regions. Phylogenetic trees were constructed with Mega 7 software adopting neighbor-joining method (bootstrap replication 1000x) for all four serotypes to determine the genotypes of dengue virus isolates.
Statistical package for the social sciences (SPSS) version 21.0 was adopted to analyze data collected from the dengue patients. Categorical variables were expressed as frequencies (percentages). Chi-square or Fisher’s Exact test was performed to analyze the significance of the categorical variables. Continuous variables were tested for normality with the Komolgorov-Smirnov test. Non-parametric analysis by Kruskal-Wallis was employed for data with non-normal distribution and presented in median and interquartile range (IQR). The normally distributed data were analyzed by One-Way ANOVA and expressed as mean and standard deviation (SD). The patients’ data were tabulated according to categorical variables including gender, serotypes, genotypes and a spectrum of clinical manifestations. Meanwhile, continuous variables refer to parameters such as age, day of fever and laboratory test results. These data were utilized to determine the distribution of dengue serotype and genotype in the study population and describe the relation with demography, clinical manifestations and laboratory parameters. The analyses were performed at 95% confidence with level of significance of p<0.05.
Serotyping results (Fig 1) from 120 study subjects revealed that more than half of the study population were infected with DENV 1 (64/120; 53.0%) followed by DENV 2 (31/120; 26.0%), DENV 3 (20/120; 17.0%), DENV 4 (4/120; 3.0%) and mixed serotype DENV 1/ DENV 2 (1/120; 1.0%). The distribution of these serotypes by year of infection is shown in Fig 2. Among our study subjects, a domination of DENV 1 was seen from the year 2014–2016 with prevalence of 35.3% (6/17), 63.2% (36/57) and 58.6% (17/29) each year, respectively. In 2017, DENV 2 was more frequently observed than other serotypes (7/17; 41.2%) among the study subjects. The number of DENV 3-infected cases from 2014–2017 were less than DENV 1 and DENV 2 except in the year 2014 with prevalence of 29.4% (5/17), 19.4% (7/57), 13.8% (4/9) and 23.5% (4/17), respectively. DENV 4-infected cases were observed more in 2014 (3/17; 17.6%) and one case in 2017 (1/17; 5.9%) while one DENV 1/ DENV 2 mix serotype case was found in 2015 (1/17; 5.9%). Further, phylogenetic analysis classified these dengue strains into genotypes (Figs 3–6). All DENV 1 and DENV 2 strains were classified under genotype I and cosmopolitan genotype, respectively. DENV 3 strains clustered into two distinct genotypes. Genotype III comprised most of the DENV 3 strains as compared to genotype I. The DENV 4 strains from the study belong to genotype I and genotype II. In addition, one dengue strain with mixed serotypes of DENV 1/DENV 2 were identified. However, only DENV 2 from the mixed strain could be amplified by PCR and sequenced. All 120 sequences were deposited in NCBI with accession numbers MG450795 –MG450914.
The 2014–2017 DENV 1 strains (Fig 2) displayed a monophyletic relationship being clustered in the genotype I group. These strains showed a distant connection with strains from Vietnam, Cambodia, China and Taiwan-imported case from the same genotype. D1/Malaysia/330877/04 and D1/Malaysia/36139/05, which were two strains from the former 2004–2005 outbreaks formed their own clade, quite distinctively from the recent outbreak strains. One recent outbreak strain (D1/Malaysia/PUR/765027/15), isolated in December 2015 was noticeably set apart from other DENV 1 genotype I strains. There was no obvious clustering within the 2014–2017 DENV 1 strains as all of them were randomly dispersed within the clades.
The DENV 2 phylogeny (Fig 3) displayed a well-defined clade formation within the Cosmopolitan genotype. The Cosmopolitan genotype of Malaysian strains formed two main clades, which are referred to as Clade 1 and Clade 2. Clade 2 was further divided into sub-clade 2a and sub-clade 2b. There were two strains within Clade 1, namely D2/Malaysia/BAG/ 680429/14 and D2/Malaysia/WCK/738138/15, which were isolated in December 2014 and August 2015, respectively. Clade 2 comprised of strains isolated during the 2014–2017 outbreak. Domination of dengue strains by year of outbreak within sub-clade 2a and sub-clade 2b was observed. Early outbreak strains from the year 2014–2015 clustered in sub-clade 2b whereas late outbreak strains from the year 2015–2017 dominated sub-clade 2a.
The DENV 3 phylogeny (Fig 4) classified the current outbreak strains into genotype I and genotype III. The 2014–2015 DENV 3 strains were commonly noticed among genotype III clade, whereas the 2016–2017 strains appeared more in genotype I clades. The dengue strains from genotype III branched into two clades whereas genotype I strains were clustered together. The DENV 4 strains in this study belonged to genotype I and genotype II (Fig 5).
Of the total 120 cases investigated, 101 cases were included in the statistical analysis while the remaining 19 were omitted. As mentioned earlier, the excluded cases were those with incomplete clinical information, co-infection with leptospirosis and small sample size for particular serotype (DENV 4 and mixed serotype). The demographic details of the 101 study subjects are shown in Table 1. Age of the patients was significantly related to dengue serotypes. The median age group for DENV 1, DENV 2 and DENV 3-infected patients was 14, 27 and 23 respectively. A significant relationship was detected between dengue serotypes and disease classification (p = 0.007) (Table 2). Dengue infections without warning signs were observed more frequently in patients who were infected with DENV 1 (29/58; 50.0%) and DENV 3 (8/16; 50.0%). Notably, DENV 2-infected patients more frequently developed severe dengue (9/27; 33.3%). In the present study, a total of 17 severe dengue cases were identified of which two were noted as fatal cases. Apart from this, dengue with warning signs were frequently displayed by patients infected with DENV 2 (13/27; 48.1%).
The investigation of dengue serotype-specific clinical manifestations demonstrated that DENV 2-infected patients were more frequently present with persistent vomiting (p = 0.010), epigastric pain (p = 0.018), severe plasma leakage (p = 0.004) and shock (p = 0.038). Additionally, myalgia and arthralgia were the two major musculoskeletal symptoms observed in DENV 3 infection (p = 0.015, p = 0.014). Although statistically insignificant, relatively high proportion of DENV 1-infected patients suffered from lethargy (9/58; 15.5%) and diarrhea (12/58; 20.7%). Furthermore, genotype I of DENV 3 was frequently found in patients with myalgia (p = 0.036). Likewise, genotype III of DENV 3 and arthralgia were found to be associated with one another (p = 0.035).
Investigation of laboratory parameters in relation to dengue serotype and genotype is shown in Table 3. None of the laboratory parameters were significant among dengue serotypes and genotypes. A majority of the study subjects that were infected with DENV 1 (57/68; 98.3%) and DENV 2 (26/27; 96.3%) had primary infection. The percentage of secondary infection was greater in the DENV 3 (2/16; 12.5%) infected group than other serotypes. The mean platelet count was the lowest among DENV 2-infected patients as indicated by mean (±SD) of 106 x 109 (±71.5) whereas the hematocrit level was the highest among DENV 3 patients (43.0 (±5.9)).
Serotype association with disease severity in regards to primary and secondary infection was elucidated in Table 4. In the primary infection group, the percentage of severe and non-severe dengue cases was 16.5% (16/97) and 83.5% (81/97), respectively. Also, within the primary infection group, disease severity was significantly associated with dengue serotype (p = 0.014). Among the severe dengue cases within the primary infection group, a great proportion were infected with DENV 2 (9/16; 56.3%) followed by DENV 1 (6/16; 37.5%) and DENV 3 (1/16; 6.3%) whereas among the non-severe dengue cases, 63.0% (51/81) were infected with DENV 1, followed by DENV 2 (17/81; 21.0%) and DENV 3 (13/81; 16.0%). In addition, within the secondary infection group, the percentage of severe cases was 25.0% (1/4) while 75.0% (3/4) were non-severe cases. The only severe dengue case in the secondary infection group was infected with DENV 3 while the three non-severe cases were each infected with DENV 1, DENV 2, and DENV 3.
The present study focused on identifying the pattern of dengue serotype and genotype distribution from the year 2014–2017 in Malaysia and attempted to investigate the clinical spectrum of patients in relation to this distribution pattern. It was found that DENV 1 genotype I was the predominant serotype and genotype in the recent dengue outbreak in Malaysia. This reflected a serotype shift replacing the formerly predominant DENV 2 prior to the outbreak. Comparison of dengue serotype and genotype with disease spectrum contended that these clinical characteristics are serotype and genotype-specific. As most clinical symptoms of severe dengue infection only manifest at a much later stage of dengue infection, therefore, information on serotype or genotype-specific dengue manifestations may serve as early surrogate markers to predict disease progression.
In Malaysia, the serotype distribution of dengue virus has been inconsistent. However, there was a seemingly interesting pattern of circulation during outbreak period whereby major outbreaks were likely to follow the switching of DENV serotypes [12]. A predominantly circulating dengue serotype before an outbreak is replaced by another serotype which persists towards the end of the outbreak. Once the number of cases decline, the persistent serotype is again replaced by another serotype. For instance, during the 1996–1998 dengue outbreaks in Malaysia, both DENV 1 and DENV 2 dominated other serotypes. Prior to 1996, DENV 3 was consistently circulating. After the outbreak subsided, DENV 2 began to surge from the year 1999 onwards. Similarly, when there was a sharp increase in the number of dengue cases in 2001 and 2002, DENV 2 in the prior years was replaced by DENV 3. Then, DENV 3 was overtaken by DENV 1 from year 2003 onwards as the number of cases declined [2, 13, 14]. In the present study, DENV 1 was primarily present at the study area from 2014–2017. This indicated that DENV 1 had replaced a formerly predominant DENV 2. The existing trend contended that serotype replacement, most probably with DENV 3 is predicted to take place once the current outbreak subsides. In contrast to our prediction, Tan et al (2017) reported the unlikeliness for DENV 3 to surge as serotype cyclical outbreak cycle in Malaysia has recently been disrupted with DENV 3 remained in the background of DENV 1 and DENV 2 from 2003 until now [15].
Apart from serotype replacement, emergence of new genotypes or genotypic clade replacements have been reported during dengue outbreak [16, 17, 18]. This is often attributed to positive pressure selection and viral fitness for survival. In the current study, all DENV 1 strains were classified into genotype I, which subsequently formed minute branches within the same genotype but well conserved. The monophyletic pattern indicated high genetic similarity among the DENV 1 Malaysian strains. As for DENV 2 Cosmopolitan genotype, the current outbreak strains diverged into clades, indicating minor evolution.
The DENV 3 strains in our study were divided into genotype I and genotype III. Genotype III was constantly documented in Malaysia from 2007–2013 [15, 19]. Interestingly, genotype I of DENV 3 was last documented in 2011 and also circulated in this country during dengue epidemics in 1974 and 2007–2008 [15]. Hence, the appearance of genotype I that was equal to genotype III among DENV 3 strains in our study raised the possibility of re-emergence of genotype I especially during outbreaks. DENV 4 is a rare dengue serotype in Malaysia. Three strains from genotype I and a strain from genotype II were detected in the present study. One of the DENV 4 strains clustered together with the dengue strains from Myanmar, reflecting the likelihood of an imported case.
A handful of studies have shown the differences in dengue symptoms are serotype specific, but majority of these studies are conducted outside outbreak periods. A study among adult dengue patients in Singapore from 2005–2011 found that cases infected with DENV 1 were more likely to be presented with red eyes and had higher risk of developing severe dengue. In contrast, those DENV 2-infected patients were found to have frequent joint pains and significantly low platelet count [9]. Another study was conducted among dengue patients in Peru, Bolivia, Ecuador and Paraguay from 2005–2010. The findings showed that DENV 3 had a higher prevalence of musculoskeletal and gastrointestinal manifestations while DENV 4 had a higher prevalence of respiratory and cutaneous manifestations [7]. On the other hand, a group of Indian researchers who undertook a study from 2002–2006 in New Delhi reported a significant result of hepatomegaly and abdominal pain in DENV 2-infected patients. The results of their study also noted that DENV 4 patients suffered from severe hemorrhagic manifestations [20]. All the aforementioned results showed that there were differences in clinical manifestations of dengue patients relating to the serotype owing to the differences in terms of year of study and level of endemicity in particular regions. In response to this, the present study investigated the relationship between dengue serotype and genotype, and the disease spectrum in Malaysia, a dengue hyper endemic country, during a recent outbreak period.
The present study showed a significant number of DENV 2-infected patients developed severe dengue more frequently as compared to other serotypes. These patients were also frequently presented with clinical manifestations such as persistent vomiting, epigastric pain, plasma leakage and shock. Many studies have proven that severity and DENV 2 are common especially in relation to shock manifestation [21, 22, 23]. Similarly in a retrospective study of dengue cases in Thailand, it was also found that DENV 2 was the most frequent serotype isolated from dengue hemorrhagic (DHF) and dengue shock syndrome (DSS) cases followed by DENV 3, DENV 1 and DENV 4 [24]. Although substantial evidence confirmed that DENV 2 caused severe dengue infection, one study found no differences between any DENV serotypes and severity of the disease [25].
Our data suggested that both DENV 1 and DENV 3-infected patients displayed the mildest clinical presentation. The data also signified that DENV 3-infected patients suffered from myalgia and arthralgia. When compared across the genotypes, DENV 2 cosmopolitan was highly prevalent among patients with epigastric pain and shock. Besides that, DENV 3 genotype III was significantly observed among patients with arthralgia whereas DENV 3 genotype I was common among patients with myalgia. Several studies also reported an association between DENV 3 and musculoskeletal manifestations. According to Chan et al. (2009), the findings showed a significant relationship between DENV 3 and myalgia, but the genotype information was unknown [26]. Similarly, Hasley et al (2012) also found that DENV 3-infected patients had frequent myalgia and arthralgia symptoms [7]. Therefore, we hypothesized that DENV 3 has a high preference and binding affinity to the receptors in the musculoskeletal system, narrowing down further that genotype I targets muscles whereas genotype III targets the joints.
The comparison between genotypes also exhibited that genotype III of DENV 3 may appear to be more virulent than genotype I. This is supported by the presence of two severe cases among DENV 3 genotype III-infected patients. These two cases were presented with severe organ impairment and shock (dengue shock syndrome), respectively. This finding may warrant further interest to investigate severity associated with DENV genotype III with inclusion of larger sample size. It can be postulated that since genotype III has been circulating dominantly among DENV 3 genotypes in Malaysia for extended period from its introduction in 2007 till 2013, thus it acquired sufficient survival fitness to become virulent over time as compared to genotype I which was dormant from 2012–2013 and had re-emerged during the outbreak in 2014. However re-emergence of DENV 3 genotype I has to be cautiously monitored as it could undergo adaptation to become more infectious to replace and diminish the existence of DENV 3 genotype III. This highlights the importance of studying serotype and genotype specific clinical manifestations during dengue outbreak to allow prediction of an outbreak outcome.
Previous research had claimed that DENV 1 and DENV 2 are often associated with primary infections and secondary infections, respectively [26]. Contrary to these findings, our investigation revealed insignificant differences between dengue serotypes and infection type as in primary and secondary infection. Indeed, this could be mainly due to low number of secondary infected cases in the present study. In our study, primary dengue infections were found to be more frequent than secondary infections. This finding contradicted with reports of high seroprevalence rate of secondary infection in Malaysia [27]. The seroprevalence rates in Malaysia were shown to be age dependent, whereby as the age increases the seroprevalence also increases and by 70 years old almost 80% of the Malaysian population were already infected with dengue. Our findings on secondary infection rate was different from the above study based on two aspects. Firstly, the distribution of subjects by age group in our study indicate that more than half of the study subjects belonged to younger age group (≤ 20); hence, less likely to have past exposures (S1 Table). Past studies also supported this in which a shift of the dengue infection from affecting primarily children to adults more than age 20 years old, was demonstrated [28, 29]. Secondly, in our study, exclusion of samples with incomplete information, unable to be serotyped or genotyped and the coverage of only two locations contributed to small sample size, thus, we are unable to provide a detailed description of the seroprevalence rate. However, our study findings on high primary infection rate could pose a serious implication if the next outbreak is predominantly caused by DENV 2. It is known that severe dengue is associated with secondary infection and infection with DENV 2 superimposed on previous DENV 1 infection carries the highest risk for development of severe dengue [30]. Even though the DENV 3 upsurge after the current outbreak is predicted, however as mentioned earlier, one study [15] has cautioned that this cycle was recently disrupted, raising the possibility for DENV 2 dominance. Therefore, our study findings on the high prevalence of primary infection is very crucial in the conveyance of early warning for a severe outbreak in the future.
Comparison of serotype severity after stratifying by infection types demonstrated that among primary dengue infection cases, there is a significant association between DENV 2 and the disease severity. This indicates that primary infections caused by DENV 2 may lead to more severe presentations than DENV 1. Interestingly, only one severe case was found in the secondary infection group and no significance was observed among the serotypes and severity in this group. One possible explanation, in this case, could be that the patients in the secondary infection group could have been previously infected by the same serotype (homotypic infection). Recent studies have reported the evidence of homotypic dengue re-infection [31, 32]. The findings for these studies demonstrate that, in some cases, serotype-specific immunity to DENV may be short-lived. Such cases challenge the paradigm of lifelong, serotype-specific DENV immunity following a natural infection. A greater risk to develop severe dengue is caused by heterotypic secondary DENV infection with a dengue serotype distinct from the primary infecting type [33]. Having said that, prior to the onset of the recent outbreak, DENV 1 were equally circulating with other serotypes (DENV 2 and DENV 3) from the year 2010–2012. It was only in the year 2013, there was an upsurge of DENV 2. Since the information is not available on the serotype that previously infected the patients in the secondary infection group, therefore our findings raised the possibility that these patients had been previously infected by DENV 1 also. Hence, increased disease severity was not observed. The theory of peak enhancement titer provides a second possible explanation. The enhancement of severe dengue is dependent on the pre-existing antibody titer, resulting from a different serotype infection. The pre-existing dengue antibody titer of 1: 21 to 1:80 induced high risk to develop severe dengue, whereas, a titer above 1: 320 triggered a protective effect [33]. It is more likely that the pre-existing dengue antibody titers in the secondary infection group in our study did not fall into the peak enhancement zone, thus causing less risk for severity. The IgM and IgG rapid test results were utilized to discriminate between primary and secondary infection. Several studies have evaluated the accuracy of infection status assignment using this method. One such study demonstrated a 100% sensitivity of an IgM/IgG rapid test in an attempt to distinguish between primary and secondary dengue virus infections [34]. Another study evaluated the similar rapid test kit (Dengue Duo Cassette; Panbio, Brisbane, Australia) which was also used in the present study and found that the test has good reproducibility, with the inconsistency of only 5% [35]. These findings collectively showed that the IgM/IgG rapid tests are reliable diagnostic tools for the indication of primary or secondary dengue infection.
Several strengths are inherent in this study. To the best of the researcher’s knowledge, this is the first observational study to investigate the relationship between dengue serotype and genotype with patients’ clinical manifestation during the recent outbreak peri-od among Malaysians. Additionally, our study findings may warrant further research to elucidate the strong association of dengue serotype and genotypes focusing on a larger population and localities. However, the study has limitations as well. Firstly, all retrospective studies depend on the completeness of medical data, and missing data is therefore unavoidable. In the present study, a number of subjects had to be excluded from the original count due to incompleteness of clinical data, thus resulted in a smaller sample size. Secondly, in Malaysia, many dengue specimens from hospitals were sent to the national reference laboratories for molecular diagnosis and surveillance purpose. Therefore, our collaborators from the hospitals were only able to provide those specimens that were not required for national surveillance. We need to acknowledge that there is a possibility of selection bias but, the findings in our study of resurgent DENV 1 infection during the recent outbreak period is consistent with the national surveillance of the dengue infection. In view of this, the selection bias was not substantial. Not only that, there was underrepresentation of certain serotypes in this study due to the small numbers such as DENV 4 causing it to be omitted from the statistical analysis. Lastly, the proportion of dengue subjects obtained in the year 2014 was low despite a spike in the dengue cases because the sample collection was initiated towards the end of the particular year.
Our study findings demonstrated that symptoms of dengue infected patients in Malaysia were indeed serotype and genotype-specific. DENV 1 was found to cause dengue without warning signs and mild symptoms. DENV 2 patients were more likely to present with severe dengue as compared to other serotypes. In addition, DENV 3-infected patients frequently had musculoskeletal symptoms. These findings suggested that different dengue serotypes targeted different receptors or organs to establish infection. Following our findings and to benefit from further research, we recommend continuous monitoring of dengue clinical manifestations in relation to serotype and genotype and investigate the link between DENV 3 genotype III and disease severity with larger sample size. We also propose to look into the recent prevalence of primary and secondary dengue infection among Malaysian population as this has an implication on dengue vaccine.
|
10.1371/journal.pntd.0001304 | Metagenomic Analysis of Taxa Associated with Lutzomyia longipalpis, Vector of Visceral Leishmaniasis, Using an Unbiased High-Throughput Approach | Leishmaniasis is one of the most diverse and complex of all vector-borne diseases worldwide. It is caused by parasites of the genus Leishmania, obligate intramacrophage protists characterised by diversity and complexity. Its most severe form is visceral leishmaniasis (VL), a systemic disease that is fatal if left untreated. In Latin America VL is caused by Leishmania infantum chagasi and transmitted by Lutzomyia longipalpis. This phlebotomine sandfly is only found in the New World, from Mexico to Argentina. In South America, migration and urbanisation have largely contributed to the increase of VL as a public health problem. Moreover, the first VL outbreak was recently reported in Argentina, which has already caused 7 deaths and 83 reported cases.
An inventory of the microbiota associated with insect vectors, especially of wild specimens, would aid in the development of novel strategies for controlling insect vectors. Given the recent VL outbreak in Argentina and the compelling need to develop appropriate control strategies, this study focused on wild male and female Lu. longipalpis from an Argentine endemic (Posadas, Misiones) and a Brazilian non-endemic (Lapinha Cave, Minas Gerais) VL location. Previous studies on wild and laboratory reared female Lu. longipalpis have described gut bacteria using standard bacteriological methods. In this study, total RNA was extracted from the insects and submitted to high-throughput pyrosequencing. The analysis revealed the presence of sequences from bacteria, fungi, protist parasites, plants and metazoans.
This is the first time an unbiased and comprehensive metagenomic approach has been used to survey taxa associated with an infectious disease vector. The identification of gregarines suggested they are a possible efficient control method under natural conditions. Ongoing studies are determining the significance of the associated taxa found in this study in a greater number of adult male and female Lu. longipalpis samples from endemic and non-endemic locations. A particular emphasis is being given to those species involved in the biological control of this vector and to the etiologic agents of animal and plant diseases.
| Leishmaniasis is a vector-borne disease with a complex ecology and epidemiology. It has three main clinical forms of which visceral leishmaniasis (VL) is the most severe, as it is fatal if untreated. It is caused by a protist parasite, Leishmania spp., and is transmitted to humans by phlebotomine sandflies. The best method to interrupt any vector-borne disease is to reduce man-vector contact. Vector-targeted strategies are particularly attractive because the vectorial capacity to transmit infectious diseases to humans is proportional to vector density and, in an exponential way, to vector survival. Biological control is an effective means of reducing or mitigating pests through the use of natural enemies and is more environmentally friendly than traditional insecticide treatments. Nevertheless, there is very scanty information on the biological control of sandflies and their potential control agents. In this context, a detailed knowledge of the microorganisms that are associated with these vectors would aid in the development of novel strategies for controlling them. This is the first study to survey the taxa associated with leishmaniasis vectors and, more importantly, with any infectious disease vector, using an unbiased and high-throughput approach.
| Leishmaniasis is a vector-borne neglected infectious disease of worldwide incidence and its most severe clinical form is visceral leishmaniasis (VL). Each year VL causes an estimated 500,000 new cases and more than 59,000 deaths [1], a death toll that is only surpassed by malaria among the parasitic diseases [2]. Furthermore, both figures are approximations since VL is frequently not recognized or not reported [3]–[4]. Leishmaniasis is transmitted through the bite of two phlebotomine sandfly genera, Phlebotomus in the Old World and Lutzomyia in the New World. In Latin America VL is caused by Leishmania infantum chagasi and transmitted by Lutzomyia longipalpis [5]. This phlebotomine sandfly is only found in the New World, with a wide distribution from Mexico to Argentina [6]. The geographical distribution of leishmaniasis has undoubtedly expanded and is now being reported in areas that were previously non-endemic. The worldwide phenomenon of urbanisation, closely related to the sharp increase in migration, is one of the major risk factors that is making leishmaniasis a growing public health concern for many countries around the world [7] and Argentina is not an exception. Between 1925 and 1989 only 14 leishmaniasis human cases were reported in Argentina and none was attributed to Le. chagasi. Moreover, there were only two isolated reports of Lu. longipalpis (in 1953 and 2000) which were not associated with VL [8]. Nevertheless, this situation has changed dramatically, mostly due to an indiscriminate advance of urbanisation, and the first Argentine VL outbreak was recently reported [9]. From 2006 to date the morbidity and mortality toll of this disease have amounted to 83 human cases (35% corresponding to children under ten years of age), 7 deaths and more than 7,000 infected dogs (National Health Surveillance System, Epidemiology Bureau, National Ministry of Health, Argentina).
In the natural environment, phlebotomine larvae feed on organic matter from soil [10], while adults from both sexes feed on sugars from plant sources [11]–[12]. Only female adults need blood to obtain necessary proteins for the development of their eggs [5]. It is widely accepted that many insects derive their microbiota from the surrounding environment, such as the phylloplane of food plants or the skin of the animal host, and although the degree of persistence of strains of the ingested species is unknown, these microorganisms can influence the insect life cycle [13]. To comprehensively understand the biology of insects, microorganisms must be considered as a very important component of the ecological system [14]. Moreover, an inventory of the associated microbiota of phlebotomine sandflies, especially of wild specimens, would aid in understanding the annual and regional variations recorded for this disease [15] and in the development of novel strategies for controlling these vectors, among others [13]. One serious obstacle for the biological control of VL sandfly vectors is that their precise breeding sites are poorly known. Furthermore, its practical application seems to be limited to the adult VL vector stage [16] because, as Lu. longipalpis larvae appear to be thinly dispersed [17], this complicates the employment of biolarvicides in the field.
There is scanty information on the microbial colonisation of Lu. longipalpis and it is not yet clear if they possess an indigenous community. Previously, midgut bacteria were examined from wild and laboratory reared Lu. longipalpis populations [18]–[20] which showed a predominance of Gram negative bacteria. Various genera found ubiquitously in the environment (water, soil and debris) were identified in these studies, including Acinetobacter, Serratia, Pseudomonas, Stenotrophomonas, Flavimonas and Enterobacter. These bacteria have also been found associated with the gut of several other insects [21]–[24], suggesting they are a part of the natural or transient microbiota. Prior studies on guts and malpighian tubes from wild P. papatasi and P. tobbi showed a high incidence of mycoses which were similar to Aspergillus sclerotiorum and Saccharomyces cerevisiae [25]. Various types of virus have also been found infecting phlebotomine sandflies [26], including Vesiculovirus [27]–[28] and Cytoplasmic Polyhedrosis Virus [29]. Furthermore, in addition to Leishmania, Trypanosoma, Endotrypanum and possibly other trypanosomatids [30], neotropical sandflies may harbour other parasites including microsporidians [31]–[32], gregarines [33]–[36], some Plasmodium spp. that parasitise lizards [37] and nematodes [38]–[41]. Nevertheless, there is little information on the pathological effects these parasites may produce in their sandfly hosts.
Metagenomics facilitates the culture-independent analysis of microbial communities [42], an approach which does not require prior assumptions about the composition of the target community. Metagenomic sequencing of communities containing eukaryotes, in particular protists, is mostly cost-prohibitive because of their enormous genome sizes and low gene coding densities [43]. Nevertheless, from an ecological perspective, excluding eukaryotes from a metagenomic analysis compromises the ability to assess the microbial community in its entirety. A possible approach to bypass the problem of large amounts of non-coding eukaryotic sequence data consists in obtaining molecular data at the RNA level. Given the recent VL outbreak in Argentina and with the ultimate goal of identifying possible biological control agents, this study used unbiased high-throughput pyrosequencing technology [44] to compare the diversity of the taxonomic groups associated with wild male and female adult Lu. longipalpis from endemic (Posadas, Misiones) and non-endemic (Lapinha Cave, Minas Gerais) VL locations in Argentina and Brazil, respectively. As in this study phlebotomine sandflies were considered environmental samples, the term “associated with” was used here in its broadest sense, referring to a wide variety of possible interactions ranging from casual associations due to random environmental contact (e.g., plant pathogenic fungi spores adhering to the hairy surface of the sandflies when sugar-feeding on plants) to closer pathogenic or symbiotic interactions (e.g., protists that parasitise phlebotomines or permanent gut microbiota, respectively). This analysis revealed the presence of sequences from bacteria, fungi, protists, plants and metazoans.
This study was carried out in strict accordance with the recommendations in the Manual for the Use of Animals/FIOCRUZ (Manual de Utilização de Animais/FIOCRUZ) of Fundação Oswaldo Cruz, FIOCRUZ, Ministry of Health of Brazil (National decree Nr 3,179). The protocol was approved by the Ethics Committee for the Use of Animals of the Fundação Oswaldo Cruz - FIOCRUZ, Ministry of Health of Brazil (Nr 242/99).
Lu. longipalpis specimens from the non-endemic VL location, Lapinha Cave (Minas Gerais, Brazil), situated in the Sumidouro National Public Park, were kindly provided by Dr. Paulo Pimenta (Laboratory of Medical Entomology, Centro de Pesquisas René Rachou, Fundação Oswaldo Cruz, FIOCRUZ). Sandflies from this location were chosen as reference because they have been extensively studied. Lu. longipalpis specimens from the endemic VL location, Posadas (Misiones, Argentina), where they occur in high density, were kindly provided by Dr. María Soledad Santini and Mr. Enrique Adolfo Sandoval (Research Network for Leishmaniasis in Argentina, REDILA, and Posadas Municipality Quality of Life Department).
Captures were made using CDC light traps [45] on the 15th and 26th of May 2009 in the Lapinha Cave and in Posadas, respectively. The Lapinha Cave (S19 33 42.42 W43 57 34.96) is a network of interconnected caves located in a vast tropical savanna ecoregion called cerrado, characterised by great plant and animal biodiversity. The trap was left 50–80 cm above ground level in an external small annex cave (2 mt long) where a chicken was kept to attract the sandflies and as a source of food (see Table 1 for a detailed description of the site). Posadas, the densely populated capital city of the province of Misiones, is located in the subtropical fields and grasslands ecoregion. In the Posadas area this ecoregion contacts the Paranaense forest and has a savanna-type landscape. The trap was installed in the peridomicile of a worst-case scenario homestead (domestic animals, dense vegetation, nearby spring) (S27 23.266 W55 53.403) (see Table 1 for a detailed description of the site).
Sandflies were transported alive in a nylon cage to the corresponding laboratories in Belo Horizonte (Minas Gerais) and Posadas (Misiones) and no mortality was registered on arrival. Other insect species were captured together with the sandflies including hymenopterans, lepidopterans and mosquitoes. Sandflies were killed at low temperature, identified and separated according to sex, and stored alternatively in Tri-Reagent (Molecular Research Center Inc., Cincinnati, OH) or RNAlater® (Qiagen). A total of four groups of 100 sandflies each, two per location, were separated and named according to: SS1, females from the Endemic VL location (EVL females); SS2, EVL males; PP1, females from the Non-Endemic VL location (NEVL females); and PP2, NEVL males.
Individual samples were ground in Tri-Reagent (Molecular Research Center Inc., Cincinnati, OH) with a Teflon pestle and total RNA was immediately extracted, according to the manufacturer's instructions. Total RNA was amplified using a modified sequence-independent amplification protocol [46]. Briefly, M-MuLV Reverse Transcriptase (Fermentas, Vilnius, Lithuania) was used for a first-strand reverse transcription which was initiated with a random octamer linked to a specific primer sequence (5′-GTT TCC CAG TAG GTC TCN NNN NNN N-3′) [47]. cDNA was then amplified with the Expand Long Template PCR System (Roche) using a 1∶9 mixture of the above primer and a primer targeting the specific primer sequence (5′-CGC CGT TTC CCA GTA GGT CTC-3′) [48]. The following profile was used: initial denaturation cycle at 94°C for 2 minutes; five low stringency cycles with denaturation at 94°C for 30 seconds, 25°C for 30 seconds and 68°C for 6 minutes, were followed by 30 cycles at 94°C for 30 seconds, 55°C for 30 seconds and 68°C for 6 minutes and a final extension cycle at 68°C for 5 minutes. Pooled samples were submitted for high-throughput pyrosequencing (Macrogen Inc., Korea).
Reads were submitted to the NCBI Sequence Read Archive (SRA) (submission SRA026595) under accessions SRR089611 (adult EVL female Lu. longipalpis; Posadas, Misiones, Argentina; SS1), SRR089612 (adult EVL male Lu. longipalpis; Posadas, Misiones, Argentina; SS2), SRR089613 (adult NEVL female Lu. longipalpis; Lapinha Cave, Minas Gerais, Brazil; PP1) and SRR089614 (adult NEVL male Lu. longipalpis; Lapinha Cave, Minas Gerais, Brazil; PP2).
Reads ranged in size from approximately 100 to 1200 base pairs (bp) (350 bp average). Raw sequence reads were trimmed to remove sequences derived from the amplification primer. With the purpose of reducing database search efforts and improving the homology detection sensitivity [49], Cd-hit [50] was used to generate non-redundant nucleotide datasets but these represented less than 1% in every case (data not shown). For this reason, singlet sequences were used for the nucleotide database search. Non-redundant (nt) and non-human, non-mouse ESTs (est-others) NCBI databases last modified on 23/04/10 and 25/04/10, respectively, were downloaded locally (ftp://ftp.ncbi.nlm.nih.gov/blast/db/). After trimming, singlet sequences were compared to these databases using BLASTN (nucleotide homology) [51], with a 1e-50 cutoff E-value. The resulting BLAST alignments were analysed and classified according to their taxonomical hierarchies using custom applications written in Mathematica (Wolfram Mathematica 7; available upon request). 16S sequences were confirmed by alignment to type-species 16S rRNA sequences from the Ribosomal Database Project (http://rdp.cme.msu.edu/) [52]–[53]. Hits for every taxon were individually revised and confirmed and only those which showed unequivocal results were included in the final analysis. Fisher's Exact Test [54] (p<0.05) was used to establish the significance of sequences in the different samples using a custom application written in Mathematica (Wolfram Mathematica 7; available upon request).
The vast majority of reads obtained for each sample corresponded to Lu. longipalpis sequences (∼85%) and an important fraction showed no significant hits in the homology searches against the different databases (∼14%). Hits which corresponded to taxa other than Lu. longipalpis represented less than 0.2% of the total reads.
Results for each taxon were organised separately in this section. Figure 1 emphasises the treatment of these vectors as environmental samples. It shows an overview of the workflow used in this study, summarising and associating information on the sampling sites and on the taxa identified by sequence homology in each adult Lu. longipalpis sample. A detailed ecological description of both sampling sites is given in Table 1. Figure 2 integrates and summarises results for all the samples, indicating the taxa that were identified in each case, the species that were found for each taxon and the number of sequences for each species. Table S1 lists all the reads that showed significant hits and a brief description of each hit.
BLASTN analysis of the high-throughput sequencing data identified bacteria in females from both locations (SS1 and PP1) and in NEVL males (PP2) (Figures 1 and 2). Bacteria were identified mostly by homology to completely sequenced bacterial genomes (7 reads), followed by rRNA genes (4 reads) and lastly to plasmid sequences (2 reads) (Table S1). Ten different bacterial types were identified, six of which showed homology at the species level (five to genomic sequences and one to plasmid sequences) and four to diverse uncultured environmental samples (three to 16S rRNA genes and one to genomic sequences). The bacterial composition was different and unique in every case and included sequences from Ralstonia pickettii, Anoxybacillus flavithermus, Geobacillus kaustophilus, Streptomyces coelicolor, Propionibacterium acnes, Acinetobacter baumannii, uncultured Veillonella sp. and uncultured bacterium clones isolated from environmental samples (cow faeces, wetland soil and water) (Figure 2; Table S1).
The totality of identified bacteria showed a predominance of Gram negative rods (53.8%, 7 reads) and a significant proportion of Gram positive bacteria (38.5%, 5 reads), in accordance with previous studies [18]–[20], [55]. Of all the bacterial sequences that were identified in this study, only A. baumannii, which was found in NEVL males (PP2), had been previously identified in adult female Lu. longipalpis. This species had been isolated from female laboratory reared specimens from the same non-endemic VL location (Lapinha Cave) [20] and from wild female specimens from endemic VL locations in Brazil (Jacobina, Bahia, and São Luís, Maranhão) [18].
Fungi were only found in NEVL males and females (PP1 and PP2) (Figure 1). A total of four different species was identified by homology to rRNA genes (Figure 2; Table S1). These differed between males and females and have not been found to date associated with phlebotomines. The identified species included Peronospora conglomerata, Cunninghamella bertholletiae, Mortierella verticillata and Toxicocladosporium irritans (Figure 2; Table S1).
Protist sequences were only identified in EVL female and male specimens (SS1 and SS2) (Figure 1), of which the vast majority (99.8%) were found in males (Figure 2; Table S1). Protists were identified by homology to sequenced rRNA genes (360 reads, 61.4%), cDNA (208 reads, 35.5%) and chromosomal sequences (18 reads, 3.1%) (Table S1). Ten species and one genus of apicomplexan parasites were identified that parasitise Diptera (Ascogregarina taiwanensis, Psychodiella chagasi), birds (Eimeria tenella, Sarcocystis falcatula, Sarcocystis cornixi), mammals (Cryptosporidium muris, Sarcocystis arieticanis, Besnoitia besnoiti, Plasmodium falciparum, Plasmodium berghei) and reptiles, birds and mammals (Sarcocystis sp.) (Figure 2; Table S1). Most of the apicomplexan sequences (64.8%, 379 reads) were homologous to the mammalian parasites C.muris, S. arieticanis, B. besnoiti, P. falciparum and P. berghei. Of these, more than half (53.6%, 203 reads) were homologous to published cDNA libraries and the rest to rRNA genes (41.7%, 158 reads) and chromosomal DNA (4.7%, 18 reads). The second most numerous group of apicomplexan sequences was homologous to the avian parasites E. tenella, S. cornixi and S. falcatula rRNA genes (28.2%, 165 reads) and the rest were homologous to the dipteran parasites A. taiwanensis and P. chagasi rRNA genes (7%, 41 reads).
Metazoan sequences (mammals, birds and reptiles) were also found in all the samples and included Homo sapiens, Gallus gallus and Anolis carolinensis (Figures 1 and 2). Homo sapiens was identified by homology to genomic chromosomal sequences (16 reads). Gallus gallus was identified by homology to genomic chromosomal sequences (46 reads), cDNA (18 reads) and rRNA genes (1 read). Anolis carolinensis was identified by homology to cDNA (1 read) (Table S1). Human sequences were found in males and females from both locations, whereas chicken sequences were found in NEVL females (PP1) and EVL males (SS2) and lizard sequences were only found in NEVL females (PP1) (Figures 1 and 2).
A total of ten different plant species were identified in males and females from both locations, namely Elaeis guineensis, Capsicum annuum, Juglans hindsii, Artemisia annua, Brassica napus, Vitis vinifera, Solanum tuberosum, Nicotiana tabacum, Oryza sativa and Rhapidophyllum hystrix (Figures 1 and 2). All plant sequences were identified by homology to cDNA libraries (56 reads), except for R. hystrix which was identified by homology to rRNA genes (1 read) (Table S1). EVL males and females showed a greater number of species (5 and 6 species, respectively), followed by NEVL males (4 species) and lastly NEVL females (3 species) (Figure 2). However, these differences were significant (p<0.05) only between females from both locations (Table 2). EVL females showed the highest number of plant sequences (23 reads), followed by EVL and NEVL males (13 reads) and lastly NEVL females (8 reads) (Figure 2). The only case in which these differences were not significant (p<0.05) was between EVL females and NEVL males (Table 3).
C. annuum (bell pepper) was found in EVL males and females and in NEVL males. E. guineensis (African oil palm) was found in EVL females and NEVL males and females. S. tuberosum (potato) was found in EVL males and females and in NEVL females. J. hindsii (Northern California walnut) was found in EVL males and females and A. annua (sweet wormwood) was found in males from both locations. R. hystrix (needle palm) and B. napus (rapeseed) were only found in EVL females and V. vinifera (grapevine) was only found in EVL males. O. sativa (rice) was only found in NEVL females and N. tabacum (tobacco) was only found in NEVL males (Figure 2).
This is the first study to survey taxa associated with an infectious disease vector applying an unbiased and comprehensive metagenomic approach. To ensure an unbiased description of the microbial community, the rationale chosen for this study included the extraction of total RNA and sequence-independent amplification. Total RNA was extracted from wild adult male and female Lu. longipalpis from an endemic (Posadas, Misiones) and a non-endemic (Lapinha Cave, Minas Gerais) VL location in Argentina and Brazil, respectively, and submitted to high-throughput pyrosequencing [44]. Given the high background level of vector sequences (∼85%), this approach proved to be very sensitive since it enabled the identification of taxa present in percentages up to 0.00036%. Moreover, as the different taxa were identified by homology to both rRNA and mRNA, the chosen approach was adequate for the objectives of this study.
The bacterial community identified in females from both locations and in NEVL males was distinct in every case. The only results in common with previous studies of gut microbiota from wild and laboratory reared female Lu. longipalpis and laboratory reared female P. duboscqi [18]–[20], [55], included the prevalence of Gram negative bacteria and the identification of A. baumanni, which in this study was found in NEVL males. Although previous reports established an essential basis for phlebotomine gut microbiota current knowledge, in these studies bacteria were identified using standard bacteriological methods. Consequently, those descriptions did not consider the remaining 99% of unculturable environmental microbes [56]. Hence the differences with this study, which applied a culture independent unbiased high-throughput approach that bypassed cloning of environmental DNA.
Interestingly and in accordance with results from this study, in previous reports the proportion of bacteria isolated from wild dipterans has been low. Studies on the midgut microbiota of wild mosquitoes, isolated bacteria from less than 50% of the specimens and the numbers of bacteria varied between individuals [57]–[58]. In a more recent study which used culture dependent and independent screening of field-collected Anopheles, bacteria were found in 15% of the mosquitoes, few of the mosquitoes harboured more than one bacterial species and only one species was found in more than one mosquito [23].
Only one bacterial type was found in EVL females (SS1), which corresponded to an unculturable bacterium originally isolated from cow faeces (Figure 2). Furthermore, a sequence match against RDP [52] indicated high similarity with Alistipes sp., a Gram negative anaerobic bacteria found in human faeces (Figure 2). Five bacterial species were found in NEVL females (PP1), four of which were Gram positive (Figure 2). Of these species, R. pickettii, A. flavithermus, G. kaustophilus and S. coelicolor, were originally isolated from contaminated lake sediment, waste water [59], deep-sea sediment [60] and soil [61], respectively. Interestingly, A. flavithermus and G. kaustophilus are thermophilic. Even though R. pickettii 12D was originally isolated from contaminated lake sediment, it is a ubiquitous microorganism found in water and soil [62] and is emerging as an opportunistic pathogen found in a wide variety of clinical samples [63]. P. acnes [64] is a universal inhabitant of human skin and is found at high population densities on the fat-rich areas of the face, scalp and upper trunk [65]. Four bacterial types were found in NEVL males (PP2), 50% of which were Gram negative (Figure 2). One of these bacterial types was the multidrug-resistant A. baumannii [66], which is recovered from natural environments and has emerged as an important opportunistic pathogen worldwide [67]. Another of the bacterial types corresponded to uncultured Veillonella sp. isolated from human skin [68]. The other two bacteria were uncultured bacterial types originally recovered from environmental samples. In one case, BLASTN analysis indicated homology both to an uncultured bacterium from a water sample (Atlantic Ocean) and to Leifsonia xyli, a sugar-cane pathogen [69]. The other bacterial sequence corresponded to a proteobacterium clone isolated from wetland soil [70].
In summary, bacteria identified in this study are ubiquitous in the diverse environments these sandflies frequent (faeces, soil, water, sediment, plants, human skin) and which were present in both sampling sites (Figure 1, Table 1). Hence, possibly they were indicative of the behavioural patterns and feeding habits of these sandflies and are probably part of their transient microbiota. However, more in depth research is required to determine these interactions.
Four different species of fungi were found in NEVL Lu. longipalpis (PP2 and PP1), which differed between males and females (Figure 2). In previous reports for P. papatasi and P. tobbi, mycoses with a high incidence rate were found in the guts and malpighian tubes of wild specimens. Similar fungi cultured from guts of laboratory reared P. papatasi were identified as A. sclerotiorum and S. cerevisiae [25]. Microsporidians, which are highly pathogenic for some insects [71], have also been found parasitising neotropical sandflies [32], [72]. The two species identified in this study in NEVL females were P. conglomerata [73], a plant pathogen (mildew), and C. bertholletiae [74], a common soil fungus and a rare cause of zygomycosis in humans. On the other hand, the species found in NEVL males were M. verticillata [75], a genus commonly found in soil and a zygomycete which also causes zygomycosis in animals, and T. irritans [76], which belongs to a genus of foliar pathogens [77]. Given the very high vegetation density in the Lapinha Cave area (Figure 1, Table 1), a possible scenario is that plant pathogenic spores adhered to the sandflies' hairy surface during sugar-feeding on infected plants. This suggested Lu. longipalpis has a putative capacity of casual dispersal of plant pathogens, among others (see below). In conclusion, fungi identified in this study are found ubiquitously in the environments frequented by sandflies (plants and soil), which were abundant in the sampling site (Lapinha Cave), and so were probably indicative of their sugar-feeding habits and behavioural patterns.
Protist sequences were only found in EVL male and female specimens, of which the vast majority were found in males (Figure 2). Nearly 90% of the identified apicomplexans corresponded to coccidians (genera Cryptosporidium, Eimeria, Sarcocystis and Besnoitia, 88.7%) and the rest to gregarines (genera Ascogregarina and Psychodiella, 7%) and haemosporidians (Plasmodium spp., 4.3%). The absence of leishmanial sequences was not unexpected considering the rate of infection of sandflies with Leishmania is generally very low (0.01–1%) [78], even in endemic areas [5].
Gregarines have been reported in over 20 species of sandflies and Ascogregarina spp. have only been described in mosquitoes [79]. Given A. taiwanensis sequences were found in EVL males in this study, this could indicate that the parasite also infects Lu. longipalpis. Genus Psychodiella comprises 3 species with host specificity to phlebotomine sandflies: P. chagasi, P. saraviae and P. mackiei [35], [79]. In the New World, only P. chagasi and P. saraviae have been found parasitising Lutzomyia spp. and P. chagasi seems to infect a large range of neotropical species [33]–[36]. In this study, P. chagasi sequences were found in EVL males. The exact pathology caused by gregarines is unknown, but in Lu. longipalpis the parasite can reduce longevity and egg production and the level of parasitaemia can reach over 80% in laboratory colonies [80]. Notwithstanding, the use of P. chagasi as a control method in the field has not been considered efficient because the parasite seems to have a limited range and a minimal effect on sandfly biology under natural conditions [81]. The fact that in this study P. chagasi was found in randomly caught wild specimens, suggested it could be a more efficient control method under natural conditions than what was previously reported.
The free-living oocyst stage of coccidians is discharged by infected animals through their faeces. Sandflies are found around human habitations and breed in specific organic wastes, exploiting the accumulation of organic matter produced by domestic animals and poor sanitary conditions such as faeces, manure, rodent burrows and leaf litter [5]. Since the EVL sampling site (Posadas) was a worst case-scenario homestead which included dense vegetation, various domestic animals and abundant organic matter (Figure 1, Table 1), this could account for the presence of these parasites in EVL males and females. On the other hand, female sandflies suck blood from different animal species including humans, bovines, pigs, equines, dogs, opossums, birds, various rodents and reptiles [5], [82] and, additionally, some Plasmodium spp. that parasitise lizards have been found in sandflies [37]. In the field it is common to see lek-like aggregations of males and females assembled on or near hosts where blood feeding and mating occur [83]–[84]. This behaviour could account for the presence of haemosporidian sequences (blood borne parasites) in EVL males. Nevertheless, as these sequences were found only in males, this suggested males, and not females, would be the primary source of Plasmodium spp. Furthermore, P. falciparum was recently identified by PCR in faecal samples from gorillas [85] and considering the EVL sampling site had a significant amount of organic matter (Figure 1, Table 1), it is highly feasible that males acquired these microorganisms from human faeces. Alternatively, EVL males could have acquired these microorganisms by contact with other vectors bearing P. falciparum (i.e., Anopheles spp.) during transportation. In any case, these results suggest that, due to their behavioural patterns, Lu. longipalpis could be implicated in the casual dispersal of parasites of medical and veterinary importance.
Human sequences were found in males and females from both locations, whereas chicken sequences were found in NEVL females and EVL males and lizard sequences were only found in NEVL females (Figures 1 and 2). Lu. longipalpis is ubiquitous in dwellings where sanitary conditions are poor and domestic animals, such as dogs, chickens and pigs, are kept in and around the houses. In this kind of environment, the sandfly tends to congregate at outdoor sites, including animal sheds, where leks easily form on abundant, stationary hosts [7], [84]. In this context, as previously mentioned, the EVL location (Posadas) was a worst case-scenario homestead that kept dogs, chickens and a cat, had dense vegetation and a nearby spring (Figure 1, Table 1). In the NEVL location (Lapinha Cave), a chicken was kept to attract sandflies and as a source of food (Figure 1, Table 1). Furthermore, 35 species of lizards can be found in the Minas Gerais region [86]. As female sandflies blood feed on different animal species such as birds, reptiles and humans [5], [82], the presence of these sequences in females was not unexpected. Contrariwise, it was unexpected to find human and chicken sequences associated with males. Nonetheless, this could be due to their previously mentioned behavioural patterns of aggregation and courtship, where male sandflies are often seen over the host where they form leks, attracting females for a blood meal and increasing their chance for mating [83]–[84]. Alternatively, the trap itself was another area of close contact between male and female sandflies and with other potential vectors of medical importance. Consequently, males could have acquired these sequences by contact during transportation.
A total of ten different plant species were identified in males and females from both locations. Capsicum annuum (bell pepper), Elaeis guineensis (African oil palm) and Solanum tuberosum (potato) were identified in three of the four Lu. longipalpis samples. Juglans hindsii (Northern California walnut) was found in both EVL samples and Artemisia annua (sweet wormwood) was found in both male samples. Rhapidophyllum hystrix (needle palm) and Brassica napus (rapeseed) were only found in EVL females and Vitis vinifera (grapevine) was only found in EVL males. Oryza sativa (rice) and Nicotiana tabacum (tobacco) were only found in NEVL females and males, respectively (Figure 2). As adults from both sexes feed on sugars from different plant sources [12], this diversity could be indicative of the different feeding preferences and/or food source availability. Moreover, as sugar meals are not composed primarily by cells, plant RNA could also have originated from other sources. Namely, pollen dispersed by wind could have adhered to the sandflies' hairy surface or, alternatively, these vectors could be casual pollinators during sugar-feeding.
For a more comprehensive understanding of these results, a few limitations of the chosen approach should be considered. In the first place, homology searches are circumscribed to the number and quality of sequences in the databases at the time of analysis. The relatively high number of sequences which showed no significant hits (∼14%) was a clear indication of this. Moreover, if the query corresponds to a given organism that has not yet been sequenced, the hit will probably coincide with the most related organism found in the database. Notwithstanding and given this situation, the results from the homology search will provide a close approximation to the real case-scenario. Another aspect is that, similarly to a previous study [48] and in order to obtain as much environmental data as possible, specimens were neither surface cleaned nor dissected to extract their guts. As they were not surface cleaned, some (or all) of the identified taxa could have been surface contaminants, acquired during transportation by contact with other captured species, i.e. hymenopterans, lepidopterans and mosquitoes, or during manipulation in the lab. In the latter case, even though samples were manipulated with extreme care, this was still a potential source of contamination. Nevertheless, if contamination occurred during manipulation, it was plausible to expect the same contaminating species in males and females from the same location (when specimens were identified and separated according to sex) or from both locations (when total RNA was extracted). The only species present in all four samples was Homo sapiens and, consequently, contamination during manipulation was a possibility for these reads. Notwithstanding, as some biological control agents act by surface contact, such as Beauveria bassiana [39], [87], and since the ultimate goal of this study was to identify possible biological control agents for this neglected infectious disease vector, had the specimens been surface cleaned, this information could have been lost together with other valuable environmental data. On the other hand, as the gut was not separated from the rest of the specimen and, consequently, they were not analysed independently, it was not possible to classify the observed taxa in putative surface contaminants and gut inhabitants, among others. Therefore, the possible role of putative permanent gut residents could not be inferred, such as influence on the insect development cycle or on the parasite transmission ability. Nevertheless, even a careful extraction process would not preclude the possibility of cross-contamination between the gut and the rest of the specimen and/or loss of information. In this sense, the chosen approach ensured that no data was lost and, notwithstanding the aforementioned limitations, enabled the identification of taxa that could putatively influence sandfly development and which have become the target of ongoing studies to determine their significance and location in the sandfly.
Finally, the diversity of bacterial, fungal, protist, plant and metazoan sequences found in this study in wild adult Lu. longipalpis from endemic and non-endemic locations, mostly confirmed their feeding habits and behavioural patterns. Nevertheless, it also suggested that these vectors could possibly be a chance source of dispersal of various animal and plant diseases, such as coccidiosis and malaria. This is particularly significant since the geographical distribution of this vector is undoubtedly expanding [7]. The fact that RNA was obtained from these animal and plant pathogens would indicate that they were biologically active, but this cannot be determined with the present results and further studies must be performed to establish the significance of these findings. The identification of gregarines in wild Lu. longipalpis specimens could indicate that these parasites are a more efficient control method under natural conditions than what was previously suggested [81]. This is specially meaningful as studies on biological control of phlebotomines are still scarce and its practical application seems to be limited to the adult VL vector stage [16]. The employment of biolarvicides in the field is difficult due to the diversity of habitats in which this vector can reproduce and evidence that Lu. longipalpis larvae appear to be thinly dispersed and not concentrated in any particular microhabitat [17]. Nevertheless, as the number of samples analysed in this study was limited, a greater number of specimens must be studied to establish the significance of these results. Current studies are underway to analyse the presence and establish the significance of the taxa found in this study in a greater number of adult male and female Lu. longipalpis samples from endemic and non-endemic locations. A particular emphasis is being given to those taxa implicated in the biological control of this vector and to the etiologic agents of animal and plant diseases.
|
10.1371/journal.pgen.1007519 | The lncRNA male-specific abdominal plays a critical role in Drosophila accessory gland development and male fertility | Although thousands of long non-coding RNAs (lncRNA) have been identified in the genomes of higher eukaryotes, the precise function of most of them is still unclear. Here, we show that a >65 kb, male-specific, lncRNA, called male-specific abdominal (msa) is required for the development of the secondary cells of the Drosophila male accessory gland (AG). msa is transcribed from within the Drosophila bithorax complex and shares much of its sequence with another lncRNA, the iab-8 lncRNA, which is involved in the development of the central nervous system (CNS). Both lncRNAs perform much of their functions via a shared miRNA embedded within their sequences. Loss of msa, or of the miRNA it contains, causes defects in secondary cell morphology and reduces male fertility. Although both lncRNAs express the same miRNA, the phenotype in the secondary cells and the CNS seem to reflect misregulation of different targets in the two tissues.
| In many animals, the male seminal fluid induces physiology changes in the mated female that increase a male’s reproductive success. These changes are often referred to as the post-mating response (PMR). In Drosophila, the seminal fluid proteins responsible for generating the PMR are made in a specialized gland, analogous to the mammalian seminal vesicle and prostate, called the accessory gland (AG). In this work, we show that a male-specific, long, non-coding RNA (lncRNA), called msa, plays a critical role in the development and function of this gland, primarily through a microRNA (miRNA) encoded within its sequence. This same miRNA had previously been shown to be expressed in the central nervous system (CNS) via an alternative promoter, where its ability to repress homeotic genes is required for both male and female fertility. Here, we present evidence that the targets of this miRNA in the AG are likely different from those found in the CNS. Thus, the same miRNA seems to have been selected to affect Drosophila fertility through two different mechanisms. Although many non-coding RNAs have now been identified, very few can be shown to have function. Our work highlights a lncRNA that has multiple biological functions, affecting cellular morphology and fertility.
| Recent studies have shown that the genomes of many higher eukaryotes contain a large number of non-coding transcripts [1] [2] [3] [4] [5] [6]. Elucidating the function of these “non-coding” transcripts is now the topic of intense research. So far, much of the research done on these non-coding RNAs (ncRNAs) has concentrated on one class of small ncRNAs, called microRNAs (miRNAs). miRNAs are short, 22-nucleotide-long RNAs that guide Argonaute family proteins to target mRNAs via anti-sense base-pairing in order to repress the mRNAs’ expression post-transcriptionally [7–9]. Hundreds of miRNAs are encoded in the genomes of most, higher eukaryotes (~400 in humans, 140 in Drosophila, 110 in C. elegans), and many miRNAs are conserved through evolution [8]. Although it is estimated that >60% of protein-coding transcripts are regulated by at least one miRNA [10], removal of individual miRNAs rarely show an overtly visible phenotype [11, 12] [8, 13, 14]. Recent studies in Drosophila, however, has shown that the removal of many miRNAs result in subtle behavioral phenotypes that had previously gone unnoticed [15] [16]. It is now clear that, as a family, miRNAs can play subtle but important roles in a diverse number of processes ranging from development to aging [17–21].
Another class of ncRNAs is comprised of non-coding transcripts with lengths of more than 200 nucleotides, called long non-coding RNAs (lncRNAs) [2, 22, 23]. Even though lncRNAs represent the largest class of non-coding transcripts [1], with a few exceptions, the functions of most lncRNAs remain largely unknown [22, 23]. Genetic manipulation of model organisms has proven to be a powerful tool to discover the functions of ncRNAs [12]. Previously, we, and others, characterized a particular lncRNA in Drosophila, the iab-8 lncRNA [24–29]. This lncRNA is transcribed from the Drosophila bithorax complex, primarily in the posterior central nervous system, beginning in early development. It is over 90kb in length and is both spliced and polyadenylated (Fig 1). One function of this lncRNA is to act as a template for the production of a miRNA that is encoded within its intronic sequence. This miRNA, miR-iab-8, has been characterized as primarily targeting the homeotic genes, abd-A and Ubx, along with their cofactors, hth and exd [24, 26, 27, 29]. The biological consequence of the loss of the iab-8 lncRNA is male and female sterility, thought to be due to defects in the innervation of the abdominal and/or reproductive tract muscles of the fly [24]. These reproductive defects were shown to be the result of overexpression of the Hox targets of the iab-8 miRNA [9, 29].
In 2011, Graveley et al, described a transcript that appeared to be a variant of the iab-8 lncRNA. This lncRNA starts from a promoter located just downstream of the Fab-7 boundary (within the iab-6 cis-regulatory domains that controls Abd-B) and, like the iab-8 lncRNA, also contains the precursor of mir-iab-8 [30]. As this transcript was expressed exclusively in the adult male abdomen, they named this lncRNA male-specific abdominal (msa) (Fig 1). Here, we show that msa is expressed in the secondary cells of the male accessory gland and that its function is required for secondary cell development and maximal male fecundity.
Like the mammalian seminal vesicle and prostate gland, the Drosophila male accessory glands make many important components of the seminal fluid [31]. Each gland is composed of a monolayer of secretory cells surrounding a central lumen. There are two types of binucleate secretory cells in the accessory gland. Ninety-six percent of the cells are “main cells”; the remaining four percent are “secondary cells” [32] [33].
Seminal fluid proteins that are produced by the accessory gland increase male reproductive success by inducing post-mating responses (PMR) in mated females. The PMR is the suite of the behavioral and physiological changes that occur in the female after mating, and include, among many other things, an increase in egg laying/production and a rejection of the courtship by subsequent males [31, 34]. PMR phenotypes that last longer than the first ~2 days post-mating are called the long-term PMR (LTR). Recently, we, and others, showed that the secondary cells produce proteins that are essential for the LTR [35, 36] [37, 38].
Here, we report that the msa lncRNA is expressed in the secondary cells. We show that expression of this lncRNA, and of the miRNA (miR-iab-8) encoded in one of its introns, is required for secondary cell development and thus for the male’s ability to induce long-term post-mating responses in his mate. Interestingly, the major targets for this miRNA in the secondary cells do not include some of the miRNA’s known targets in the CNS. Thus, a single miRNA plays two roles in the process of male fertility but probably through two different mechanisms in the two tissues.
Graveley et al. (2011) isolated a male-specific transcript from abdominal tissue whose promoter mapped to the iab-6 domain of the Drosophila bithorax complex [30]. Examining this sequence more carefully, we discovered that the first exon of msa fell within an enhancer region that we previously demonstrated to be required for Abd-B expression in the secondary cells of the male accessory gland (Fig 1) [35]. Loss of this enhancer (iab-6cocu) eliminates ABD-B expression from these cells and causes both cytological and reproductive phenotypes [35, 36]. We were able to confirm the role played by Abd-B in this process by showing that the iab-6cocu phenotype could be partially rescued using an Abd-B expressing transgene and that an Abd-B RNAi construct was able to create an iab-6cocu-like phenotype when expressed in the secondary cells. Interestingly, in those experiments both the rescue and RNAi-induced phenotype were weaker than expected [35]. As the iab-6cocu mutation deletes the promoter and first exon of msa, we asked if the incomplete nature of both the rescue and RNAi phenotype could be due to a role of the msa lncRNA (see Fig 1).
The iab-6cocuD1 mutation is a 1.1kb deletion of the iab-6cocu enhancer (Fig 1) and is the smallest iab-6cocu deletion that we have made. Secondary cells of iab-6cocuD1 males display abnormal cytological phenotypes (Figs 2B and S2B) and lack ABD-B expression (see S1 Fig). In wild type accessory glands, the secondary cells have a characteristic morphology: they are round and contain a substantial number of large vacuolar structures within their cytoplasm (Figs 2A and S2A) [33] [35]. In iab-6cocuD1 males, the secondary cells are less round (often appearing more hexagonal (main cell-like) in shape) and lose or severely reduce the size of the vacuolar structures (Figs 2B and S2B). In order to functionally test if the msa transcript is made in the secondary cells and to determine if its loss results in a secondary cell phenotype, we tested if mutations that should disrupt the msa transcription unit led to morphological phenotypes in the secondary cells. To do this, we examined the accessory glands of flies containing different BX-C mutations (Fab-864, iab-611, and iab-4186) [39] [40] [41] over a deficiency of the whole BX-C [Df(3R)P9] [42] (Figs 1, 2 and S2). Two mutations (iab-4186A and iab-611) are chromosomal rearrangements with breaks that lead to truncated versions of the msa transcript. The Fab-864 mutation removes the previously characterized promoter of the iab-8 lncRNA and should indicate if the larger iab-8 lncRNA is functionally important in the secondary cells [27]. To more easily visualize the potential phenotypes, all experiments were performed in the presence of a GFP reporter that is secondary cells specific in the AG (Abd-B-Gal4, UAS-GFP referred as to AGFP in Fig 2). In secondary cells, this reporter fills the nuclei and cytoplasm with GFP, but is excluded from the vacuoles [35].
Our results show that both the iab-611and iab-4186 breaks show variably expressed defects in secondary cell morphology (vacuolar structures become smaller) when placed over a BX-C deficiency (Figs 2C, 2D, S2C and S2D) but that secondary cells from Fab-864/ Df(3R)P9 males appear normal (Figs 2E and S2D). The fact that the Fab-864 mutation did not have an effect on secondary cell morphology (Fig 2E) indicates that the full-length iab-8 lncRNA probably does not play a critical role in secondary cell development. However, the effects of the iab-611and iab-4186 breaks are consistent with the msa lncRNA playing a crucial role in secondary cell development.
To verify that the msa transcript is indeed expressed in the secondary cells, we placed an mCherry reporter within its second exon, and examined mCherry expression. Throughout the lifecycle of transgenic flies carrying this reporter, we detected mCherry only within the secondary cells of the adult male’s AG (Fig 3A), indicating that msa’s expression is secondary cell-specific.
As the msa transcript shares most of its sequence with the iab-8 lncRNA, it should contain the precursor of the miR-iab-8 miRNA between its fifth and sixth exons. We reasoned that if the msa lncRNA is important for secondary cell morphology then some of that function could be due to expression of miR-iab-8. If this were so, then we would expect that mutations that truncate the msa transcript downstream of the miRNA should show little or no morphological defects. To test this, we examined the iab-386A break, which breaks the transcript just downstream of the miRNA (Fig 1). Flies carrying the iab-386A break over a BX-C deficiency do not show an iab-6cocuD1–like secondary cell morphological phenotype (small or missing vacuoles), suggesting that miR-iab-8 or an element located between the iab-386A and the iab-4186 breakpoints, is critical for secondary cell morphology (Figs 2F and S2F).
To show that miR-iab-8 is expressed in secondary cells, we used a GFP miRNA sensor, specifically designed to detect miR-iab-8 [25] [29]. This miRNA sensor contains binding sites for miR-iab-8 in the 3’UTR of a ubiquitously expressed GFP reporter. Thus, in cells in which miR-iab-8 is expressed, the GFP reporter is silenced, leading to GFP-negative cells in an otherwise GFP-positive background. Examining the secondary cells of flies carrying the GFP sensor showed that within the AG, only the secondary cells lack GFP expression. This indicates that miR-iab-8 is indeed expressed in the secondary cells of the male accessory gland (Fig 3B). We next tested if the GFP sensor is affected in the iab-6cocuD1 mutation that removes the msa promoter and first exon. As seen in Fig 3C, iab-6cocuD1 mutants display high levels of GFP expression throughout the accessory gland, indicating that miR-iab-8 expression is abrogated or severely impaired in these mutant cells.
From these results, we conclude that the msa lncRNA is expressed in the secondary cells of the male AG through its promoter in the iab-6cocuD1 region. Furthermore, given that we previously showed that deletions removing a large portion of the msa transcript upstream of the miRNA (but leaving its promoter intact, like iab-4,5,6DB [35] [43]) causes no noticeable secondary cell phenotype, we believe that msa probably fulfills most of its function in these cells via the miR-iab-8 miRNA embedded within its sequence.
Our findings suggest that the iab-6cocuD1 deletion is actually a double mutant in the sense that it affects both Abd-B and msa/miR-iab-8 expression. We thus wanted to know how much of the iab-6cocuD1 phenotype is due to the loss of miR-iab-8. Because homozygous miR-iab-8 mutations cause sterility due to a role in the CNS and affects gene expression in multiple tissues [24], we examined the effects of the loss of miR-iab-8 specifically in secondary cells by placing a mir-iab-8 deletion in trans to iab-6cocuD1. Previous data from our lab indicates that the miR-iab-8 deletion should leave the msa transcript intact; the miRNA precursor is located inside an intron and the closely related iab-8 lncRNA is still made in miR-iab-8 deletion lines [27]. Therefore, these transheterozygotes should be mutant for miR-iab-8 in the secondary cells but still express ABD-B (and the rest of the msa transcript) (S1E Fig). To make any resulting phenotypes easier to see, we again included a secondary cell-specific GFP marker in all flies. The results of these experiments can be seen in Fig 4.
Our results indicate that loss of miR-iab-8 alone significantly impairs secondary cell development, but not as severely as the iab-6cocuD1 mutation itself [Fig 4B and 4C; median vacuole size (area of the vacuole at its widest point) control = 24.43 μm2, iab-6cocuD1 homozygotes = 2.025 μm2, iab-6cocuD1/miR-iab-8 = 3.615 μm2. The two mutant genotypes are significantly different from the control (p < .0001 (Kruskal-Wallis one-way analysis of variation followed by a post hoc Dunn’t test) [44, 45]) and from each other (p = .0085)]. We also assessed the contribution of Abd-B by crossing an Abd-B null allele (Abd-BD16) to iab-6cocuD1. These males also showed reduced vacuole size (Fig 4D) (12.41 μm2 relative to heterozygous cells (p < .0001)), though the effects were milder than those of either iab-6cocuD1 homozygotes or iab-6cocuD1/miR-iab-8 mutants (p < .0001). These results indicate that Abd-B may play a lesser role in secondary cell development. However, the Abd-B locus is a region of the genome that is highly susceptible to transvection events [46] [47] [48]. Indeed, it seems that within the secondary cells, some transvection is occurring, as we observe low but detectable ABD-B levels in iab-6cocuD1/ Abd-BD16 secondary cells (Fig 4E and 4F). Thus, the iab-6cocuD1/Abd-BD16 phenotype may be mitigated by interactions between the Abd-B enhancers on the Abd-BD16 chromosome with the Abd-B promoter on the iab-6cocuD1chromosome.
To further differentiate the contribution of the Abd-B relative to the miRNA in the iab-6cocuD1 phenotype, we examined two additional mutations of the msa transcript created using the CrispR-Cas9 system (Fig 1). Consistent with expectations, in a line where the msa transcript was completely deleted, both Abd-B and miR-iab-8 (based on the miRNA sensor[25]) were completely undetectable in the accessory glands. The phenotype of males heterozygous for this allele validated our hypothesis for the function of the lncRNA: males carrying the msa deletion over iab-6cocuD1 showed a secondary cell phenotype similar to that of iab-6cocuD1 homozygotes, with no large vacuoles (Fig 5A).
In the course of generating the msa deletion, we fortuitously recovered a second mutation that let us further explore the role of the miRNA. This mutation is a near perfect inversion of the intervening 68kb sequence (Fig 1). Surprisingly, males carrying this inversion over iab-6cocuD1 displayed a variable secondary cell phenotype, with some cells looking like iab-6cocuD1 homozygotes and others showing a markedly weaker phenotype (some vacuoles, though mostly smaller) (Fig 5B and 5B’). Abd-B protein was undetectable in all secondary cells from these males (Fig 5E). Interestingly, the GFP-based miR-iab-8 sensor [25] showed that some secondary cells expressed the miRNA, while others did not (Fig 5C and 5D). We then tested whether absence or presence of the miRNA correlated with the severe vs. weak phenotype. Because we could not use our normal GFP reporter to judge the relative severity of the phenotypes in the cells expressing the sensor, we used phase contrast microscopy. As seen in Fig 5F and 5F’, secondary cells that express the iab-8 miRNA still have visible vacuoles (appearing as rougher patches within the cells under phase contrast), whereas no vacuoles are seen in secondary cells that do not contain the miRNA. These results lead to the conclusion that absence of the miRNA, in the absence of Abd-B, leads to a secondary cell phenotype like that seen in iab-6cocu homozygotes, whereas the presence of the miR, in the absence of Abd-B leads to a milder phenotype (some small vacuoles). Overall, these results support our results using the Abd-BD16 mutation that suggested that Abd-B plays a significant, though perhaps lesser role in secondary cell development.
After mating to wild type males, females exhibit characteristic changes in behavior called the post-mating response (PMR) (reviewed in [31]). These changes in behavior include an increase in egg production and deposition, as well as a decrease in receptivity to remating. Persistence of these changes for >1 day post-mating, is due, in part, to the action of secondary cell produced proteins [35]. To determine if the PMR changes seen in iab-6cocuD1 mutants are mediated by the iab-8 miRNA, we compared the long-term post-mating responses (LTR) induced by iab-6cocu/miR-iab-8 males to both iab-6cocu homozygous males and heterozygous control males. We assessed the LTR using two assays: (1) we tested for the long-term increase in egg-laying and (2) we tested for the long-term decrease in remating receptivity. Because the quantitative levels of these phenomena can be sensitive to genetic background, we carried out assays on both iab-6cocuD1 and a second iab-6cocu mutation, iab-6cocuD5 (the original iab-6cocu deletion that is 2kb larger than iab-6cocuD1) [35].
In the fertility/fecundity assay (Fig 6A and 6B), mates of transheterozygous iab-6cocu/miR-iab-8 mutants showed a significantly decreased LTR relative to mates of heterozygous control males: mates of miR-iab-8 deficient males laid significantly fewer eggs after the initial short-term increase in egg-laying. This phenotype in the mates of iab-6cocu/miR-iab-8 transheterozygous males was identical to that of females mated to homozygotes for either of the iab-6cocu mutants.
Results with the receptivity assay were also consistent with the finding that the LTR is disrupted in mates of mir-iab8 transheterozygous males (Fig 6C and 6D.) Mates of either transheterozygous iab-6cocu/miR-iab-8 mutants showed a significant increase in receptivity over heterozygous control males in the four-day receptivity assay, although there were slight numerical differences between the results with the two iab-6cocu alleles; mates of iab-6cocuD5/miR-iab-8 males showed a stronger effect than mates of iab-6cocuD1/miR-iab-8.
Our results indicate that miR-iab-8 plays a major role in the male’s ability to induce the LTR in his mate. Previous RNAi results [35] showed that Abd-B also plays an important role in the male’s ability to induce an LTR. Thus, we conclude that the iab-8 miRNA and Abd-B both play a role in the male’s ability to induce the LTR in his mate.
The PMR experiments highlight the role of the secondary cells in propagating the long-term response. A number of proteins have been implicated in regulating the LTR. Two of these proteins are CG1652 and CG1656 (annotated in flybase as Lectin-46Cb and Lectin-46Ca, repectively) [49, 50]. Previously, we showed that these proteins migrate differently on SDS-PAGE gels when isolated from iab-6cocu mutants or wild-type AGs [35]. In order to test if this is also true for iab-6cocu/miR-iab-8 transheterozygous males, we performed Western blot analysis on AG extracts from wild type, iab-6cocuD1 homozygous and iab-6cocuD1/miR-iab-8 transheterozygous males. As seen in Fig 6E, the CG1652 and CG1656 proteins migrate faster when isolated from either iab-6cocuD1 homozygous males or iab-6cocuD1/miR-iab-8 transheterozygous males relative to these proteins from extracts of wild type males. In fact, we noticed that both the CG1652 and CG1656 proteins seem to migrate slightly faster in extracts from iab-6cocu/miR-iab-8 AGs than from iab-6cocuD1AGs (Fig 6E). In Gligorov et al., we showed that the genotype specific difference in CG1656 migration was due to changes in N-linked glycosylation; treatment of extracts with PGNase F led to proteins that migrated at equal velocities [35]. We repeated these treatments on extracts from iab-6cocuD1/miR-iab-8 transheterozygous AGs. Although PGNase treatment equalized CG1656 migration from in iab-6cocuD1 and wild-type extracts, CG1656 from iab-6cocuD1/miR-iab-8 transheterozygous AGs continued to migrate more rapidly (Fig 6F). These results suggest that there is a difference in secondary cell characteristics between the two mutant genotypes.
The strongest known target of the iab-8 miRNA is the abd-A gene [24, 26, 27]. Indeed, in the embryonic CNS, it has been shown that the iab-8 lncRNA represses the expression of the abd-A gene through two mechanisms: through the iab-8 miRNA and through apparent transcriptional interference at the abd-A promoter [27]. iab-8 lncRNA mediated abd-A repression in the CNS has been implicated in the female sterility phenotype associated with removal of the iab-8 lncRNA [27]. In order to determine if the msa transcript is important for regulating abd-A in the secondary cells, we stained for ABD-A protein in iab-6cocuD1 homozygous accessory glands. Fig 7 shows that in iab-6cocuD1 homozygous accessory glands, both ABD-A and ABD-B are undetectable. This is consistent with our RNA-seq results that show that the small amounts of abd-A transcript seen in wild type AGs drops ~4.5-fold in iab-6cocuD1mutants [36]. Thus, ABD-A does not seem to be regulated by the msa transcript in the AG and therefore, misexpression of ABD-A is not responsible for the iab-6cocu phenotype.
Other known targets of miR-iab-8 are the transcripts encoding the hox protein, ULTRABITHORAX (UBX) and its general cofactor EXTRADENTICLE (EXD) [29]. Both of these proteins have been shown to be misregulated in the CNS in miR-iab-8 mutants and have been shown to be involved in the sterility phenotype associated with this deletion [29]. As our RNA-seq data suggest that both Ubx and Exd are transcribed in the accessory gland [36], we decided to check for the expression of both of these proteins in the secondary cells of wild type and iab-6cocuD1 mutants.
Immunostaining for either UBX and EXD failed to detect nuclear localized UBX or EXD protein in the secondary cells of both genotypes. While UBX protein could not be detected at all, EXD staining revealed a punctate cytoplasmic staining (Fig 8). Although we cannot rule out that the EXD staining is simply background, the conditions used are able to visualize EXD protein in control tissues (S3 Fig). We did not further investigate EXD as a potential target as the function of EXD is known to be nuclear and its nuclear localization is known to depend on its partner HTH [51] [52] [53] [54] [55]. Since our RNA-seq results indicate that HTH is not expressed in the secondary cells of wild type or iab-6cocuD1 mutants [36], any EXD misregulation is unlikely to cause the iab-6cocuD1 mutant phenotype.
Overall, from our ABD-A, UBX and EXD staining results, it seems that the known targets of the iab-8 miRNA are not the cause of secondary cell defects seen in iab-6cocuD1 mutants. Given that the overexpression of these genes has been directly linked to the CNS mediated sterility phenotypes, we conclude that the primary targets of the iab-8 miRNA are probably different in the CNS and the accessory glands.
Previously, we showed that the iab-6cocuD1 mutation removes a secondary cell specific enhancer for Abd-B and that this mutation causes the loss of Abd-B expression in the secondary cells. Furthermore, this mutation causes both a cytological phenotype in these cells and a reduction in the long-term post-mating response (LTR) in the mates of these mutant males [35]. Sequence analysis of a lncRNA discovered by large-scale RNA-seq [30], called male-specific abdominal (msa), indicate that the ~1.1 kb iab-6cocuD1 deletion also removes the promoter and first exon of the msa lncRNA. By genetic and molecular analyses, we show that the expression of the msa lncRNA allows the mir-iab-8 miRNA to be produced in the secondary cells and that this miRNA is responsible for some of the male’s ability to induce an LTR in his mate. Based on the data presented here, and from our previous work, it is clear that both Abd-B and the miRNA play overlapping, but not completely-redundant roles in secondary cell development; reduction of either Abd-B or the mir-iab-8 miRNA alone show weaker phenotypes than the removal of both elements together (Fig 4, also see [35]).
That there is some non-redundancy between the effects of the ABD-B transcription factor and the mir-iab-8 miRNA in secondary cells can also be seen with the LTR-mediating proteins CG1652 and CG1656. In both iab-6cocuD1 and iab-6cocuD1/miR iab-8 mutants, we observe shifts in the gel-migration of these proteins relative to wildtype. However, both proteins migrate slightly differently in extracts of iab-6cocuD1/miR iab-8 vs. extracts of iab-6cocuD1 homozygotes, indicating differences between these genotypes. We previously linked the subtle shifts in the migration pattern of CG1656 in iab-6cocu mutants to changes in N-linked glycosylation [35]. PGNase F treatment of extracts from iab-6cocuD1 and iab-6cocuD1/miR iab-8 mutants does not remove the subtle difference in SDS-PAGE migration, thus further highlighting the differences between the two genotypes and likely, the role of Abd-B vs the miRNA.
Earlier work from our lab and others have shown that miR-iab-8 is important for male and female fertility through its role in the developing CNS [24] [27, 29]. In the early posterior CNS, miR-iab-8 is important for the repression of specific homeotic genes and homeotic gene cofactors (abd-A, Ubx, exd and hth). Failure to repress these targets in the posterior CNS leads to male and female sterility, as males lack the ability to curl their abdomens for mating and females lack the ability to lay eggs [24]. These phenotypes seem to be due to CNS defects that prevent proper innervation of particular abdominal muscles [24] [29]. The results we present here suggest that miR-iab-8 has at least some different primary targets in the secondary cells that are also needed for fertility. We have not yet been able to determine a primary targets for miR-iab-8 in the secondary cells, though we have examined a number of genes that are upregulated in iab-6cocu mutants and contain predicted miRNA binding sites [36]. Given that many miRNA loss-of-function phenotypes are thought to be caused by mildly affecting the expression level of many target genes [8, 56], our lack of success in finding a primary target could indicate that removing miR-iab8 causes defects in the secondary cells through the mild overexpression of a network of targets.
As mentioned above, we now know that the iab-8 miRNA plays a dual role in male fertility. One role is in the CNS and is accomplished through the regulation of the hox genes and their cofactors. Here, we have described a second role for mir-iab-8; in the secondary cells of the male AG, it plays an important fertility function that seems to be through the regulation of a different set of genes. Given the high conservation of mir-iab-8 in arthropod hox complexes [57] and the conservation of its target sites in the homeotic genes and their effectors [58], we believe the homeotic function of the miRNA is likely its ancestral role. In this context, it is intriguing to speculate on the genesis of the msa transcript. In both flies and humans, male gonadal tissues have the highest levels of ncRNA expression [59] [60]. This has been suggested to reflect the high concentration of transcription factors in this tissue regulating cryptic promoters in intergenic regions [61] [62]. If the ncRNA provided advantages in fertility [59] [62–64] [65] it could be selected; perhaps such a scenario led to selection for secondary cell expression of msa.
The msa promoter seems to be tied to an Abd-B enhancer. Interestingly, Abd-B class hox genes have evolutionarily conserved functions in the male reproductive tissues. For example, egl-5, the C. elegans Abd-B ortholog is expressed in the male worm seminal fluid producing organs and is sufficient to induce markers associated with the male-specific, seminal fluid-producing cell fates in hermaphrodites [66]. In mammals, Abd-B class hox genes have been shown to be expressed in the male seminal fluid creating organs like the prostate and the seminal vesicle and have been shown to be critical for the production of secreted gene products. [67, 68]. Based on these conservations, it may be that the accessory gland function of msa/miR-iab-8 arose from the co-opting of the Abd-B secondary cell enhancer (iab-6cocuD1) by a neighboring, potentially cryptic, promoter. As the creation of the msa transcript would not disturb hox gene regulation in the secondary cell (since its normal targets do not seem to be expressed in these cells), its appearance could have been tolerated, adding increased genetic flexibility for selection. One can imagine that transcripts of secondary cell-expressed genes whose repression was beneficial to male fertility might then have acquired/retained regulation by miR-iab8. Based on this, it would be particularly interesting to look for species that do not express the msa lncRNA and to examine predicted accessory gland targets for changes in mir-iab8 binding sites.
Flies were raised at 25°C on standard yeast-cornmeal-agar or yeast-glucose-agar media and crossed using standard fly methods. Mutants and transgenic constructs used for this study include: Δmir-iab-8121−8[24], iab-6cocuD1, iab-6cocuD5, Abd-B-Gal4, UAS-nGFP/CyO; iab-6cocuD1 (this study), Df(P9) (Lewis, 1978), iab-4186, iab-611 [39], Abd-BD16[69], miR-iab-8 sensor [29], miR-iab-8 sensor; iab-6cocuD1, Fab-864 [41], iab-386A[40], iab-5,6rescue [70].
All molecular biology was performed using enzymes from NEB (Ipswich, MA USA) or Promega (Promega). The iab-6cocuD1 mutation was created in a fashion similar to the original iab-6cocu mutation in Gligorov et al. [35, 36] [35]. In short, a kanamycin cassette was amplified using the following primers: FI (5’-GGCAGCACGAATAGTTTAGTTTATTTTAGCCATAGCTCAAGAACGACAGCGAATACAAGCTTGGGCTGCAGG-3’) and RI (5’-GGTGAATAATTTTTATTGCCGTAAATCACTGTGTCAATTGTGGTTGTAATCTCGCCCGGGGATCCTCTAGAG-3’). This cassette was then used to recombineer a plasmid that could be used for the InSIRT technique to target the iab-6 domain for site specific mutation [70].
The mCherry reporter in exon 3 of the iab-8 lncRNA was made using by site specific integration into an attP site within the BX-C that removes the iab-6 domain (iab-5,6C.I) [70]. The integration construct was made by two successive recombineering steps using the Counter-Selection BAC Modification Kit (Gene Bridges) on pKsY-iab6H [70]. Amplification of the rpsl-neo cassette with homology regions flanking exon 3 was performed using the primers: Rpsl-neo F: TATACTTTATGCCCTTCCAGTTTGATTACACATCGACCCCTGGAGCGAGCCAAACGGCCTGGTGATGATGGCGGG and Rpsl-neo R:TATGAAATATGTTAAGATGGAGACTCAC-CTGATGCAGCTGCCGTCGGGTTAAGTCTCAGAAGAACTCGTCAAGAA. The resulting PCR fragment was used to insert an rpsl-neo cassette into exon 3. Next, mCherry was amplified using the primers: MCherry F: TATACTTTATGCCCTTCCAGTTT-GATTACACATCGACCCCTGGAGCGAGCCAAACGGTACCATGGTGAGCAAGGG and MCherry R: TATGAAATATGTTAAGATGGAGACTCACCTGATGCAGCTGCCGTCGGG-TTAAGTCGCCCCAAGGGGTTATGCTAG. The resulting PCR fragment was used to recombineer the rpsl-neo exon 3 construct to replace the rpsl-neo with the exon-3-mCherry fusion. This plasmid was then integrated into the bithorax complex using the InSiRT technique [70].
msa deletions and inversions were created using the CRISPR-Cas9 system. Guide RNAs were created in the embryo from plasmids containing guide sequences and by injecting U6-promoter target guide-gRNA fusion plasmids. The templates for the guide RNA plasmids were created as g-blocks (IDT, Coralville, Iowa USA) and cloned into pGemTeasy (Promega). The target guide sequences are GGTGGCAAAATATCAAACAA(TGG) and GATGAGAGCAATAGTAGAAG(AGG)(PAM sequences in parentheses). Guide plasmids were injected using standard microinjection procedures into vasa-Cas9, lig4169 flies. The vasa-Cas9 transgene was made by PCR. The vasa promoter and 5’ UTR was made by PCR using primers 5’vasapromS: 5’ GATATCTTTGGACACGTGGCATAAACAAGCC 3’ and NcoI vasa5’UTRAS: 5’CCATGGTATTGATATTTTTTTTTTAATTTGGCCTGCCTTTC3’. The vasa 3’UTR was made by PCR using the primers NcoI-ApaI vasa3’UTRS: 5’CCATGGAAGGGCCCAATGTATGGACATAGATTTCAAATAATTAAATGTAATGC3’ and SacI vasa3’UTRAS: 5’GAGCTCAACACGAAGAGCAGCAGTGTGGTGG3’. First the vas promoter/5’UTR was cloned into pGemTeasy. Clones in the proper direction were then cut by ApaI, blunt ended using Mung bean exonuclease (New England Biolabs), then cut with EcoRV and religated. The Vasa3’UTR PCR fragment was then cut with NcoI and SacI. The pGemTeasy vector with the vasa promoter/5’ UTR was cut with NcoI and SacI and the cut vasa 3’UTR fragment was cloned into it. The Vasa 5’3’UTR construct was then placed into pTnT (Promega). The vasa 5’3’ construct was then cut with NotI, blunt ended and religated. The Cas9 coding sequence was excised from plasmid X260 (a kind gift of Feng Zhang)[71] using NcoI and NotI and ligated into the pTnT vasa construct cut with NcoI and PspOMI. The whole vasa-Cas9 cassette was then removed by a partial EcoRI digest and cutting with XhoI. This fragment was then ligated into a pATTB cut with EcoRI and XhoI. The resulting vector was then injected into the ZH2A attP platform (FlyC31.org[72]). After isolating transformants, the lines were recombined with a lig4169 mutation [73].
For imaging accessory glands, 3–5 day old virgin male flies were dissected in Grace’s Insect Media and fixed for 20 minutes using 3.7% formaldehyde in PBST or Grace’s Insect Media. Samples were blocked in PBST buffer containing 5% horse serum. Primary antibodies were incubated in the blocking solution overnight at 4°. Secondary antibodies were incubated for 1–2 hours at room temperature in blocking solution. All washes were done with PBST and generally consist of 3 rinses of 30 seconds followed by 3 washes of 20 minutes. All samples were mounted in Vectashield with DAPI (Vector Labs). Mouse anti-Abd-B (Developmental Systems Hybridoma Bank) supernatant was pre-absorbed against Drosophila embryos and then diluted to 1:100 before use. Goat anti-ABD-A (Santa Cruz Biotechnology) and Goat anti-Exd (Santa Cruz Biotechnology) were both used at a 1:50 dilutions. Mouse Anti-Ubx (Developmental Systems Hybridoma Bank) was used at a 1:10 dilution. Secondary antibodies were Alexa Fluor 488 or 555 antibodies (Invitrogen AG) and were used at a 1:1000 dilution. GFP fluorescence was visualized directly.
Microscopy was performed on a Zeiss Axioplan fluorescence microscope using an X-Lite 120 lamp or a Zeiss LSM 700 confocal microscope.
Accessory glands were dissected from 3–5 day old virgin males, placed into tubes with SDS-sample buffer, homogenized, and boiled for 5 min. Extract representing the equivalent of one male’s accessory glands was loaded onto 12% or 4–20% SDS-PAGE gels, purchased from (Invitrogen/Thermo Fisher Scientific). The proteins were run slowly (20–30 volts) through the stacking gel and first half of the separating gel. Later, when the proteins were in the separating gel, the voltage was turned up to at 100V. Transfer of the proteins onto PVDF membranes was performed using a Biorad wet transfer cell for 1 hour at 100V. Antibodies used for the western blots were affinity purified rabbit anti-CG1656 (1:500 dilution) [74], affinity purified rabbit anti-CG1652 (1:250) [74] and rabbit anti-Acp36DE (1:30000)[75]. Secondary antibodies were goat anti-rabbit antibodies conjugated to either alkaline phosphatase (Biorad) or horseradish peroxidase (Promega). Signals for the alkaline phosphatase conjugated antibodies were visualized according to the NBT/BCIP staining kit (Roche), while peroxidase conjugated antibodies were visualized according to the Supersignal West Pico PLUS Chemiluminescent Substrate kit (Invitrogen/Thermo Fisher Scientific). PGNase treatments of AG protein extracts were performed as in Gligorov et al. {Gligorov, 2013 #176}
Three to four day old virgin males, raised at 25°C of the following genotypes were collected: Abd-B-Gal4, UAS-nGFP/CyO; iab-6cocuD1[36], Abd-B-Gal4, UAS-nGFP/+; iab-6cocuD1/ +, Abd-B-Gal4, UAS-nGFP/+; iab-6cocuD1/Δmir-iab-8121−8 and Abd-B-Gal4, UAS-nGFP/+; iab-6cocuD1/Abd-BD16. The accessory glands were dissected in Grace’s Insect Media and fixed in 3.7% formaldehyde for 20 minutes. After washing with PBST, the accessory glands were mounted on slides in vectashield mounting medium. Vacuoles were visualized using GFP fluorescence, captured by a Leica LSM700 confocal microscope.
Vacuoles were measured using the ImageJ software; the stacks were examined and the area of every vacuole seen in a cell was measured in the stack at its point of largest diameter. For each genotype, the vacuoles of whole cells were measured from a number of different glands. The recorded data were then statistically analysed using Kruskal-Wallis analysis of variance with post-hoc Dunn’s tests using the Prism 7.0 software (Graphpad Software, Inc.).
Fertility/fecundity assays were performed as previously described [35] with Wilcoxon non-parametric tests used to compare results for mates of different genotypes in total and on individual days. rmANOVA was used to evaluate overall 10-day trends. All statistical analysis was performed with the JMP9 software [76]. For each genotype, mates of between 15 and 22 males were tested (see figure legend for exact numbers).
Receptivity assays were performed as described in [35]. Comparisons of remating frequencies between females mated to different genotype males was evaluated using a Wilcoxon ranked sums test (WRST) using JMP9 software [76]. For each genotype, mates of between 15 and 22 males were tested (see figure legend for exact numbers).
|
10.1371/journal.pgen.1006521 | Tbx5 Buffers Inherent Left/Right Asymmetry Ensuring Symmetric Forelimb Formation | The forelimbs and hindlimbs of vertebrates are bilaterally symmetric. The mechanisms that ensure symmetric limb formation are unknown but they can be disrupted in disease. In Holt-Oram Syndrome (HOS), caused by mutations in TBX5, affected individuals have left-biased upper/forelimb defects. We demonstrate a role for the transcription factor Tbx5 in ensuring the symmetric formation of the left and right forelimb. In our mouse model, bilateral hypomorphic levels of Tbx5 produces asymmetric forelimb defects that are consistently more severe in the left limb than the right, phenocopying the left-biased limb defects seen in HOS patients. In Tbx hypomorphic mutants maintained on an INV mutant background, with situs inversus, the laterality of defects is reversed. Our data demonstrate an early, inherent asymmetry in the left and right limb-forming regions and that threshold levels of Tbx5 are required to overcome this asymmetry to ensure symmetric forelimb formation.
| Externally, the human form appears bilaterally symmetric. For example, each of our pairs of arms and legs are the same length. This external symmetry masks many asymmetries found in internal organs. In most people the heart is found on the left side of the chest. The stomach, liver and spleen are also positioned asymmetrically. The authors of this study demonstrate, using a mouse model, that bilateral symmetry of the arms is not a default, passive state but that mechanisms are in place that ensure symmetrical formation of the left and right limbs. Bilateral symmetry of the arms is achieved by the action of a gene Tbx5 that masks the effects of signals that acted earlier during embryogenesis, many days before limb formation, and imposed asymmetries on the forming internal organs. Maintaining bilateral symmetry of the arms is important for them to carry out their normal functions but this process can go wrong. Holt-Oram syndrome patients have upper limb defects, including shortened arms. Consistently the defects are more severe in their left arm than right. This birth defect is caused by disruption of the TBX5 gene. By linking the action of Tbx5 to symmetrical limb formation, the authors provide an explanation for why Holt-Oram syndrome patients have more severe defects in the left arms than right.
| The external body plan of most metazoans is bilaterally symmetric. How this symmetry is achieved has fascinated biologists for centuries and is exemplified by Leonardo Da Vinci’s description of the “Vitruvian Man” that describes some of the uniform proportions of the human body. In vertebrates, external bilateral symmetry masks significant internal asymmetries such as the position of the heart and liver and number of lobes of the lungs. While there has been progress in identifying how asymmetry of internal organs is generated [1], we know very little about how symmetry in bilateral structures, such as the limbs is established. A classic study that compared the lengths of the skeletal elements of left and right embryonic chick wings found there was very little difference in their size [2] demonstrating the exquisite fidelity in the genetic programmes regulating limb development, a finding all the more intriguing by the lack of evidence for molecular crosstalk between the developing left and right limb buds.
Holt Oram Syndrome (HOS) [OMIM 142900] is caused by mutations in TBX5, a T-box transcription factor expressed in the forelimb and heart. The clinical features of HOS include heart defects and a spectrum of upper/fore limb defects ranging in severity from total aplasia of the radial elements to relatively mild defects such as a tri-phalangeal thumb [3]. A characteristic feature of HOS patients is that although the severity of limb defects can vary between patients, the left limb is more severely affected than the right in most patients [3]. HOS is a dominant disorder and heterozygous affected individuals carry mutations in TBX5 predicted to result in truncated proteins that fail to fold or are rapidly degraded [4] and are most likely loss-of-function alleles. HOS defects are therefore thought to arise as a result of TBX5 haploinsufficiency and indicate that both copies of the gene are required for normal function.
Attempts to recapitulate the upper limb defects associated with HOS in the mouse have previously been unsuccessful. Conditional heterozygous deletion of Tbx5 using the limb-restricted Prx1Cre (Tbx5lox/+;Prx1Cre) does not produce any obvious forelimb phenotype, while homozygous conditional deletion of Tbx5 (Tbx5lox/lox;Prx1Cre) leads to a total loss of forelimb bud initiation and subsequent forelimb formation [5]. We have developed two alternative strategies to model the left-biased, asymmetrical limb defects of HOS patients in the mouse. One strategy utilizes a gene deletion-gene replacement strategy [6] in which Tbx5 is conditionally deleted using the Prx1Cre line and hypomorphic levels of a Prx1-Tbx transgene are expressed in the forelimb regions of the same embryo. The alternative strategy uses a mosaic cre line, Prx1Cre(98), to conditionally delete Tbx5 from cells in the right and left forelimb-forming LPM. Both approaches successfully recapitulate the types of limb abnormalities observed in HOS patients and significantly, the left bias in the severity of the limb defects. We also demonstrate that the disruption of INV, a gene crucial for the establishment of the left-right pathway [7] in the presence of hypomorphic levels of Prx1-Tbx lead to the formation of right sided forelimb defects. Additionally, optimal levels of Fgf10 expression in the presence of hypomorphic Prx1-Tbx levels are not sufficient to rescue symmetrical forelimb bud initiation and formation. These data demonstrate that although the left and right limb ultimately develop to be bilaterally symmetric structures they arise from regions of the left and right embryo flank that have inherent asymmetries. We further show that threshold levels of Tbx5 are required to overcome these asymmetries to ensure bilaterally symmetric forelimb formation.
To generate a model mouse of HOS, we used a gene deletion-gene replacement strategy (S1A Fig) [6] so that hypomorphic levels of a Tbx transgene are delivered to the forelimb-forming region. Previously, we have shown that homozygous deletion of Tbx5 using the limb-restricted Cre-deleter line, Prx1Cre, produces a forelimb-less phenotype [5]. This defect can be completely rescued by misexpression of either Tbx5 or Tbx4 transgenes driven by the same Prx1 regulatory element [8]. One chimeric construct, Prx1-5N5T4C#1, a fusion of the N-terminal and T-Box DNA binding domains of Tbx5 and the C-terminal domain of Tbx4, is also able to rescue forelimb bud formation [8]. These results indicate that Tbx5 and Tbx4 have common functions in initiation of fore and hindlimb outgrowth and Tbx5, Tbx4 and Tbx5/Tbx4 chimera proteins can act equivalently. Another transgenic line, Prx1-5N5T4C#2 (hereinafter referred to as Prx1-Tbx), harbours the same transgene as Prx1-5N5T4C#1 but in a different locus and expresses hypomorphic levels of the chimera protein compared to endogenous Tbx5 (S1B–S1E Fig). This hypomorphic misexpression of the transgene only partially rescues forelimb bud initiation defects caused by the conditional deletion of Tbx5 (Tbx5lox/lox;Prx1Cre;Prx1-Tbx) and recapitulates many features of the upper limb defects seen in HOS patients, for example an anterior bias to the structures affected (including triphalangeal thumb, absent thumb), defects in the scapula and in more severe examples, phocomelia (Fig 1A–1C). Another clinical feature of HOS is the broad range in the severity of limb defects, even within a family carrying the same TBX5 mutation, ranging from almost complete absence of the upper limb to triphalangeal thumb. This is also reproduced in our mouse model (Fig 1C). Despite the range in the severity of the limb defects between different embryos, in individual Tbx5lox/lox;Prx1Cre;Prx1-Tbx mutant embryos the forelimb defects are consistently more severe in the left forelimb than the right. In our mouse model this was observed in 100% of the embryos studied (n = 11) (S2 Table). A transverse section of Z/AP/+;Prx1Cre embryo shows the bilaterally symmetrical Cre activity in the left and right forelimb buds [9] (Fig 1D). Furthermore, we compared Cre mRNA expression levels between the left and right forelimb buds and did not detect statistically significant difference (p > 0.05 with two-tailed Student’s t-test) (Fig 1E). This result is consistent with the fact that the Prx1Cre driver has been used in combination with numerous other conditional mutant alleles and all limb defects reported to date are bilaterally symmetric. In addition, equivalent, bilateral chimera transcript levels are present in Prx1-Tbx transgenic embryos (p > 0.05 with two-tailed Student’s t-test) (Fig 1E), indicating Tbx activity that partially and asymmetrically rescues limb outgrowth is bilaterally symmetric in Tbx5lox/lox;Prx1Cre;Prx1-Tbx mutant embryos.
In a previous study it was calculated that 70% of HOS individuals display asymmetrical upper limb defects, and in 91% of these, the left side is more severely affected than the right [3]. In contrast, 100% of the Tbx hypomorphic mutant mice we obtained showed left-biased defects. A possible explanation for this discrepancy in penetrance is that the original analysis of HOS patients was carried out before genetic lesions in TBX5 were identified as responsible for HOS and some of the original patients may not have had pathogenic TBX5 mutations. We searched the electronic patient record at Great Ormond St Hospital and identified 32 patients with a clinical suspicion of Holt-Oram syndrome. In 11 cases TBX5 mutations had been identified, 8 of which were pathogenic mutations (S1 Table). 7 cases show left-biased forelimb defects and one case is symmetrical (Fig 1G–1K) [10–13]. We did not observe right-biased defects in any patients with confirmed pathogenic TBX5 mutations. Therefore in our patient series when pathogenic mutations in TBX5 have been confirmed the clinical presentation includes left-biased severity with almost 100% penetrance consistent with what we observed in our mouse model.
During forelimb bud initiation, Tbx5 induces the expression of Fgf10 in the limb mesenchyme and Fgf10 subsequently induces Fgf8 expression in the overlying ectoderm. Sall4 is also proposed as a target of Tbx5 [14,15]. We examined whether the expression patterns of these genes are different between the right and left forelimb buds of Tbx5lox/lox;Prx1Cre;Prx1-Tbx mutant embryos. At E10.5 mutant embryos can be divided into two groups depending on the severity of phenotypes, which correlate with the variation in forelimb defects observed later at E17.5. In the first group, there is no obvious left limb bud and small right limb buds are formed (Fig 2B, 2E and 2H). Fgf10, Fgf8 and Sall4 are not detected in the left lateral plate mesoderm (Fig 2B, 2E and 2H arrowheads). In contrast, low levels of Fgf10 and Sall4 are detectable in the right forelimb buds and Fgf8 is present in the right AER. In this group of mutants the left limb bud lacks the first emergence of the bud, rather than a subsequent regression. In the second group, they form small left limb buds (Fig 2C, 2F and 2I, arrowheads). The right limb buds are larger than left ones but still smaller than control limb buds. Fgf10 expression in the left forelimb bud is lower than the right (Fig 2C), whereas there is no detectable difference in Sall4 expression between the left and right forelimb buds (Fig 2I). Fgf8 expression in the left AER is disrupted in contrast to the right AER where it is expressed throughout (Fig 2F).
Together, these results demonstrate that hypomorphic levels of a Prx1-Tbx transgene cause the limb outgrowth defect by failing to establish the correct positive feedback loop of Fgf10 and Fgf8 and this defect is consistently more severe in the left forelimbs than the right.
We tested if mosaic deletion of Tbx5 in the limb mesenchyme induces similar defects as hypomorphic Tbx mutants. We hypothesized that following mosaic deletion of Tbx5, only a subset of cells would express Fgf10 so that the total amount of secreted Fgf10 available in the forelimb LPM is reduced to hypomorphic levels. For this purpose, we used the Prx1Cre(98) transgenic line [16]. In this mouse, the same Prx1Cre transgene is integrated in a different locus, resulting in a mosaic and delayed Cre recombinase activity in the early forelimb bud (Figs 3A–3F and S2). Cre activity is not detected at 16 somites stage in Prx1Cre (98) (Fig 3B), when Cre from Prx1Cre is already active (Fig 3A), but its activity is detectable in the forelimb region at 22 somites stage and at E10.5 in a mosaic manner (Figs 3D, 3F and S2). Cre activity appears to be at equivalent levels in the left and right forelimbs when analysed with a cre reporter and stained histologically (S2 Fig). Furthermore, we confirmed the symmetrical mRNA expression levels between the left and right forelimb buds (p > 0.05 with two-tailed Student’s t-test) (Fig 3G). This mosaic and delayed deletion of Tbx5 by Prx1Cre(98) is sufficient to cause forelimb defects (Fig 3H and 3I).
These Tbx5lox/lox;Prx1Cre (98) embryos display similar, although milder, defects as Tbx5lox/lox;Prx1Cre;Prx1-Tbx mutants including triphalangeal thumb, absent thumb and defects in the scapula and humerus (Fig 4A and 4B). Again, although a range in the severity of defects is observed in different embryos, consistently defects are observed with greater severity and higher frequency on the left side than the right (n = 18) (Fig 4B and S2 Table). An exception in these milder phenotypes was triphalangeal thumb that is observed with similar frequency on both sides. In mutant embryos at E10.5, forelimb buds are smaller than those of control embryos (n = 12) (Fig 4C–4H). Fgf8 expression is disrupted on the left forelimb (Fig 4F). As the defects in this mutant are milder than those of Tbx5lox/lox;Prx1Cre;Prx1-Tbx embryos, clear reduction of Fgf10 and Sall4 expression patterns were not observed (Fig 4D and 4H). These results demonstrate that mosaic and delayed deletion of Tbx5 from the left and right forelimb LPM leads to left-biased asymmetric forelimb defects.
Together, using two different genetic strategies, we demonstrate that bilateral hypomorphic levels of Tbx activity cause left-biased forelimb defects. These results reveal an inherent asymmetric difference in the left and right forelimb LPM and that an optimal level of Tbx activity is required to buffer the asymmetric difference to ensure bilateral, symmetric limb outgrowth.
We tested if the inherent asymmetric difference in the left and right limb-forming LPM is downstream of the axial left-right pathway. The establishment of the left-right axis is critical for the asymmetric patterning of the visceral organs and the disruption of the left-right pathway can cause situs inversus, a reversal of internal organ asymmetry [17–19]. Previous studies have demonstrated that the biased asymmetry of genetic and drug-induced forelimb defects is subject to mirror reversal in situs inversus embryos [20,21]. Therefore, we tested whether the left bias of forelimb defects in Tbx5lox/lox;Prx1Cre;Prx1-Tbx mutants is reversed in situs inversus embryos. We used homozygous mutants of INV as the loss of this gene induces situs inversus rather than randomising situs as reported in the IV mutant [7,22,23]. We generated 30 INV/INV embryos, 21 of which show situs solitus and the other 9 embryos show situs inversus, indicating that in our hands and in common with other reports the penetrance of situs inversus in INV/INV embryos is lower than originally reported [7,24].
Situs solitus, or normal organ asymmetry, in E14.5 Tbx5lox/lox;Prx1Cre;Prx1-Tbx mutants is shown by the left-sided position of the heart (black asterisk, Fig 5A). In this example, the left forelimb is more severely affected (3 digits) than the right forelimb (4 digits) (red arrow, Fig 5B). Because of the low penetrance of situs inversus in INV homozygous mutants we observed, we were only able to obtain 3 Tbx5lox/lox;Prx1Cre;Prx1-Tbx;INV/INV mutants that show situs inversus, indicated by right-sided heart position (black asterisk, Fig 5C). In all embryos, the right limb is more severely affected (Figs 5D, S3 and S2 Table). In the embryo shown in Fig 5D the right forelimb has 3 digits with the most anterior bifurcated (red arrow) whereas the left forelimb has 4 digits.
This indicates that the left-bias of the forelimb defects caused by hypomorphic levels of the Prx1-Tbx transgene are reversed in embryos with situs inversus and embryos now display a right-sided bias in forelimb defects.
Since Fgf10 is a direct target of Tbx5 during forelimb initiation, we tested whether this factor is mediating the mechanism to buffer asymmetry between the left and right forelimb-forming regions. If so, optimal levels of Fgf10 in the Tbx5lox/lox;Prx1Cre; Prx1-Tbx mutant background would be sufficient to rescue bilaterally symmetric limb formation.
To carry out this assay, we used a Cre-inducible Fgf10-expressing line, Z/EGFgf10 (see Materials and Methods). By crossing with the Prx1Cre transgenic, transgene-derived Fgf10 can be expressed throughout the forelimb-forming region. First we tested the ability of the Z/EGFgf10 transgenic to rescue Fgf10 mutant phenotypes. Fgf10-/- null mutants lack all forelimb skeletal elements as the establishment of the positive feedback loop of Fgf10 and Fgf8 is required for limb bud formation and subsequent outgrowth [25] (S4D–S4F Fig). Fgf10 expression from the Z/EGFgf10 transgene, is able to fully rescue the forelimb defects in Fgf10-/- null mutants (n = 2/2) (S4G–S4I Fig and S2 Table), indicating the levels of Fgf10 produced by Z/EGFgf10 are sufficient to support normal limb formation.
We used the Z/EGFgf10 to test if Fgf10 expressed from this transgene is able to rescue the asymmetric defects observed in Tbx5lox/lox;Prx1Cre;Prx1-Tbx mutants. Tbx5lox/lox;Prx1Cre;Prx1-Tbx;Z/EGFgf10 mutants have milder forelimb outgrowth defects compared to Tbx5lox/lox;Prx1Cre;Prx1-Tbx (Fig 6A and 6B), consistent with previous studies showing that Fgf10 acts downstream of Tbx5 during forelimb initiation. Skeletal analysis of four mutants, however, demonstrates that some defects, including absent thumb, bifurcated digit, hypoplastic scapula and short humerus are still observed (Fig 6C–6E). Significantly, these defects are more severe and more frequently observed on the left side (n = 4/4) (S2 Table), indicating left-biased asymmetric differences are still present. These results demonstrate that in the presence of bilaterally symmetric, hypomorphic Tbx expression levels, raising levels of Fgf10 ligand (to a level sufficient to fully rescue limb formation in the Fgf10-/- null mutants) cannot rescue symmetrical forelimb outgrowth.
Here, we demonstrate that inherent asymmetry between the left and right LPM is buffered by threshold levels of Tbx5 that ensure bilaterally symmetric forelimb formation (Fig 7).
The simplest explanation for how bilateral symmetry in limb formation is achieved is by default and that the gene programmes controlling limb formation can operate with equal fidelity on the left and right sides of the embryo. Over the last 20 years however, it has become clear that the pathway that breaks L/R symmetry in the very early embryo and ultimately establishes the asymmetry in visceral organs has an impact on the LPM. Components of the L/R pathway, Nodal, Lefty and Pitx2 are expressed in the left LPM at stages prior to limb formation, however, the effects of these genes on bilateral symmetry of the limb have not been appreciated, previously. Our results indicate that bilaterally symmetric limb development is not the ground state but instead the ‘memory’ of L/R pathway genes expressed in the left LPM can interfere with the ability of LPM cells to establish Fgf10-Fgf8 positive feedback loop essential for limb outgrowth, and that threshold levels of Tbx5 are required to buffer this developmental history. Furthermore, we demonstrate that one direct target of Tbx5, Fgf10, does not have the ability to mask the bilateral asymmetry, suggesting that Tbx5 carries out this buffering role upstream of Fgf10.
Other bilaterally symmetric structures in the embryo, such as the somites, also need to override the influence of early asymmetric gene expression associated with the L/R pathway. RA signalling is essential to synchronize somite formation between left and right sides by antagonizing Fgf8 to ensure symmetric FGF signalling activity on the both sides of embryos [26–28]. Together with our study, these results suggest that buffering the effects of the left-right pathways is a general strategy employed during the formation of bilaterally symmetric structures in the developing embryo. Significantly, however, the mechanisms by which bilateral symmetry is achieved are different in different structures, such as the limbs and somites. During somitogenesis, the L/R asymmetric signals are counter-balanced by asymmetric expression of Nr2f2, a nuclear receptor that promotes the transcriptional activity of the RA signaling pathway [29]. In contrast, in the limb, bilaterally symmetric Tbx5 expression at levels above a threshold are able to buffer the effects of the L/R pathway. A buffering mechanism that operates through bilaterally symmetric gene expression may be more robust than one that relies on establishing asymmetric gene expression to counterbalance a ‘historic’ gene expression asymmetry.
While Tbx hypomorph mutants and HOS patients are unique in that they display left-biased forelimb defects, a few other mutant mice with oriented asymmetric forelimb outgrowth defects have been reported [21,30–32]. Retinaldehyde dehydrogenase 2 (Raldh2) mutants rescued by maternal administration of RA display more severely affected left forelimbs than right [32], similar to Tbx hypomorph mutants. Furthermore, the defects in these rescued forelimbs include triphalangeal thumbs, lack of thumbs and loss of humerus as seen in Tbx hypomorph mutants. Since RA signalling directly regulates Tbx5 transcription [33], Tbx5 expression may be at hypomorphic levels in these rescued embryos, which would explain the left-biased forelimb defects.
Two transgene insertion mutant mice, legless and footless, display right-biased forelimb defects [21,30,31]. The Legless mutation causes hypomorphic expression of Sp8, a zinc finger transcription factor expressed in the AER [34], and the footless mutation causes hypomorphic expression of R-spondin 2, a secreted protein also expressed in the AER [35]. Since these genes are both co-expressed in AER and Sp8 is required for R-spondin2 expression, hypomorphic levels of Sp8 in the legless mutant is likely the cause of hyopmorphic levels of R-spondin2. Although how hypomorphic levels of R-spondin2 causes the forelimb asymmetry is not understood, this mutant demonstrates that bilateral disruption of the Wnt pathway can produce right-biased asymmetric forelimb defects. Further studies will reveal how input from different signaling pathways is buffered to ensure that their ultimate output, in the form of limb structures is bilaterally symmetric.
Asymmetrical limb defects can be induced by teratogens such as cadmium, acetazolamide, MNNG and acetoxymethyl-methylnitrosamine [36–41]. The mechanisms causing these asymmetrical defects are not understood however. The affected mice do not show anterior biased limb defects or other clinical features of HOS as seen in Tbx5 hypomorph mutants suggesting these teratogens are not acting by disrupting the Tbx5 pathway or act on only one component of what could be several different pathways regulated by Tbx5.
The consensus is that HOS is caused by haploinsufficiency rather than pathogenic mutations producing dominant-negative acting forms of the protein. HOS pathogenic mutations have been identified throughout the coding sequence and are not exclusively localized to either the DNA binding T domain or regions thought to interact with cofactors. Functional studies of mutations that lie within the T domain show these forms of the protein have reduced DNA binding ability and diminished binding with interaction partners consistent with these mutations leading to reduced transcriptional activation of target genes leading to functional haploinsufficiency [42].
HOS is unique among congenital abnormalities affecting the limb in that a consistent feature of the clinical presentation is the left-biased severity in the defects [3]. Our results show that the asymmetry in the severity of limb phenotype results as a consequence of an effect on the very earliest events of limb bud formation when the precursors of limb structures are recruited. These events are sensitive to the inherent asymmetry in the limb-forming LPM when sub-threshold, hypomorphic levels of Tbx proteins are present. Our analysis of the penetrance in genetically defined HOS patients and our observations in our HOS mouse model indicate left-biased defects present with almost 100% penetrance. Our results provide an explanation for origins of the left-biased severity of HOS defects and confirm it as key a diagnostic criterion to indicate HOS.
Shoulder/girdle involvement is a cardinal feature of HOS and the humerus can also be affected and this is reflected in our mouse mutants. Overall, we observed more severe shoulder and humerus defects in our mouse mutant series than have been described in HOS patients and this may reflect the generally more severe severity range in limb defects we observe in our mouse mutant compared to HOS. In clinical descriptions of HOS, greater emphasis has been placed on description of more distal structures such as radius and hand that are expected to have greater impact on patient limb function. Shoulder girdle and humerus abnormalities are often poorly described making it difficult to compare to the defects we see in the mouse.
The electronic patient record at Great Ormond Street was reviewed and approved by the hospital clinical audit committee (ref 1337). Data were extracted into a linked anonymised database. Animal work was carried out under an appropriate Home Office licence and approved by the local ethics panel (AWERB).
Mice were staged according to Kaufman 2001 [43]. Noon on the day a vaginal plug was observed was taken to be 0.5 embryonic days (E) of development. Prx1Cre [9], Prx1Cre(98) [16], conditional Tbx5 [44], INV [7], Rosa26RLacZ [45] and the chimeric transgenic lines [8] have all been described previously. The Z/EGFgf10 line was produced using the Z/EG backbone [46] provided by C. Lobe. Briefly a mouse Fgf10 cDNA was inserted into the Z/EG construct upstream of the IRESeGFP cassette. The full construct was transfected into ES cells. Cells that had successfully integrated the construct were selected with G418, screened for single integration events by Southern Blott and surviving clones were used to generate chimeras from which founder animals were derived.
Whole mount in situ hybridisation protocols were carried out as previously described [47]. A minimum of 4 mutant embryos were analysed for each genotype at each stage. mFgf10 [48] and mFgf8 [49] probes were reported previously. We used a full-length mouse Sall4 clone as a probe template.
E14.5 and E17.5 embryo skeletons were stained using alizarin red (for bone) and alcian blue (for cartilage) [50].
RNA was extracted from the 10 forelimb buds of embryos according to the manufacturer’s instructions using the RNeasy mini kit (Qiagen) and cDNA was subsequently prepared by using SuperScript III Reverse Transcriptase (Invitrogen).
Primers were generated using the PrimerBlast application available online (http://www.ncbi.nlm.nih.gov/tools/primer-blast). The following primer pairs were used: Cre recombinase Fwd 5’- GAACGAAAACGCTGGTTAGC -3’ Rev 5’- CCCGGC AAAACAGGTAGTTA -3’, Prx1-Tbx Fwd 5’- GAGACAGCTTTTATCGCTGTG -3’ Rev 5’- CATCGCTGCCCCGGAATCCCT -3’, GAPDH Fwd 5’- TGTCAGCAATGCATCCTGCA -3’ Rev 5’- CCGTTCAGCTCTGGGATG AC -3’. The two-tailed Student’s t-test was used for statistical analysis.
Wild type and mutant forelimb buds (E10.5) were homogenized in RIPA buffer. A rabbit polyclonal Tbx5 antibody raised against a peptide spanning an N-terminal region of the protein was used (details are available on request). Densitometry was performed by scanning the original films and then analyzing the bands with ImageJ (NIH).
The electronic patient record at Great Ormond Street was searched for the terms ‘Tbx5’ and ‘Holt-Oram’. Data on the cardiac, genetic and upper limb changes of affected patients were extracted by G.M. into a linked anonymised database. Patients without recorded Tbx5 mutation analysis were cross-checked with the national reference laboratory in Nottingham, and missing results retrieved. This review was approved by the hospital clinical audit committee (ref 1337).
|
10.1371/journal.pgen.1005710 | GDNF Overexpression from the Native Locus Reveals its Role in the Nigrostriatal Dopaminergic System Function | Degeneration of nigrostriatal dopaminergic system is the principal lesion in Parkinson’s disease. Because glial cell line-derived neurotrophic factor (GDNF) promotes survival of dopamine neurons in vitro and in vivo, intracranial delivery of GDNF has been attempted for Parkinson’s disease treatment but with variable success. For improving GDNF-based therapies, knowledge on physiological role of endogenous GDNF at the sites of its expression is important. However, due to limitations of existing genetic model systems, such knowledge is scarce. Here, we report that prevention of transcription of Gdnf 3’UTR in Gdnf endogenous locus yields GDNF hypermorphic mice with increased, but spatially unchanged GDNF expression, enabling analysis of postnatal GDNF function. We found that increased level of GDNF in the central nervous system increases the number of adult dopamine neurons in the substantia nigra pars compacta and the number of dopaminergic terminals in the dorsal striatum. At the functional level, GDNF levels increased striatal tissue dopamine levels and augmented striatal dopamine release and re-uptake. In a proteasome inhibitor lactacystin-induced model of Parkinson’s disease GDNF hypermorphic mice were protected from the reduction in striatal dopamine and failure of dopaminergic system function. Importantly, adverse phenotypic effects associated with spatially unregulated GDNF applications were not observed. Enhanced GDNF levels up-regulated striatal dopamine transporter activity by at least five fold resulting in enhanced susceptibility to 6-OHDA, a toxin transported into dopamine neurons by DAT. Further, we report how GDNF levels regulate kidney development and identify microRNAs miR-9, miR-96, miR-133, and miR-146a as negative regulators of GDNF expression via interaction with Gdnf 3’UTR in vitro. Our results reveal the role of GDNF in nigrostriatal dopamine system postnatal development and adult function, and highlight the importance of correct spatial expression of GDNF. Furthermore, our results suggest that 3’UTR targeting may constitute a useful tool in analyzing gene function.
| Intracranial delivery of GDNF has been attempted for Parkinson’s disease (PD) treatment but with variable success. For improving GDNF-based therapies, knowledge on physiological role of endogenous GDNF at the sites of its expression is important. However, due to limitations of existing genetic model systems, such knowledge is scarce. Here, we utilize an innovative genetic approach by targeting the 3’UTR regulation of Gdnf in mice. Such animals express elevated levels of Gdnf exclusively in natively Gdnf-expressing cells, enabling dissection of endogenous GDNF functions in vivo. We show that endogenous GDNF regulates dopamine system development and function and protects mice in a rodent PD model without side effects associated with ectopic GDNF applications. Further, we report how GDNF levels regulate kidney development and identify microRNAs which control GDNF expression. Our study highlights the importance of correct spatial expression of GDNF and opens a novel approach to study gene function in mice.
| Exogenously applied glial cell line-derived neurotrophic factor (GDNF) promotes the survival, function, and neurite growth of nigrostriatal dopamine (DA) neurons both in vitro and in vivo [1,2]. The classic motor deficit in Parkinson’s disease is characterized by a gradual loss of nigrostriatal DA neurons, leading to a reduction in striatal dopamine levels, resting tremor, rigidity, and an inability to initiate voluntary movement [3]. Intracranial delivery of GDNF has been tested in clinical trials for treating Parkinson’s disease (PD); however, both the efficacy and the side effects of this treatment vary widely [3–6]. Increasing the therapeutic efficacy of GDNF requires a better understanding of its physiological role; however, our knowledge regarding the postnatal role of GDNF is currently limited. Knockout mice that lack Gdnf or its receptors (Gfrα1 and Ret) die at birth due to a lack of kidneys, but with intact nigrostriatal DA system which undergoes developmental maturation during the first three post-natal weeks [7,8]. It has been reported that 50% reduction in GDNF levels in adult Gdnf conditional knock-out mice has profound consequences on midbrain dopamine neuron survival upon aging [9]. However, our recent study with Gdnf conditional knock-out mice utilizing three Cre systems including the repetition of the experiments performed in [9] did not reveal loss of DA neurons after GDNF deletion or reduction at any age [10]. Based on current evidence it is possible that GDNF either has no physiological role in the brain DA system, that GDNF reduction or deletion in the brain is compensated by another mechanism, or that GDNF regulates the DA system at the functional level, rather than at the level of supporting the survival of the DA cell bodies in the midbrain. Moreover, although GDNF is known to be essential for initiating kidney development [7], our understanding of the role of endogenous GDNF in kidney maturation has remained limited.
Here, we report generation and analysis of mice carrying Gdnf hypermorphic (Gdnfhyper) allele, generated by insertion of a cassette containing bovine growth hormone polyA signal after the stop-codon in Gdnf endogenous locus preventing transcription into Gdnf wild type 3’UTR. These mice have increased–but spatially unchanged–expression of the endogenous Gdnf gene. While Gdnfhyper/hyper mice die by postnatal day 18 (P18) due to kidney defects, Gdnfwt/hyper mice are healthy and only display mild occasional reduction in kidney size. Gdnfwt/hyper animals revealed that GDNF has an important role in the postnatal nigrostriatal system development and adult function and clarified which aspects of the nigrostriatal dopaminergic system structure and function are regulated by GDNF. They also enabled analysis of GDNF function in kidney maturation beyond the first steps in renal development.
In the process of generating a conditional knockout (or “floxed”) Gdnf allele [10], we noted that the 3’UTR of Gdnf is relatively long and evolutionarily conserved (Fig 1A). Since Gdnf 3’UTR inhibits reporter gene expression in a cell line [11] we decided to analyze Gdnf 3’UTR function in vivo by insertion of an FRT-flanked puΔtk cassette [12] after the stop codon in the Gdnf locus in embryonic stem (ES) cells. The puΔtk cassette contains the bovine growth hormone polyadenylation (bGHpA) signal, which induces termination of transcription and is commonly used in gene-trap experiments in mice (Fig 1B). We used a luciferase-based reporter assay to confirm that the bGHpA signal prevents transcription into the Gdnf 3’UTR in our construct (S1A and S1B Fig) and yields correctly sized fusion mRNA (S1C Fig). Using a reporter gene assay, we found an 8-fold increase in luciferase expression from the construct containing Firefly-puΔtk proceeded by Gdnf 3’UTR (relative to Firefly-Gdnf 3’UTR) in a cell line derived from human embryonic kidney cells (HEK293) and a 2-fold increase in a cell line derived from human brain cells (U87) (S1D Fig). We also observed similar inhibitory effects on reporter gene expression, regardless of whether the Gdnf 3’UTR was cloned downstream of a sea pansy (Renilla reniformis) or Photinini firefly (Photinus pyralis) luciferase in both cell lines (S1E Fig), suggesting that the inhibition of expression by Gdnf 3’UTR is not limited to one cell type or dependent on the preceding gene. Blocking transcription with actinomycin D revealed that the Firefly-puΔtk yields a more stable gene product than Firefly-Gdnf 3’UTR, suggesting that negative regulation via Gdnf 3’UTR occurs at the post-transcriptional level (S1F Fig).
Next, we generated mice carrying the Gdnf-FRT-puΔtk-FRT-Gdnf 3’UTR allele (S1G–S1I Fig). Based on the above experiments and a previously published study [11] the allele was hypothesized to result in elevated expression of endogenous Gdnf (Fig 1B, S1D–S1F Fig). Homozygous GdnfpuΔtk/puΔtk mice died before P18 due to extremely small, morphologically disorganized kidneys (see below). In contrast, kidney defects were mild (or absent) in heterozygous mice (S1 Table), which were born at the expected Mendelian frequency (S2 Table) and appeared otherwise phenotypically normal (see below).
We first measured the location and levels of Gdnf expression in the developing kidney, testis and developing hind limb where the expression sites of Gdnf are well established [13,14], http://developingmouse.brain-map.org. Because antibodies that selectively and specifically bind endogenous GDNF protein in histology sections are not currently available, we used in situ hybridization to detect Gdnf mRNA. During embryonic kidney development, the expression of Gdnf is limited to a structure of metanephric mesenchyme called cap condensate, which surrounds a Gdnf-negative epithelial ureteric bud (Fig 2A). A probe spanning the Gdnf-coding sequence (comprising exons 1 through 3) revealed a GdnfpuΔtk allelic dose-dependent increase in Gdnf mRNA levels, but no difference in the site of expression between wild-type, heterozygous, and homozygous mice (Fig 2B, upper panel). Thus we designated the GdnfpuΔtk allele as Gdnf hypermorphic or Gdnfhyper allele. Further analysis revealed that a probe complementary to 525 bp of the 3’ end of the Gdnf 3’UTR revealed fewer Gdnf 3’UTR-containing transcripts in Gdnfwt/hyper mice compared to Gdnfw/wt mice. As expected, no signal was detected in Gdnfhyper/hyper mice using this probe (Fig 2B, lower panel), suggesting that transcription of Gdnf wt 3’UTR is prevented by the puΔtk cassette, as it did in cell line (S1B Fig). A similar increase in Gdnf expression was measured in Sertoli cells (in the testes) and in the developing hind limbs (S2A and S2B Fig). Quantification of Gdnf mRNA and GDNF protein levels in the developing kidneys and testes also revealed a Gdnfhyper allelic dose-dependent increase in both organs (Fig 2C–2E, S2C and S2D Fig), and removal of the puΔtk cassette by crossing with Deleter-FLP mice (Fig 1B; “Gdnf 3’UTRrest/rest mice”) restored GDNF expression to wild-type levels (Fig 2E, S2C and S2D Fig). Thus, we concluded that the Gdnfhyper allele indeed resulted in elevated expression of GDNF in peripheral tissues.
To further analyze the Gdnf transcript in Gdnfhyper mice we applied Northern blotting to analyze Gdnf mRNA in testis. We observed a major band at the expected size of the predicted Gdnf-puΔtk fusion transcript (S2E Fig). We also sequenced the Gdnf transcripts from testis in Gdnfwt/wt, Gdnfhyper/hyper, and Gdnf 3’UTRrest/rest mice using RT-PCR. We used a forward primer spanning Gdnf exon 2 and a reverse primer spanning puΔtk (starting 428 bp downstream from the TGA stop codon of Gdnf) (S2F and S2G Fig); our analysis revealed a Gdnf-puΔtk fusion transcript with the predicted size and sequence (S2G Fig).
Ureteric bud outgrowth–the first step in kidney development–is induced by GDNF [7]. Thus, homozygous Gdnf-knockout mice lack kidneys, and 20–30% of heterozygous GDNF-knockout mice have only one kidney [15]. In vitro, exogenous GDNF promotes the growth of ectopic ureteric buds in embryonic tissue explants and expansion of endogenous ureteric buds tips [16]. However, how endogenous GDNF regulates subsequent steps in kidney development in vivo is poorly understood. In vitro GDNF protein application studies [16] and results from the Gdnf knockout mice [15] suggest that the overexpression of Gdnf in our hypermorphic mice might cause enlargement of the kidneys. In contrast, we found that overexpression of GDNF negatively regulates kidney size and morphological maturation (Fig 2F and 2G; S1 Table) in a concentration-dependent manner (Fig 2B–2E). An analysis of the early steps in kidney development (i.e., in E11.5 through E11.75) in Gdnfhyper/hyper mice revealed that excess GDNF (Fig 2B, upper panel) induces hypertrophic ureteric bud formation and impairs the development and elongation of the stalk (Fig 2H). At mid-gestation (i.e., E13.5), these enlarged ureteric buds and absent or shortened stalks are still observed, and the initial signs of reduced kidney size emerge (Fig 2I). During late embryogenesis, the kidneys in Gdnfhyper/hyper mice fail to reach normal size, resulting in severe renal hypodysplasia with disorganized medulla-cortex compartmentalization, reduced cortical and medullar areas, and cysts in the collecting ducts (Fig 2G). An analysis of kidney function by measuring serum electrolytes revealed that the kidneys of Gdnfhyper/hyper mice function poorly (S2H and S2I Fig). However, in Gdnfwt/hyper mice the kidney function remains relatively normal (S2H–S2K Fig). Finally, renal development was restored by crossing with Deleter-FLP mice (Gdnf 3’UTRrest/rest; S2L Fig). These data indicate that in kidney development, the correct GDNF levels are important as both lack of GDNF and excess GDNF result in failure of kidney development and function. Importantly, in Gdnfwt/hyper mouse the kidneys were functional and mice were healthy, providing us a unique animal model to study the postnatal function of GDNF in the CNS.
Next, we set to analyze the site of Gdnf expression in the CNS of GDNF hypermorphic mice. We utilized in situ hybridization analysis of Gdnf mRNA in thalamic nuclei and spinal cord and observed identical spatial expression patterns between genotypes (Fig 3A and 3B). In the striatum, where Gdnf mRNA levels peak at P12.5 [17,18], the spatial pattern of Gdnf expression was limited to a discrete set of sparsely distributed cells, interspaced with Gdnf nonexpressing cells, consistent with previous reports [17,19] and similar between Gdnfwt/hyper and Gdnfwt/wt mice (Fig 3C). In the striatum, Gdnf is expressed primarily by parvalbumin (Pvalb)—expressing inhibitory neurons [17]. We therefore used RNAscope, a novel high-sensitivity in situ hybridization method for double-staining Pvalb and Gdnf mRNAs in histological sections [20]. First, we verified the specificity of RNAscope probes for Pvalb and Gdnf mRNA in the adult cerebellum and E14.5 kidney (S2M Fig), where the expression of Pvalb and Gdnf, respectively, has been well characterized. Pvalb/Gdnf double-staining of the striatum of adult mice confirmed that the expression of Gdnf in Gdnfwt/hyper mice is largely retained to Pvalb-expressing cells, similar to wild-type controls (Fig 3D and 3E).
To assess the level of Gdnf mRNA in several CNS regions including dorsal striatum, olfactory bulb, hypothalamus, ventral striatum, ventral midbrain and cerebellum in P7.5 Gdnfwt/hyper and Gdnfhyper/hyper animals and in 2.5–3 month old adult Gdnfwt/hyper mice we utilized quantitative RT-PCR. We observed an increase of Gdnf mRNA in Gdnfwt/hyper and Gdnfhyper/hyper mice in comparison to wild-type littermate controls (Fig 3F and 3G). To examine whether the increase in Gdnf mRNA level is reflected in the GDNF protein levels we used ELISA assay and noted that GDNF protein levels were increased by approximately two fold in the dorsal striatum of adult Gdnfwt/hyper mice (Fig 3H). Notably, we observed no elevation in Gdnf mRNA levels in the substantia nigra of adult Gdnfwt/hyper mice (Fig 3G). We conclude that elevation in Gdnf mRNA expression in Gdnfhyper mice occurs in most brain structures which naturally express Gdnf in the brain, and that Gdnf mRNA and protein levels in the dorsal striatum of adult Gdnfwt/hyper mice are two fold increased.
In the first three postnatal weeks, DA neurons experience several major developmental changes, including maturation of striatal target innervation and programmed cell death in the substantia nigra pars compacta [7,8]. The postnatal function of endogenous GDNF in the brain’s dopaminergic system is poorly understood [10]. To assess whether increased GDNF levels in the GDNF hypermorphic mice result in changes in nigrostriatal DA system during development, we first analyzed the striatal levels of phosphorylated extracellular signal-regulated kinase (Erk) a known target in GDNF signaling [7] using western blotting. We found that the levels of phosphorylated ERK2 were increased in the striatum of Gdnfwt/hyper and Gdnfhyper/hyper mice at P7.5 (Fig 4A), indicating increased GDNF signaling. Next, we analyzed rostral brain DA levels and the number of DA neurons in the substantia nigra pars compacta. We found that at P7.5, DA levels in the rostral brain were increased by 25% to a similar extent in both Gdnfwt/hyper and Gdnfhyper/hyper mice, and normalizing Gdnf levels by crossing the mice with Deleter-FLP animals restored DA to wild-type levels (Fig 4B). Compared to wt mice, the number of DA neurons in the substantia nigra pars compacta revealed a similar 15% increase in Gdnfwt/hyper and Gdnfhyper/hyper animals at P7.5 (Fig 4C). Finally, the levels of DA metabolites in the rostral brain were similar between genotypes at P7.5 (S3A and S3B Fig). The development of the DA system was not linked with kidney development, as kidney function was severely impaired in the Gdnfhyper/hyper mice, but not in the Gdnfwt/hyper mice while the DA system parameters were comparable between the Gdnfwt/hyper and Gdnfhyper/hyper mice. This finding was further supported by the lack of correlation between serum urea and brain DA levels in individual heterozygous animals at P7.5 (S3 Table). We conclude that increased GDNF levels increase striatal DA levels and DA cell number in the substantia nigra pars compacta at P7.5.
To assess whether DA system was changed in adult animals, we studied Gdnfwt/hyper mice at 2.5–4 months of age and noted that the DA levels in the striatum of Gdnfwt/hyper mice were increased by 25% compared to wild-type littermates (Fig 4D). Analysis of striatal DA metabolites (S3C and S3D Fig) revealed a 35–40% increase in 3,4-dihydroxyphenylacetic acid (DOPAC) levels (S3C Fig), suggesting that also DA release in Gdnfwt/hyper mice may be increased. To assess the effect of the absence of one Gdnf allele in the comparable genetic background, we also measured DA levels in heterozygous GDNF-knockout mice (Gdnfwt/KO) [10]. Consistent with previous studies [21] the levels of DA in Gdnfwt/KO mice were similar to wild-type levels (Fig 4D). Restoring wild-type expression levels of GDNF by crossing Gdnfwt/hyper mice with Deleter-FLP mice also restored DA levels (S3E Fig). In the substantia nigra pars compacta the number of DA neurons in the Gdnfwt/hyper mice was 15% increased relative to the controls (Fig 4E and 4F), indicating that the moderate 15% increase in the number of substantia nigra pars compacta DA neurons noted at P7.5 is retained in adulthood. Similarly, we found that the number of dopaminergic terminals in the dorsal striatum is increased by 15% in the Gdnfwt/hyper mice (Fig 4G, S3F Fig). Immunohistochemical examination of nigrostriatal DA system revealed no gross anatomical differences between the Gdnfwt/hyper and Gdnfwt/wt mice (S3G and S3H Fig) and the size of the striatum appeared unaffected (S3I Fig).
Collectively, these data indicate that an increase in GDNF expression levels increases the adult number of DA cells in the substantia nigra pars compacta and the number of dorsal striatal DA terminals by 15%. In addition, increase in GDNF expression increases striatal tissue DA and its metabolite DOPAC levels by 25% and 35–40%, respectively, suggesting that DA release in Gdnfwt/hyper mice may be enhanced.
To determine whether increasing the levels of endogenous GDNF affects the function of the nigrostriatal DA system in adult mice, we performed fast-scan cyclic voltammetry measurements in acute striatal slices at 3–4 months of age, and measured the clearance rate of extracellular exogenous DA in the striatum using in vivo amperometry at the same age. To study various functional aspects of the DA system, we used a range of stimulation patterns for the voltammetry measurements (S3J Fig). We observed that stimulations in the striatal tissue of Gdnfwt/hyper mice released about 35–45% more DA (Fig 4H and 4I) with a steeper rising slope (S3K Fig) when compared to Gdnfwt/wt mice. No differences were found between the genotypes in the rise time and decay parameters of the DA events (S3L–S3N Fig). We observed no difference in paired stimulus depression of DA release (Fig 4J), and we found no difference in DA release probability (Fig 4K).
Since more DA is released but DA reuptake parameters are comparable between the genotypes, increased DA uptake in the striatum of Gdnfwt/hyper mice can result at least in part, from the observed 15% increase in the number of dopaminergic terminals in the striatum. However, increased DA transporter (DAT) levels and/or increased DAT activity could also contribute to the observed increase in DA re-uptake. To test these possibilities, we measured the clearance rate of extracellularly applied DA in the striatum using in vivo amperometry. We found that compared to Gdnfwt/wt mice, DAT activity in the Gdnfwt/hyper mice was increased at least up to five-fold in a DA concentration-dependent manner (Fig 4L). Next, we examined whether this difference is due to differences in the levels of DAT between genotypes. We measured total and surface levels of DAT protein in the striata at P7.5 and total DAT levels in the striata of adult mice; we found no differences between the genotypes (S4A–S4C Fig). Taken together, our findings suggest that increasing endogenous GDNF levels increases striatal tissue DA content and striatal DA release and re-uptake without affecting the DA release probability. The observed increase in striatal DA re-uptake in Gdnfwt/hyper mice is most likely explained by the combined effects of increased DAT activity and 15% increase in the number of striatal dopaminergic terminals.
In order to investigate whether increased GDNF levels result in functional consequences in the DA system in behaving animals we administered amphetamine, a dopaminergic stimulant. Amphetamine is taken up by the DAT into the presynaptic terminal, where it is then loaded into synaptic vesicles. Amphetamine releases DA from these vesicles into the cytoplasm and reverses its active transport across the presynaptic membrane; the result is increased DA concentration in the synaptic cleft and increased locomotor activity in treated animals [22]. Compared to Gdnfwt/wt littermates, Gdnfwt/hyper mice had increased amphetamine-induced locomotor activity (Fig 4M), suggesting that amphetamine has a stronger effect in terms of driving higher extracellular DA levels in the striatum of Gdnfwt/hyper mice. To test this hypothesis, we measured extracellular dopamine levels in the striatum following local amphetamine delivery via in vivo microdialysis. Consistent with our hypothesis, we observed an increase in amphetamine-induced extracellular DA levels in the striata of Gdnfwt/hyper mice (Fig 4N). We also measured the effect of amphetamine in striatal slices using cyclic voltammetry. Because amphetamine depletes DA from nerve terminals, cyclic voltammetry can detect the gradual decrease in stimulated DA release following amphetamine application. Consistent with increased DAT function in Gdnfwt/hyper mice, we found that amphetamine depletes the synaptic DA storages with a faster time course in Gdnfwt/hyper mice compared to Gdnfwt/wt littermate controls (Fig 4O).
Based on these results, the increased amphetamine-induced locomotor activity in Gdnfwt/hyper mice is likely due to a combination of three effects. First, the increased DAT activity in Gdnfwt/hyper mice likely causes amphetamine to accumulate in the nerve terminals more rapidly. Second, the 15% increase in the number of striatal dopaminergic terminals increases the number of striatal DA release sites. And third, more DA is released by the terminals. Collectively, these DA system features explain the increased amphetamine-induced locomotor activity in Gdnfwt/hyper mice.
Currently, no mouse model is available that phenocopies the slow disease progression of patients with PD [23]. Thus, the most widely used animal models for studying PD are based on the toxins MPTP (1-methyl-4-phenyl-1,2,3,6-tetrahydropyridine) and 6-hydroxydopamine (6-OHDA), both of which are taken up specifically into DA neurons via the DAT [23]. Given that we found at least up to five-fold increase in DAT activity in Gdnfwt/hyper mice, we expect Gdnfwt/hyper mice to be sensitized to DAT-based toxins, unless the trophic effect of increased endogenous GDNF expression dampens or reverses the phenotype. We found that relative to the controls, Gdnfwt/hyper mice are five-fold more sensitive to striatal 6-OHDA injection. More specifically, following striatal 6-OHDA injection we found five-fold aggravated decrease in striatal DA levels and two-fold aggravated decrease in substantia nigra pars compacta DA neuron numbers in Gdnfwt/hyper mice relative to the littermate controls (Fig 4P and 4Q). To overcome the confounding effect of enhanced DAT activity in Gdnfwt/hyper mice in DAT based PD models, we looked for alternative models of PD. Abnormal aggregation of proteins is a generally accepted pathological process common to most neurodegenerative disorders, including PD. Consistent with this notion, intracranial application of proteasome inhibitors such as lactacystin can induce a PD-like phenotype in both rodents and fish [24,25]. We found that unilateral lactacystin injection just above the substantia nigra induced significant side bias (measured using the corridor test) in Gdnfwt/wt mice; however, lactacystin-injected Gdnfwt/hyper mice did not develop this bias (Fig 4R). In addition, DA and its metabolite levels were better preserved in the striatum of lactacystin-injected Gdnfwt/hyper mice compared to lactacystin-injected wild-type mice (Fig 4S). However, the number of DA cells in the substantia nigra pars compacta was comparably reduced by lactacystin injection in both genotypes (Fig 4T), suggesting that the protective effect from lactacystin-induced PD in Gdnfwt/hyper mice occurs at the functional level in the striatum.
Despite its dopaminotrophic benefits both in PD models and in clinical trials, the delivery of GDNF to the nigrostriatal DA system also induces adverse side effects, including hyperactivity [5,26–29], reduced levels of striatal tyrosine hydroxylase (TH)–the rate-limiting enzyme in DA synthesis [5,30], reduced food intake, and loss of body weight [6,26]. Importantly, none of these effects were observed in our Gdnfwt/hyper mice as assessed by open field test, measurements of food intake in physiological cage, bodyweight measurements of adult mice at 3–4 months of age (Fig 4U–4W) and by western blotting and immunohistochemical measurements of striatal TH levels at P7.5 and in adult mice at 3–4 months of age (S4D–S4G Fig).
To get insight on the mechanism how Gdnf 3’UTR negatively regulates gene expression (S1D–S1F Fig) we looked for trans-acting factors that regulate transcripts containing the Gdnf 3’UTR. Sequence analysis revealed three putative binding sites for RNA-binding proteins (RBPs) and conserved binding sites for several miRNAs in the Gdnf 3’UTR (Fig 5A). We found that the RBPs tristetraprolin (TTP, a negative regulator of brain-derived neurotrophic factor BDNF [31]); embryonic lethal abnormal vision-like protein 1 (ELAVL1); and the AU-rich element-binding protein AUF1 had little or no effect on the expression of a reporter construct containing the Gdnf 3’UTR (S4H Fig), consistent with a previous report [11].
Next, we investigated the effect of miRNAs on regulating Gdnf expression. We examined the miRNAs miR-133a, miR-133b, miR-125a-5p, miR-125b-5p, miR-30a, miR-30b, miR-96, miR-9, and miR-146a, which were selected based on their co-expression with Gdnf in several brain areas [17,19,32,33]; see also www.microrna.org and presence of conserved binding sites within Gdnf 3’UTR (Fig 5A; www.targetscan.org). Our analysis of miRNA expression revealed that miR-9, miR-133a, miR-133b, miR-125a-5p, miR-125b-5p, miR-30a, miR-30b, and miR-146a are all expressed in the developing forebrain, adult dorsal striatum and in the developing kidney (S4 Table). Next, we transfected HEK293 cells with the above-mentioned putative Gdnf-regulating pre-miRNAs. Compared to the control miRNAs, the specific miRNAs negatively regulated the expression of a reporter construct containing the Gdnf 3’UTR by 30–50% (Fig 5B). Next, we examined whether the in silico (genie.weizmann.ac.il, www.targetscan.org) predicted conserved miRNA sites mediate the interaction between these miRNAs and the Gdnf 3’UTR. We paid particular interest to miR-9 and miR-96, as previous data obtained from two genome-wide screens suggested that these miRNAs interact with Gdnf mRNA in the mouse brain; and overexpressing them in human cell line suppresses the expression of GDNF (summarized in [34]). Mutating some of the predicted miRNA seed sites (Fig 5C; see S1 Materials and Methods for details) in the Gdnf 3’UTR either reduced or abolished the ability of miR-9, miR-96, miR-133a, and miR-146a to inhibit expression (Fig 5D), suggesting a direct interaction between these miRNAs and some of the predicted sites in the Gdnf 3’UTR. Based on in silico analysis, puΔtk contains approximately half the number of potential miRNA-binding sites and about 10% of the conserved miRNA sites present in the Gdnf 3’UTR (S4I and S4J Fig). Consistent with this prediction, we found that miRNAs had no effect on expression of a reporter construct containing the puΔtk-Gdnf 3’UTR cassette (S4K Fig). To analyze the effect of miR-9, miR-96, miR-133b, and miR-146a on endogenous GDNF expression, we transiently overexpressed these miRNAs in U87 cells (a human glioblastoma cell line that expresses endogenous GDNF at detectable levels). We found that compared to the control miRNAs, transient co-expression of these four miRNAs reduced the Gdnf mRNA level (Fig 5E) and GDNF protein level (Fig 5F) without affecting cell survival (S4L Fig).
Dicer is required for the maturation of miRNAs. To gain further evidence that the GDNF expression is regulated by miRNAs, we next examined the effect of deleting Dicer on endogenous GDNF expression in primary Dicerflox/flox cortical neurons [35] using adenovirus-mediated delivery of Cre. Deleting Dicer resulted in up-regulation of endogenous GDNF expression in cortical neurons (Fig 5G). Finally, to gain insight into how endogenous miR-9, miR-96 and miR-146a impact endogenous GDNF mRNA levels in human cells we tested six different shRNA constructs targeted to each miRNA for their ability to derepress endogenous GDNF expression in HEK293 cells (S4M Fig). Based on this analysis, we identified the three shRNA constructs targeting miR-9, miR-96 and miR-146a with the highest potency for derepressing endogenous GDNF mRNA expression (Fig 5H). An siRNA against Dicer was used as a positive control (Fig 5H). We conclude that miR-9, miR-96, miR-133, and miR-146a interact directly with their binding sites in the Gdnf 3’UTR; moreover, miR-9, miR-96, and miR-146a regulate the expression of GDNF in vitro.
Due to the limitations associated with existing genetic tools, the function of endogenous GDNF has remained poorly understood [7,10]. Despite multiple attempts by several research groups, transgenic animals in which GDNF expression is restricted to cells that normally express GDNF are not available. The dramatic consequences of ectopic GDNF expression—for example, on the development of the urogenital tract [36]—make it difficult to draw conclusions regarding the function of endogenous GDNF based solely on ectopic GDNF expression in the brain.
GDNF hypermorphic mice provided us with an opportunity to study the function of endogenous GDNF with the focus on the postnatal nigrostriatal DA system and in renal development.
We found that increasing the endogenous levels of GDNF increases the number of DA neurons in the substantia nigra pars compacta during the development, and that this increase in DA neuron numbers is retained in adulthood. Highlighting the importance of the correct expression site, GDNF has no such effects when overexpressed in the mouse brain under the promoter not specific to GDNF-expressing neurons [8].
In the dorsal striatum, increase in GDNF levels resulted in increased number of dopaminergic terminals, increased levels of DA and enhanced DA release and reuptake, explaining the enhanced amphetamine-induced locomotor activity in GDNF hypermorphic mice. Since the nigrostriatal DA system is well known for its capacity to compensate for changes [37], the observed increase in DA transporter activity in GDNF hypermorphic mice likely reflects a measure to counterbalance elevated striatal DA content and release to maintain normal extracellular DA levels. We observed no change in striatal DA transporter levels, suggesting that DAT activity in GDNF hypermorphic mice is regulated by other mechanisms, such as post-translational modifications or protein-protein interactions. Since currently methods allowing temporal analysis of those endpoints in the mouse striatum are not available, the question of how GDNF regulates DAT activity remains to be resolved.
Notably, we found that increased GDNF levels have a protective effect in supranigrally delivered lactacystin model of Parkinson’s disease; this effect is not mediated by protecting DA cell bodies in the substantia nigra pars compacta, but by enhancing the dopaminergic function in the striatum. These findings are consistent with a previous study reporting the lack of effect of selective GDNF deletion or reduction on the survival of DA cell bodies in the substantia nigra pars compacta upon aging [10] and with a study reporting that deletion of GDNF receptor RET does not modulate MPTP toxicity on the dopaminergic system but is required for regeneration of striatal dopaminergic axon terminals [38]. Collectively, these results suggest that in adult mice GDNF acts as a local trophic factor for DA axons in the striatum.
Furthermore, our Gdnfwt/hyper mice do not develop any of the adverse side effects usually associated with ectopic GDNF expression, which can include hyperactivity, loss of body weight, and a decrease in the levels of tyrosine hydroxylase—the rate-limiting enzyme in DA synthesis. Because GDNF is not applied ectopically in our model, the common feature in both experimental animals and human studies—specifically, the massive sprouting of DA fibers towards the site of GDNF delivery, with unknown consequences with respect to side effects, treatment efficacy and behavior—was not observed in Gdnfwt/hyper mice. Together, these results imply that measures that promote elevation in endogenous GDNF levels in the striatum may have clinical potential in the treatment of Parkinson’s disease. However, because we found increased DAT activity in GDNF hypermorphic mice, the simultaneous use of GDNF with DAT-based toxins or drugs should be carefully considered and calls for further investigation.
To date, data on GDNF function is mainly gathered using various ectopic GDNF application methods in rodent and primate models and using constitutive or conditional GDNF knock-out mice. Our results bring an important new dimension since we analyze the effect of the elevation of endogenous GDNF. In Table 1 we illustrate the qualitative difference between ectopic and endogenous GDNF sources.
With respect to the role of GDNF in renal development, we found that excess GDNF levels negatively regulate kidney growth and morphogenesis. This result is contrary to expectations based on the majority of in vitro studies, where the effect of ectopic GDNF on embryonic urogenital block can be followed for the few days, and likely arose from an overstimulation of ureteric bud growth, a key process in renal development.
A comparison between our Gdnfwt/hyper mice and MEN2B mice, which express a constitutively active form of the GDNF receptor RET [41], revealed several common features and key differences between these two models. For example, striatal DA levels, TH-positive cell numbers, and striatal DAT activity are increased in both mutants. In contrast, only MEN2B mice develop increased levels of DAT and TH, and reduced spontaneous locomotion [42–44]. Thus, increasing endogenous GDNF levels is fundamentally different than constitutively activating RET; this finding may have broad-reaching implications with respect to drug design and studies of receptor-ligand biology.
Previously published genome-wide screens suggested that overexpressing miR-9 and miR-96 reduce the levels of GDNF mRNA in a human cell line and that these miRNAs interact with the Gdnf mRNA in the mouse brain [34]. We identified binding sites for miR-9 and miR-96 in the 3’UTR of Gdnf; in addition, we identified binding sites for miR-133 and miR-146a. We also found that reducing the levels of Dicer, an enzyme required for the maturation of miRNAs, derepresses endogenous GDNF expression both in human cells and in mouse primary neurons. Moreover, overexpressing miR-9, miR-96, miR-133b, and miR-146a represses the expression of endogenous GDNF mRNA and protein in a human cell line. Finally, we found that shRNAs against miR-9, miR-96, and miR-146a derepress endogenous GDNF mRNA levels in human cell line. Taken together, these results confirm the miRNA target in vitro [45]. However, every miRNA can have several hundred mRNA targets; thus, assigning specific observations to the direct effect of a given miRNA acting on one target in vivo is currently not possible. Future work with a conditional targeting of Gdnf 3’UTR is necessary to overcome this limitation. Once the effects of adult-onset GDNF elevation in the striatum are characterized, we can then utilize various anti-miRNA strategies to evaluate a given miRNA’s role in GDNF-induced changes in the DA system. Our data also suggest that miR-125a-5p, miR-125b-5p, miR-30a, and miR-30b are possible regulators of GDNF expression. Whether the effect of these miRNAs on GDNF expression is direct or indirect warrants further research.
Interestingly, we found that GDNF derepression in GDNF hypermorphic mice was stronger in the kidneys than in the brain and in the brain areas in which Gdnf mRNA levels are higher. The reasons underlying this finding are currently unknown; however, it may be related to the relative ratio between the mRNA and miRNAs; in addition, miRNAs can have different effects on their target mRNAs depending on the tissue context [46].
Using GDNF hypermorphic mice we found that endogenous GDNF regulates postnatal development and function of nigrostriatal dopamine system. Moreover, some of the identified GDNF functions overlapped with results from earlier studies with ectopic GDNF, whereas others were novel, highlighting the importance of correct spatial expression of GDNF. We also found that about two-fold elevation in endogenous GDNF levels protects mice in lactacystin-based model of Parkinson’s disease without side effects associated with ectopic GDNF applications. Whether increasing endogenous GDNF levels is a viable strategy for developing new therapeutic approaches for treating Parkinson’s disease and other diseases is an important question. Finally, since negative regulation via 3’UTR is shared by many genes, our data pinpoints that 3’UTR-s could provide an important target for genetic studies in vivo. More specifically, prevention of transcription of negatively regulated 3’UTR-s could provide a measure to elevate endogenous gene expression while avoiding mis-expression commonly associated with transgenesis.
Detailed descriptions of all materials and methods are provided in S1 Materials and Methods.
The animal experiments were performed according to the EU legislation harmonized with Finnish legislation and have been approved by the National Animal Experiment Board of Finland (permit no. ESAVI/3770/04.10.03/2012).
Cell culture and molecular biology assays were performed using routine methods in the field; please see S1 Materials and Methods for details.
Briefly, 5667bp 5’ homologous arm spanning the second intron of the Gdnf gene, 6055bp 3’ homologous arm and GDNF protein coding part of Gdnf exon 3 including the stop codon were amplified with PCR from Gdnf-containing PAC (RP21-583-K20, CHORI) and cloned into pFlexible [12,47], to generate Gdnf targeted allele. ES clones were screened with standard Southern blotting. Mice were maintained in 129Ola/ICR/C57bl6 mixed genetic background.
4 μg of lactacystin (AG Scientific) in 4 μl of PBS was injected just above the SN at: antero-posterior (AP) -3.3 mm; medio-lateral (ML) -1.2 mm and dorso-ventral (DV) -4.6 mm. The animals were subjected to corridor test 5 weeks after the injection, and sacrificed for tissue isolation and IHC analysis.
In situ hybridization was performed using a probe spanning Gdnf exons or 525bp in the 3’ end of Gdnf 3’UTR. RNAscope [20] probes detecting Gdnf (red) and Parvalbumin (PV, blue) mRNA were custom made by Advanced Cell Diagnostics.
Dopamine release was evoked on acute striatal slices with electrical stimulations and measured with a carbon fiber electrode calibrated with known dopamine concentrations. The signal was amplified with Axopatch 200B amplifier (Molecular Devices), digitized (ITC-18 board; InstruTech) and analyzed with a computer routine in IGOR Pro (WaveMetrics).
In anesthetized mice (urethane 1.7–1.9 g/kg, i.p.; Sigma) the electrode was mounted in parallel with a micropipette used for application of dopamine. Recordings were performed at two rostrocaudal striatal tracks in each hemisphere, at AP +0.3 or +1.0 mm; ML ±1.8 mm, using Fast Analytical Sensing Technology (FAST-16) system (Quanteon). At each recording site, data was collected from three depths below the dura: at -2.0, -2.5, and -3.0 mm.
A microdialysis guide cannula (MAB 4.1, AgnTho’s AB) was inserted into the dorsal striatum (AP +0.6 mm; ML +1.8 mm and DV -2.2 mm) of mice under isoflurane anaesthesia. After obtaining a stable baseline, the Ringer solution was switched into 100 μM of D-amphetamine for 60 minutes. The dialysis flow rate was 2 μl/min. Concentration of dopamine was analyzed using HPLC.
Striatal tissues lysates were prepared and analyzed immediately after sacrificing with GDNF Emax ImmunoAssay System (Promega).
Open field test was performed in three independent cohorts of comparable size (N = 10–12 male mice per genotype with littermate controls). Metabolic monitoring was performed using Comprehensive Lab Animal Monitoring System (CLAMS).
Statistical analysis for pairwise comparisons was performed using Student’s t-test with two tailed distribution using the unequal variance option. Data from amperometry was analyzed by one-way ANOVA followed by Bonferroni post hoc test. Behavioral data were analyzed using factorial ANOVA design with genotype and cohort as between-subject factors, where appropriate. Post hoc analysis after significant ANOVA was carried out using Student-Newman-Keuls test. Data from CV was analyzed by two-way repeated measures ANOVA, which in the amphetamine analysis was followed by multiple comparisons (Sidak´s). All numerical results are reported as mean ± standard error of the mean. SPSS (IBM Corp., Armonk NY, USA) or STATISTICA 11 (StatSoft Inc., Tulsa) were used for analysis.
|
10.1371/journal.pbio.1001407 | Generation of Functional Blood Vessels from a Single c-kit+ Adult Vascular Endothelial Stem Cell | In adults, the growth of blood vessels, a process known as angiogenesis, is essential for organ growth and repair. In many disorders including cancer, angiogenesis becomes excessive. The cellular origin of new vascular endothelial cells (ECs) during blood vessel growth in angiogenic situations has remained unknown. Here, we provide evidence for adult vascular endothelial stem cells (VESCs) that reside in the blood vessel wall endothelium. VESCs constitute a small subpopulation within CD117+ (c-kit+) ECs capable of undergoing clonal expansion while other ECs have a very limited proliferative capacity. Isolated VESCs can produce tens of millions of endothelial daughter cells in vitro. A single transplanted c-kit-expressing VESC by the phenotype lin−CD31+CD105+Sca1+CD117+ can generate in vivo functional blood vessels that connect to host circulation. VESCs also have long-term self-renewal capacity, a defining functional property of adult stem cells. To provide functional verification on the role of c-kit in VESCs, we show that a genetic deficit in endothelial c-kit expression markedly decreases total colony-forming VESCs. In vivo, c-kit expression deficit resulted in impaired EC proliferation and angiogenesis and retardation of tumor growth. Isolated VESCs could be used in cell-based therapies for cardiovascular repair to restore tissue vascularization after ischemic events. VESCs also provide a novel cellular target to block pathological angiogenesis and cancer growth.
| Angiogenesis—the growth of blood vessels—is essential for organ growth and repair, but also occurs during tumorigenesis and in certain inflammatory disorders. All blood vessels are lined by endothelial cells (ECs)—thin, flattened cells that form a continuous monolayer throughout the entire circulatory system. The cellular origin of new vascular ECs during blood vessel growth in angiogenic situations in adults is a matter of debate. New ECs could develop, in principle, from as yet undiscovered stem cells, as is well documented for the differentiated cells of skin or epithelia, or by the duplication of existing differentiated ECs. Here, we provide evidence for the existence of vascular endothelial stem cells (VESCs) that reside in the adult blood vessel wall endothelium. VESCs constitute a small subpopulation of ECs capable of clonal expansion, while other ECs have a very limited proliferative capacity. When isolated, these VESCs can produce tens of millions of endothelial daughter cells, and a single transplanted VESC can generate in vivo functional blood vessels that connect to host blood circulation. Isolated VESCs could be used in cell-based therapies for cardiovascular repair to restore tissue vascularization following ischemia and could also be pursued as a novel cellular target of inhibition to block pathological angiogenesis, for example during tumor growth.
| The early blood vessels of the embryo and yolk sac in mammals develop by aggregation of de-novo-forming angioblasts into a primitive vascular plexus (vasculogenesis). Blood vessels arise from endothelial precursors, which share an origin with hematopoietic progenitors [1]–[3]. In adults, the growth of blood vessels is essential for organ growth and repair. The best-known pathological conditions in which angiogenesis is switched on are malignant [4], ocular, and inflammatory disorders [5]. Endothelial cells (ECs) are thin, flattened cells that line the inside of blood vessels in a continuous monolayer in all blood vessels through the entire circulatory system. ECs are best identified by their specific location and function, but there are also various cell-surface molecules (such as vWF, CD31, CD34, CD105, vascular endothelial cadherin [VE-cadherin], vascular endothelial growth factor receptor 1 [VEGFR-1], VEGFR-2, Tie-1, Tie-2) that characterize their phenotype [6],[7]. Recently, Weissman and coworkers by performing genetic fate mapping and clonal analysis of individual cells showed that the endothelial stem/progenitor cells involved in adult angiogenesis must be local, non-hematopoietic, and non-circulating tissue resident cells [8]. However, the definite cellular origin of the new ECs necessary for adult neoangiogenesis has remained unknown [8]–[16]. Creation of new ECs in adult tissues could in principle occur by their so far undiscovered tissue resident stem cells, as is well documented for the differentiated cells of skin or epithelia [17]–[19], or by the duplication of existing differentiated ECs, as has been described for pancreatic beta-cells [20].
In passaged human aortic ECs not all cells in the monolayers proliferate at an equal rate [21]. Previous work has also indicated that very low numbers of cells with endothelial characteristics and high proliferative potential may be found in umbilical cord blood or in peripheral blood [22]–[26]. Together, these earlier findings suggest that all ECs in adult tissues may not have an equal potential to produce progeny. Therefore, we wanted to learn if there exists a rare vascular endothelial stem cell (VESC) population that is capable of producing very high numbers of endothelial daughter cells and is responsible for neovascular growth in adults.
In preliminary experiments we found that rare endothelial colony-forming cells (CFCs) were routinely detected when ECs were isolated from single cell suspensions prepared by enzymatic digestion of adult mouse tissues, and cultured in vitro in low-cell density adherent semi-solid methylcellulose matrix colony assays supplemented with the endothelial growth factor VEGF. To study the possibility that endothelial CFCs might reside in the vascular wall endothelium, we perfused adult C57BL/6J mice with PBS to wash out the circulating hematopoietic cells, and subsequently isolated CD31+CD105+ ECs from the lung vasculature and tested them in vitro in colony assays. After removal of the remaining contaminating hematopoietic cells by immunomagnetic lineage depletion using a standard combination of antibodies designed to remove mature hematopoietic cells from a sample based on their hematopoietic surface-marker (CD5, CD45R [B220], CD11b, Gr-1 [Ly-6G/C], 7-4, and Ter-119) expression, endothelial CFCs could be recovered from the tissues at higher frequencies. In these isolated lineage depleted (lin−) CD31+CD105+ mouse lung ECs, CFCs were detected at a mean frequency of 0.14% (standard deviation [SD]±0.051; n = 10), corresponding to approximately one and a half colony-forming units (CFUs) per thousand isolated ECs (Figure 1A). These EC colonies grow beneath the methylcellulose matrix adhered to the plastic bottom of the culture dish (Figure 1A). The formed colonies express phenotypic EC markers CD31, CD105, VE-cadherin, and vWF while the cells are negative for the pan-hematopoietic marker CD45 (Figure 1C).
To further study in vitro the endothelial CFCs, we isolated CD31+CD105+ ECs from transgenic C57BL/6-Tg(ACTB-EGFP)1Osb/J mice where all the tissues, including the ECs, are green fluorescent protein (GFP)+. Subsequently, we seeded a mixture of one CFU of GFP+ CD31+CD105+ ECs and 20 CFUs of wild-type (wt) CD31+CD105+ ECs on a standard 2-D EC culture and let the cells grow until the monolayer was confluent, typically for 12 d. The resulting confluent monolayers of wt ECs contained on average one circular, GFP+ EC batch per culture, demonstrating the clonal growth pattern of the ECs responsible for creating the confluent monolayer (Figure 1B).
To assess the proliferative potential of the CFCs in vitro, we picked up individual colonies, resuspended them in EC growth medium, and cultured the cells as monolayers in 2-D EC cultures. Some of the formed monolayers were propagated for over 6 mo, splitting the cultures always when they reached 90% confluence (Figure 2). The cultures were transferred in large 75 cm2 flasks after an average of ten passages (Figure S1B). When a total of 148 separate colonies originating from distinct CFCs were picked up and cultured ex vivo as monolayers, a total of five colonies could be expanded and passaged for over ten generations, and three of them were propagated for over 20 passages. During the long-term culture, each of these five colonies originating from a single CFC produced tens of millions of daughter cells (Figure 2A). The long-term cultured colony cells express EC markers CD105, VEGFR-2, and the stem/progenitor cell marker CD117 (Figure 2B). Surface marker expression of EC monolayers originating from single CFCs is summarized in Table S1, and micrographs of immunostainings against the cell-surface markers at the 24th passage are shown in Figure S1. Taken together, the experiments demonstrate that a subpopulation of colony-forming ECs exists that are capable of clonal expansion and can produce tens of millions of endothelial daughter cells when expanded as EC monolayers ex vivo.
To further characterize the colony-forming ECs we analyzed their frequency in various subpopulations of lin− ECs isolated from enzymatically digested mouse lung vasculature using fluorescence activated cell sorter (FACS). Sorting against endothelial-specific markers CD31 and CD105 and against CD117 and Sca-1, molecules that are expressed by many adult stem cell types including hematopoietic stem cells (HSCs) and prostate and mammary gland stem cells [27]–[30], was utilized first. CD117 (c-kit, stem cell growth factor receptor [SCFR]) plays an important role in adult HSCs survival and proliferation [27]. Sca-1 is a GPI-anchored cell surface protein that plays a role in modulating CD117 expression and characterizes mouse bone marrow (BM) subset that contains pluripotent HSCs [31]. In addition to certain adult stem cell types, CD117 and Sca-1 are also expressed by some differentiated adult cells. These include mast cells, dendritic cell subsets, and melanocytes (CD117) [32]–[34], and activated T cells (Sca-1) [32],[35]. We found that while CD105 and Sca-1 were expressed by most lin−CD31+ ECs in the lung, CD117+ ECs represent a more infrequent EC subpopulation (39% of all lin−CD31+ lung ECs) (Figure 3A). We therefore prepared CD117-enriched and CD117-depleted fractions of isolated lin−CD31+CD105+ ECs, and assayed them in vitro for CFCs. The CD117-enriched fraction contained CFCs with an almost 10-fold frequency (0.42%, SD ± 0.054 versus 0.045%, SD±0.037; p<0.0001) (Figure 3B) compared to CD117-depleted ECs that contained only few CFCs.
To estimate the frequency of endothelial CD117+ CFCs in yet another assay, a total of 960 freshly isolated GFP+ lin−CD31+CD105+Sca-1+CD117+ single cells were sorted into individual wells of 96-well plates together with a carrier population of lin−CD31+CD105+ wt (GFP-negative) ECs. The presence of a single GFP+ cell per well was checked under fluorescence microscope after sorting. In 7 d, the wt EC carrier population formed a wt (GFP−) EC monolayer in the wells, and the contribution of the single GFP+ cell to the EC monolayer could be studied. The wt (GFP−) EC monolayers on six of the plated 960 wells (0.6% of all wells) contained a clonal area of more than 20 GFP+ ECs (Figure 3C). The rest of the wells (99.4%) contained only a few GFP+ ECs at most (Figure 3C). This occurrence frequency for CD117+ colony-forming ECs (0.6%) is well in line with the results we obtained using the methylcellulose matrix colony assays (Figure 3B).
Further cell-surface marker analyses were performed to better characterize the isolated CD117+ EC population that encloses the progenitor ECs responsible for the clonal expansion. Isolated lin−CD31+CD105+Sca1+CD117+ cells were found to be highly immunoreactive for various established endothelial-cell markers including VE-cadherin (CD144), vascular cell adhesion molecule 1 (VCAM-1, CD106), VEGFR-2, VEGFR-1, CD104 (integrin beta 4), CD34, and for CD14, a marker that is expressed both by hematopoietic cells and by ECs [36]–[39]. No immunoreactivity was detected against hematopoietic lineage markers such as CD45 (leukocyte common antigen), CD11b (macrophage-1 antigen, Mac-1), CXCR4 (chemokine receptor type 4, CD184), F4/80 (a pan macrophage marker), CD115 (macrophage colony-stimulating factor receptor), or against smooth muscle α-actin (Figure 3D). mRNA expression profile from the isolated ECs was analyzed by using real-time quantitative reverse transcription (RT)-PCR, and corresponded to what is expected from ECs (Figure S2). Taken together, the results provide functional evidence that CD117+ ECs are enriched for endothelial CFCs, and that there exists a small subpopulation within CD117+ ECs that are capable of undergoing clonal expansion while other ECs have a very limited proliferative capacity.
lin−CD31+CD105+Sca-1+CD117+ ECs were detected in various different tissues including subcutaneous tissues, lung, liver, and kidney. In the liver, these CD117+ ECs comprised 18% (mean; SD±12; n = 5) of all lin−CD31+ ECs, while in the kidney they were observed more infrequently (mean 2%; SD±1; n = 4). CD117+ ECs were detected in capillaries as well as in arteries and veins (Figures 4A and S3). Abundant CD117+ ECs were discovered in neoangiogenic vessels in subcutaneous matrigel plugs and in B16 melanoma tumors (Figure 4B). CD117+ ECs were also detected in tumor vasculature of all randomly picked human cancer samples that were analyzed (Figure 4C; human malignant melanomas and invasive breast cancers, n = 14).
The possibility that hematopoietic lineage cells might constitute the EC CFCs was further studied using lineage depleted or CD45-enriched murine BM hematopoietic cells. In standardized murine hematopoietic CFC assays the BM hematopoietic produced classical hematopoietic colonies thus confirming their viability. In contrast, lin−CD31+CD105+Sca1+CD117+ lung ECs produced only completely dissimilar EC colonies on the same assay system (Figures 4D and S4). Importantly, the hematopoietic lin− or CD45+ BM cells could not form any colonies in the endothelial colony assay format even when plated up to 100,000 cells per plate (Figure 4D; for both populations six independent experiments were performed in duplicate). Therefore, we conclude the EC colonies we have studied here do not originate from hematopoietic cells.
Additionally, we also tested our findings by using two different genetic reporter systems for the gene expression of the receptor tyrosine kinases VEGFR-2 and Tie-2 (Figure 4E). These transgenic reporter systems were chosen because VEGFR-2 and Tie-2 are expressed primarily by vascular ECs (although also VEGFR-2+ or Tie-2+ HSCs have been described). lin−CD31+CD105+Sca1+CD117+ cells were isolated from transgenic mice with a lacZ-β-gal reporter under the VEGFR-2 promoter (C57BL/6J-Kdrtm1Jrt; lacZ mice [40]) or under the Tie-2 promoter (FVB/N-Tg(TIE2-lacZ)182Sato/J mice [41]). The isolated lin−CD31+CD105+Sca1+CD117+ cells were then studied in colony assays and analyzed for the activity of the lacZ-β-gal reporter system using fluorescence-based detection of β-galactosidase activity. The colonies both from the VEGFR-2 promoter mice and from the Tie-2 promoter mice promptly expressed the β-gal reporter thus confirming the corresponding marker gene promoter activity (Figures 4E and S4C). Taken together, we conclude that all our data collectively indicate that the EC colonies we observed here originate from vessel wall lin−CD31+CD105+Sca1+CD117+ ECs and are not produced by possible contaminating hematopoietic stem or progenitor cells. These data are also in agreement with recent results by Weissman and coworkers, which by genetic fate mapping and clonal analysis of individual cells demonstrate that endothelial stem/progenitor cells involved in adult angiogenesis must be local, non-hematopoietic, and non-circulating tissue resident cells [8].
To learn if entire blood vessels could originate from one single c-kit-expressing EC, we performed in vivo transplantations of GFP-tagged ECs originating from a single lin−CD31+CD105+Sca1+CD117+ CFC (Figure 5A). GFP-tagged lin−CD31+CD105+Sca1+CD117+ cells were first isolated from transgenic C57BL/6-Tg(ACTB-EGFP)1Osb/J mice, and the single cell suspension was plated in adherent colony-forming assays at one CFC per plate and cultured for 12–14 d. The plates that later contained only a single clonal colony were utilized in cell transplantations. The single colonies were manually picked up using a micropipette and an inverted microscope, resuspended, and transplanted in matrigel into wt C57BL/6J mice. When the matrigel plugs were analyzed 14 d later, we observed GFP-expressing CD31+ CD105+ blood vessels (Figure 5B). Perfusion of the mice with fluorescein-labeled microspheres, a standard technique to visualize endothelia of functional blood vessels [42],[43], revealed that the GFP+ vessels generated by the transplanted descendants of a single lin−CD31+CD105+Sca1+CD117+ cell were functional blood vessels connected to the host blood circulation (Figure 5B).
Self-renewal is a defining functional property of adult stem cells, which therefore have the ability to repeatedly respond to tissue injury or other growth stimulus by giving rise to substantial numbers of proliferative progenitors. To determine whether the endothelial CFCs would retain their capacity to generate functional blood vessels, we performed serial transplantations with GFP-tagged ECs. We first inoculated C57BL/6J mice with B16 melanoma tumors mixed with 15 CFUs of GFP-tagged CD31+CD105+ ECs, and performed repeated isolations and serial transplantations of lineage depleted single cell suspensions from the tumors every time after 2 wk of tumor growth. GFP+ blood vessels were observed in secondary, tertiary, and quaternary transplants (Figure 5C). These findings provide direct evidence that the GFP-tagged ECs contained VESCs with self-renewal capacity.
To study in vivo if the CD117+ EC subpopulation is enriched for endothelial stem and progenitor cells capable of creating blood vessels in adults, we transplanted CD117-depleted or CD117-enriched GFP-tagged lin−CD31+CD105+Sca-1+ ECs in matrigel (10,000 or 100,000 ECs per plug) into wt C57BL/6J mice, and analyzed the plugs 14 d later for the presence of GFP+ blood vessels (Figure 6). None of the plugs in mice transplanted with CD117-depleted ECs (n = 20) contained any GFP+ vessels (Figure 6A). Solitary GFP+ ECs from the CD117-depleted transplant were observed within the plugs, but they were infrequent (corresponding to the transplanted EC number) and did not form complete blood vessels. In contrast, all the plugs in mice transplanted with CD117-enriched ECs (n = 12) contained numerous GFP+ blood vessels (Figure 6B). Thus, the VESCs with a potential to create novel blood vessels in adults were greatly enriched in the CD117+ EC fraction.
Mutant C57BL/6J mice with a genetic KitW-sh deficit in c-kit expression were used to study the functional role of CD117 in EC colony-formation and in angiogenesis in vivo. In many adult stem cell types, CD117 plays an important role in stem cell survival and proliferation [27]. The KitW-sh mutation disrupts 5′ regulatory sequences of the c-kit gene [44] and influences c-kit expression in a tissue specific manner affecting melanoblast survival and density [45]. Additionally, c-kit expression is abolished in mast cells, and the KitW-sh mutant mice have a mast cell deficit [46]. However, the KitW-sh homozygotes are healthy and fertile, and produce normal litters [47]. We found that kit deficient C57BL/6J-KitW-sh mice have much lower numbers of CD117-expressing ECs than the wt control mice (mean 1.5%, SD±0.71 of lin−CD31+CD105+ lung ECs, n = 3; Figure 7A). In agreement, isolated ECs from KitW-sh mice had lower numbers of endothelial CFCs (mean <0.02%, SD±0.012; p<0.0001; Figure 7B) than ECs from wt mice. In vivo, KitW-sh mutation resulted in impaired tumor angiogenesis and retardation of tumor growth (Figure 7C and 7D). Blood vessel densities from kit deficient mice inoculated with syngeneic B16 melanomas were less than half of those from wt C57BL/6J mice (mean vessel count per field 14.7, SD±1.5 versus 34.0, SD±3.8; p = 0.006). Correspondingly, the tumor vasculature in the kit deficient mice contained a significantly diminished number of proliferating ECs (mean 17.3%, SD±5.3 versus 25.7%, SD±4.6; p = 0.01; Figure 7C).
Importantly, B16 melanoma tumor growth was very significantly inhibited in the kit deficient mice (Figure 7D). To dissect the possible effect from the kit deficient hematopoietic system on tumor growth, wt mice were subjected to total myeloablation and reconstituted with kit deficient KitW-sh mutant BM or with wt BM. When B16 melanomas were later, after the reconstitution of the hematopoietic system from the BM transplant, implanted to mice, no differences in tumor growth were observed between the two groups, regardless of whether the hosts had received a wt (n = 14) or a kit deficient BM transplant (n = 13; p>0.1; Figure S5). Thus, the retardation in tumor angiogenesis and cancer growth we observed in C57BL/6J-KitW-sh mutant mice is not an effect of the c-kit expression deficit in the hematopoietic system.
While infrequent proliferative cells with endothelial characteristics have been observed circulating in the peripheral blood [22]–[26], genetic fate mapping has suggested that the origin of endothelial stem/progenitor cells involved in adult angiogenesis must be local tissue resident cells [8]. Our present results provide evidence for adult endothelial stem cell hierarchy and the existence of a rare self-renewing CD117+ adult VESC that resides at the blood vessel wall endothelium (Figure 8). In all the divergent assay types we utilized to estimate the frequency of CFCs in different adult EC populations, their occurrence was always only a few CFCs per a thousand total ECs. The CD117+ EC fraction is greatly enriched and the CD117− EC fraction is greatly depleted in colony-forming ECs. Similarly, only CD117+ ECs were able to form blood vessels when transplanted in vivo. In KitW-sh mice where CD117 expression is deficient in a tissue specific manner, the number of CD117-expressing ECs and endothelial CFCs were significantly reduced, and EC proliferation and angiogenesis were impaired. This observed defective EC phenotype is not unlike the already reported deficiency in melanocyte and mast cell densities in these c-kit mutant mice [45],[46]. Taken together, our present results suggest that the ability to proliferate is not a stochastic property of ECs, and that the proliferative potential of ECs is hierarchically organized [21], with different EC subpopulations discriminated by their clonogenic potential. Our present results identify a population of EC progenitors with a high proliferative and clonogenic potential that is a small subset within CD117+ ECs (Figure 8).
While a subset of ECs had a high capacity to produce endothelial daughter cells in various experimental settings in vitro and in vitro, we observed no evidence of transdifferentiation of other cell types to ECs. The possibility that endothelial colony-forming activity might originate from hematopoietic cells was particularly carefully studied, but this was not observed in the experiments. The result is in agreement with earlier reports suggesting that adult stem/progenitor cells are lineage restricted, and that adult blood vessel endothelium does not originate from hematopoietic cells [8],[11],[12],[23],[48]–[59]. However, the present results cannot directly exclude the possible existence of other cell populations capable of endothelial growth. Colony formation assays are admittedly artificial and do not always necessarily represent what actually occurs in vivo.
Additionally, end differentiated mature vascular wall ECs that normally have very limited proliferative capacity might reversibly acquire more stem cell-like characteristics in vivo in angiogenic situations (including human cancer), maybe for example by turning on CD117 expression (Figure 4B and 4C). Indeed, it has been reported that c-kit expression occurs in a subset of angiosarcomas, probably representing oncofetal expression, i.e., reversion of the tumor cell phenotype to that of fetal ECs that may normally show c-kit expression [60]. Thus, angiogenesis might in certain situations be driven also by abundant ECs that can reversibly acquire stem cell-like characteristics and proliferative potential.
Discovery of additional cell-surface markers should allow more efficient identification and isolation of VESCs than we have currently achieved here. The daughter cells produced by VESCs could compensate for cell loss during normal lifelong cellular turnover, and be responsible for the generation of novel neoangiogenic ECs in angiogenic situations. Together, the present results suggest the possibility of cell-based therapies for cardiovascular repair using isolated, highly enriched VESCs to restore tissue vascularization [61],[62]. VESCs are also a cellular target for future—and also present— therapies that aim to restrain angiogenesis by inhibiting endothelial-cell proliferation. Therefore, stem and progenitor cells for vascular ECs and the molecular machinery governing their activation and functions are novel cellular and molecular targets for therapeutic approaches to inhibit excessive angiogenesis and to block cancer growth [4],[62].
Mouse lung ECs were isolated from lungs and other indicated tissues dissected from adult wt C57BL/6J mice (Scanbur AB), GFP-tagged transgenic C57BL/6-Tg(ACTB-EGFP)1Osb/J mice, C57BL/6J-Kdrtm1Jrt; lacZ mice [40], FVB/N-Tg(TIE2-lacZ)182Sato/J mice [41], or C57BL/6J-KitW-sh mice (all from The Jackson Laboratory). Mouse lung ECs were isolated from lungs dissected from adult mice. Mice were anesthetized with Rompun vet (Bayer) and Ketaminol vet (Intervet). The chest was opened through a midline sternotomy. The left ventricle was identified and the ventricular cavity was entered through the apex with a 27-gauge needle. The right ventricle was identified and an incision was made in the free wall to exsanguinate the animal and to allow the excess perfusate to exit the vascular space. The animal was perfused with 20 ml of PBS at approximately 10 ml/min. After being killed, the lungs were collected, fat tissue was removed, and lung tissue was minced manually. The tissue fragments were then digested with DMEM medium containing Dispase II (0.8 U/ml, Roche), collagenase H (1 mg/ml, Roche), Pen/Strep (100 U/100 g/ml), 2% FCS, 2 mM glutamine at 37°C for 1 h after which the suspension was homogenized by pipetting. The homogenate was filtered through a 100-µm nylon mesh filcon (BD Biosciences) and subsequently through a 30-µm nylon mesh falcon (BD Biosciences) and pelleted by centrifugation (300g for 6 min). Erythrocytes were lysed in lysis buffer containing 10 mM KHCO3, 155 mM NH4Cl, 0.1 mM EDTA, (pH 7.5) at room temperature for 2 min. The cell pellets were resuspended in DMEM medium containing Pen/Strep (100 U/100 g/ml), 2% FCS, and 2 mM glutamine. For lineage depletion, Mouse Lineage Cell Depletion Kit (Miltenyi Biotec) was used according to manufacturer's instructions. Subsequently, the cells were incubated with anti-mouse CD16/CD32 blocker (BD Pharmingen) according to the manufacturer's instructions to reduce FcγII/III receptor-mediated antibody binding stained for FACS. The antibodies used included APC anti-mouse CD31 (BD Pharmingen) and PE anti-mouse CD105 (eBioscience), V450 Sca-1 (BD Biosciences), and PE-Cy7 CD117 (eBioscience). Cells were analyzed and sorted using BD FACSAria flow cytometer and cell sorting system (BD Biosciences). Compensation adjustments were performed with single color positive controls. Dead cells and cell debris were excluded by gating the population according to the forward and side light scatters. The positive cells were compared the cells positive in stainings with IgG isotype controls (BD Pharmingen). Before each sorting, laser compensations were adjusted automatically with FACSDiva software version 4.1.2 (BD Biosciences). For single cell sorting, automated cell deposition unit (ACDU) in FACSAria was used for single cell sorting of populations onto 96 multi-well plates. The presence of a single GFP+ cell per well was checked under fluorescence microscope after sorting. A carrier population of 2,500 freshly isolated wt (GFP−) lin−CD31+CD105+Sca-1+CD117+ cells per well was plated together with the GFP+ single cells. The carrier population was used because the single cells were not expected to be able to survive completely alone in the well. For some of the preliminary in vitro assays, CD31+CD105+ ECs were isolated using anti-fluorochrome multisort kit (Miltenyi Biotec) according to the instructions of the manufacturer. The Provincial State Office of Southern Finland approved all the animal experiments.
Freshly isolated ECs were plated in duplicate in 1 ml of 0.8% methylcellulose containing 15% FCS, 1% L-glutamine, 1% BSA, 10−4 mM 2-mercaptoethanol, 0.2 mg/ml human transferring, 0.01 mg/ml rh insulin (all from Stemcell Technologies) supplemented with 100 ng/ml recombinant murine VEGF (Invitrogen). Colonies were studied and counted at days 7, 10, and 14. Scoring of colonies was performed with an inverted microscope. Colonies containing 15 or more cells at day 7 were counted. In preliminary experiments the isolated cell populations were plated at various plating densities, and an ideal plating density for each population was determined.
BM cells were collected by flushing femurs and tibias of the donor mice with 29-gauge needle into DMEM (Invitrogen) supplemented with 2-mM L-glutamin, 100 units/ml penicillin, and 100 g/ml streptomycin (Invitrogen). CD45+ BM cells were selected using CD45 microbeads, mouse (Miltenyi Biotec) according to manufacturer's instructions. Lineage depleted cells were selected using Mouse Lineage Cell Depletion Kit (Miltenyi Biotec) according to manufacturer's instructions. Freshly isolated cells were plated in duplicated in 1 ml of MethoCult GF M3534 medium (contains rm stem cell factor, rm IL-3, and rh IL-6; Stemcell technologies) supplemented with 10 ng/ml GM-CSF (Invitrogen) and 10 ng/ml M-CSF (Invitrogen). Different plating densities ranging from 1,000 to 100,000 cells per plate were studied. Colonies were studied and counted at days 7, 10, and 14.
Freshly isolated lung ECs from transgenic lacZ-carrying reporter mouse strains or from wt mice (negative controls) were plated in 1 ml of 0.8% methycellulose containing 15% FCS, 1% L-glutamine, 1% BSA, 10−4 mM 2-mercaptoethanol, 0.2 mg/ml human transferin, 0.01 mg/ml rh insulin (all from StemCell technologies) supplemented with 100 ng/ml rm VEGF (Invitrogen). After 7 d, methycellulose was washed by PBS. Colonies were stained using ImaGene C12FDG lacZ Gene Expression Kit (Invitrogen) according to the manufacturer's instructions. The colonies were studied and photographed with Axiovert 200 inverted epifluorescence microscope (Carl Zeiss) using a Plan-Neofluar 10× objective (NA = 0.3), Zeiss AxioCam HRc color camera and Zeiss AxioVision 3.1 software.
Freshly isolated ECs were cultured on gelatin coated plates or flasks in IMDM medium containing 15% FCS, 1% L-glutamine, 1% BSA, 10−4 mM 2-mercaptoethanol, 0.01 mg/ml rh insulin, 100 ng/ml rm VEGF, 100 ng/ml rm bFGF, and 100 ng/ml rm EGF (all from Invitrogen), and 0.2 mg/ml human transferring (Sigma-Aldrich) at 37°C in a humidified 5% CO2 atmosphere. To study cultured passaged ECs, the cells were fixed with 4% PFA, blocked with PBS buffer containing 5% serum (Vector Laboratories), and incubated with the primary antibodies overnight at 4°C and subsequently detected with fluorochrome-conjugated secondary antibodies for 30 min at RT. The primary antibodies used in immunofluorescence were rat anti-mouse VEGFR-2 (BD Pharmingen), rat anti-mouse CD45 (BD Pharmingen), FITC conjugated rat anti-mouse CD14 (BD Pharmingen), rat anti-mouse CD16/32 (BD Pharmingen), Alexa 488 anti-mouse CD115 (BD Pharmingen), rat anti-mouse to F4/80 (Abcam), monoclonal anti-actin, alpha-smooth (Sigma), and FITC conjugated anti-mouse CD34 (BD Pharmingen). The secondary antibodies used were Alexa594 anti-rat. The samples were analyzed and photographed with an Axioplan 2 upright epifluorescence microscope using 20× (NA = 0.50) and 40× (NA = 0.75) Plan-Neofluar objectives, AxioCam Hrc camera 14-bit grayscale CCD camera, and Axiovision 4.3 software (Carl Zeiss).
For MS-1 EC line (ATCC number 2279; a kind gift from Jack L. Arbiser, Emory University School of Medicine, Atlanta), and cultured passage 1 and 24 ECs, the cultures were washed with PBS, detached by 1 mM EDTA at 37°C for 10 min, and cell pellets were collected. The studied cells were incubated with anti-mouse CD16/CD32 blocker (BD Pharmingen) according to the manufacturer's instructions to reduce FcγII/III receptor-mediated antibody binding stained for FACS. The antibodies used inculed APC anti-mouse CD31 (BD Pharmingen), PE anti-mouse CD105 (BD Pharmingen), V450 anti-mouse Sca-1 (BD Pharmingen), PE-Cy7 anti-mouse CD117 (eBioscience), FITC anti-mouse CD45 (BD Pharmingen), FITC anti-mouse CD11b (BD Pharmingen), Alexa 700 anti-mouse VEGFR2 (eBioscience), Alexa 700 anti-mouse CD34 (eBioscience), FITC anti-mouse CD115 (eBioscience), FITC anti-mouse CD14 (eBioscience), FITC anti-mouse F4/80 (eBioscience), Alexa 488 anti-mouse VEGFR-1 (R&D systems), FITC anti-mouse CD104 (Clone 346-11A, Abcam), FITC anti-mouse CD106 (BD Pharmingen), FITC anti-mouse CD184 (BD Pharmingen), Alexa 488 anti-mouse CD144 (eBioscience), and FITC anti-mouse alpha smooth muscle actin (Abcam). Cells were analyzed and sorted using BD FACSAria flow cytometer and cell sorting system (BD Biosciences). Compensation adjustments were performed with single color positive controls. Dead cells and cell debris were excluded by gating the population according to the forward and side light scatters. The positive cells were compared the cells positive in stainings with IgG isotype controls (BD Pharmingen). A minimum of 30,000 total events was always analyzed. Before each analysis, laser compensations were adjusted automatically with FACSDiva software (BD Biosciences). Mean fluorescence intensity (MFI) values are presented as the mean ± SD, calculated from multiple mice over three experiments. Statistical significance was assessed by the Student's t test, and p values<0.05 were considered significant.
Total RNA from freshly isolated mouse lung ECs was extracted using RNeasy Mini kit (Qiagen) according to manufacturer's instruction. Sample integrity was analyzed with Agilent 2100 BioAnalyzer to demonstrate consistent quality. Real-time quantitative RT-PCR analyses (RT2 qPCR Primer Assays) were performed by SABiosciences-QIAGEN using the SYBR Green real-time PCR detection method. To normalize the data the Ct value of the housekeeping gene GAPDH was subtracted from the value of the gene of interest. The 2∧(-delta Ct) was used to calculate the normalized relative quantity in order to compare the relative quantification of the gene of interest.
For transplanting the progeny of a single EC, GFP-tagged cells were isolated from transgenic C57BL/6-Tg(ACTB-EGFP)1Osb/J mice, and the single cell suspension was plated in adherent colony-forming assays at one CFC per plate. The colonies were expanded for 12 d in adherent matrix. Subsequently, the plates and the colonies were studied under Axiovert 200 inverted epifluorescence microscope (Carl Zeiss) using both bright field and eGFP channels. The plates that contained a single clonal colony per plate after the expansion were utilized in cell transplantations. The colonies were under Olympus CKX31 inverted microscope and were manually collected from the dish bottom (where they grow adherent to the plastic) by scraping and carefully pipetting using a micropipette. The cells were then resuspended in PBS, and mixed with 200 µl of matrigel (Basement Membrane Matrix; BD Pharmingen) supplemented with VEGF (100 ng/ml; Invitrogen) and bFGF (10 ng/ml; Invitrogen), and injected subcutaneously into wt C57BL/6J mice. After 2–3 wk, the mice were killed and the vasculature of the plugs analyzed. In some experiments, defined numbers of freshly isolated GFP-tagged cells were used. In repeated isolations and serial transplantations in vivo, C57BL/6J mice were inoculated with syngeneic B16 melanoma cells (2×106 B16 cells in 200 µl) mixed with 15 CFUs of GFP-tagged CD31+CD105+ ECs. After 2 wk, the tumor was excised and processed to single cell suspension that was lineage depleted using Mouse Lineage Cell Depletion Kit (Miltenyi Biotec) according to manufacturer's instructions. The lineage depleted cells (containing B16 cells and 10,000 GFP-tagged ECs per tumor) were then inoculated to a new host. The isolation and retransplantation was repeated every 2 wk. Part of the tumors in each generation was processed for tissue analyses.
The B16-F1 melanoma cell line (ATCC number 6323) was maintained in DMEM supplemented with 2 mM L-glutamine, Pen/Strep (100 U/100 g/ml), and 10% FBS (PromoCell). The mice were injected subcutaneously with B16 cells (2×106 cells in 200 µl), and the tumors were let to grow for 10–20 d. Matrigel plugs (400 µl per injection Basement Membrane Matrix; BD Pharmingen) supplemented with recombinant murine VEGF (100 ng/ml; Invitrogen) were injected subcutaneously to the back of the mouse. The plugs were excised and processed for tissue analyses at 2–3 wk after injection. In some experiments, GFP+ cell populations isolated from transgenic C57BL/6-Tg(ACTB-EGFP)1Osb/J mice were mixed with B16 cells or matrigel prior to injection. In experiments where tumor growth rate was studied, 2×106 B16 cells suspended in 200 µl growth factor reduced matrigel were injected subcutaneously. The tumors were measured every other day using a vernier caliper. Tumor volume was determined using the Pi/6×L×W×W formula with L as the longest diameter and W the diameter at the position perpendicular to L.
The primary antibodies used in immunoflurescence were rat anti-mouse CD31/PECAM-1 (BD Pharmingen) rat anti-mouse CD105/endoglin (BD Pharmingen), rat anti-mouse VEGFR-2 (BD Pharmingen), rabbit anti-mouse/human von Willebrand Factor (vWF; DAKO), rat anti-mouse Sca-1 (BD Pharmingen), rabbit polyclonal anti-mouse VE Cadherin (Abcam), rabbit anti-ZO-1 (N-term) (Invitrogen), and goat anti-mouse CD117 (R&D Systems). In some stainings, rat anti-mouse CD45 (BD Pharmingen) and rabbit polyclonal anti-β-galactosidase (Chemicon International) were used as negative controls. The secondary antibodies used were Alexa594 anti-rat, Alexa594 anti-rabbit, Alexa633 anti-rat, Alexa633 anti-rabbit, Alexa 594 anti-goat, Alexa 647 anti-goat, Alexa 647 anti-rabbit, Alexa488 anti-rat, Alexa 488 anti-rabbit (all from Molecular Probes). For staining of whole mounts, all samples were fixed in 4% PFA, blocked with PBS buffer containing 5% serum (Vector Laboratories), 0.2% BSA, 0.09% Na-Azide, 0.2% BSA, and 0.3% Triton-X (Sigma-Aldrich), and incubated with the primary antibodies for 2 d at room temperature. Auto-fluorescent cartilage was removed from the ears before fixing. The samples were washed and incubated with fluorochrome conjugated secondary antibodies overnight at room temperature. Finally, the plugs were sliced, the ears were flattened, and the samples were mounted with antifading medium (Vectashield; Vector Laboratories). For fluorescent immunohistochemistry of cryosections, samples were fixed for 1 h with 2% PFA and incubated in 20% sucrose/PBS overnight. After the cryopreservation, tissues were embedded in OCT compound (Tissue-Tek; Sakura Finetek Europe) and frozen at −70°C. Sections (8–80 µm) were stained with the primary antibodies overnight at 4°C and subsequently detected with fluorochrome-conjugated secondary antibodies for 30 min at room temperature. Finally, the sections were mounted with anti-fading medium (Vectashield). To verify that detected blood vessels were actually functional vessels that were connected to the blood circulation, the mice were perfused with fluorescein-labeled microspheres essentially as described earlier [42]. Fluorescent carboxylate-modified microspheres, 0.2-µm, red fluorescent (580/605) (FluoSpheres, Molecular Probes) were used diluted 1∶6 with PBS. Mice were anesthetized with Rompun vet (Bayer) and Ketaminol vet (Intervet). Tridil (nitroglycerin, Orion Oyi, Espoo, Finland) was included in the anesthesia mixture at 50 mg/ml to allow maximal vasodilation of the peripheral vasculature. The chest was opened through a midline sternotomy. The left ventricle was identified and the ventricular cavity was entered through the apex with a 27-gauge needle. The right ventricle was identified and an incision was made in the free wall to exsanguinate the animal and to allow the excess perfusate to exit the vascular space. The animal was perfused with 4–6 ml of PBS at approximately 10 ml/min and then with the fluorescent microspheres. For staining of colonies grown in colony assays, methylcellulose was washed with PBS and cells were fixed in 4% PFA; blocked with PBS buffer containing 5% serum (Vector Laboratories), 0.2% BSA, 0.09% Na-Azide, 0.2% BSA, and 0.3% Triton-X (Sigma-Aldrich), and incubated with the primary antibodies for overnight at 4°C and subsequently detected with fluorochrome-conjugated secondary antibodies for 1 h at room temperature. The samples were analyzed with a Zeiss LSM510 laser scanning confocal microscope (Carl Zeiss; LSM 5 Software version 3.2) using multichannel (sequential) scanning in frame mode using a 40× (NA = 1.3) Plan-Neofluar and a 63× (NA = 1.4) Plan-Apochromat oil immersion objectives. Single XY-scans typically had an optical slice thickness of 0.9 µm or less. Additionally, the samples were analyzed and photographed with an Axioplan 2 upright epifluorescence microscope using 5× (NA = 0.15), 10× (NA = 0.3), 20× (NA = 0.50), and 40× (NA = 0.75) Plan-Neofluar objectives, AxioCam Hrc camera 14-bit grayscale CCD camera, and Axiovision 4.3 software (Carl Zeiss). 3D reconstitutions were created with ImageJ 1.42q using Volume Viewer plugin in Volume II mode. For bright field immunohistochemistry, 7-µm cryosections were air-dried and fixed in cold acetone for 10 min and processed essentially as described earlier [63]. The sections were first treated with 0.3% H2O2 in methanol to block endogenous peroxidase. After endogenous peroxidase blocking, the sections were rehydrated in PBS for 5 min and incubated in 5% normal serum at room temperature for 30 min, and incubated with rat anti-mouse CD117 (R&D Systems) at 4°C overnight. Negative controls were performed by omitting the primary of secondary antibodies or by using irrelevant primary antibodies of the same isotype. A subsequent incubation for 30 min with biotinylated anti-rat IgG was followed by a 30-min incubation using ABC reagents of the Vectastain Elite Rat IgG ABC Kit (Vector Laboratories) and a 10-min incubation with Chromogen Immpact DAB (Vector Laboratories). The slides were counterstained with Mayer's hematoxylin (DAKO) and mounted using VectaMount AQ medium (Vector Laboratories), and studied and photographed with Leica DM LB microscope (Leica Microsystems GmbH) using C Plan 63× objective (NA = 0.75) and Olympus DP50 color camera (Olympus Corporation). Tumor vessel density was determined from the peri- and intratumoral area in four to six fields (×100) of CD31 stainings of the tissue sections in areas with the highest vascularity (hot spots) essentially as described by Folkman and coworkers [64], and the results were recorded as the sum of all vessel counts per field. Only structures that morphologically appeared as vascular and stained with the immunomarker were taken into account. The scoring was performed blinded to avoid bias. The final score for each tumor was determined as a mean of the values determined from intra- and/or peritumoral vessel densities. For assessment of the proliferating fraction of tumor ECs in B16 tumors, proliferating (ki-67+) EC (CD31+) hotspots were identified by scanning the sections at low magnification by using 10× (NA = 0.3) plan-Neofluar objective and Axioplan2 epifluorescence microscope (Carl Zeiss). Micrographs were then taken using 40× (NA = 1.3) oil objective in four to 18 fields and Axiovision 4.3 software. Total CD31+ ECs and CD31+/ki–67+ ECs within the micrograph area (0.097 mm2) were manually calculated using Image J (NIH Image) via cell counter plug-in. The results were recorded as mean for each tumor. The scoring was performed blinded to avoid bias.
Paraffin-embedded human tumor tissue samples from patients with histologically diagnosed malignant melanoma (n = 4) or malignant breast cancer (n = 10) were drawn from the archives of the Department of Pathology, University of Helsinki. Heat-induced epitope retrieval was performed by heating the deparaffinized tissue sections in a buffered citric acid solution (Dako) for 15 min. The sections were incubated for 30 min in 0.3% hydrogen peroxidase in methanol at room temperature and for 30 min in 5% normal horse serum at room RT. CD117 expression was determined using mouse anti-human CD117 Mab (MS-289-P; NeoMarkers) diluted 1∶100 incubated overnight at 4°C at a dilution of 1∶100. A subsequent incubation for 30 min in biotinylated anti-mouse serum was followed by a 30-min incubation using reagents of the Vectastain Elite ABC kit (Vector Laboratories). Peroxidase activity was developed using DAB (Vector) for 10 min. Finally, the sections were stained with hematoxylin for 30 sec and mounted with VectaMount AQ (Vector). Negative controls were performed by omitting the primary antibody or by using irrelevant primary antibodies.
BM cells were collected by flushing femurs and tibias of the donor mice with 29G needle into DMEM (Invitrogen Corporation) supplemented with 2 mM L-glutamin, penicillin (100 U/ml) and streptomycin (100 µg/ml) (Sigma-Aldrich). Unselected BM cells (7×106) were transplanted into corresponding syngeneic wt mice via tail vein injection. The recipients were irradiated 1 d prior to transplantation by a lethal dosage of 9.1 Gy. Mice were used for experiments 5–8 wk after transplantation.
|
10.1371/journal.ppat.1007171 | Selected HLA-B allotypes are resistant to inhibition or deficiency of the transporter associated with antigen processing (TAP) | Major histocompatibility complex class I (MHC-I) molecules present antigenic peptides to CD8+ T cells, and are also important for natural killer (NK) cell immune surveillance against infections and cancers. MHC-I molecules are assembled via a complex assembly pathway in the endoplasmic reticulum (ER) of cells. Peptides present in the cytosol of cells are transported into the ER via the transporter associated with antigen processing (TAP). In the ER, peptides are assembled with MHC-I molecules via the peptide-loading complex (PLC). Components of the MHC-I assembly pathway are frequently targeted by viruses, in order to evade host immunity. Many viruses encode inhibitors of TAP, which is thought to be a central source of peptides for the assembly of MHC-I molecules. However, human MHC-I (HLA-I) genes are highly polymorphic, and it is conceivable that several variants can acquire peptides via TAP-independent pathways, thereby conferring resistance to pathogen-derived inhibitors of TAP. To broadly assess TAP-independent expression within the HLA-B locus, expression levels of 27 frequent HLA-B alleles were tested in cells with deficiencies in TAP. Approximately 15% of tested HLA-B allotypes are expressed at relatively high levels on the surface of TAP1 or TAP2-deficient cells and occur in partially peptide-receptive forms and Endoglycosidase H sensitive forms on the cell surface. Synergy between high peptide loading efficiency, broad specificity for peptides prevalent within unconventional sources and high intrinsic stability of the empty form allows for deviations from the conventional HLA-I assembly pathway for some HLA-B*35, HLA-B*57 and HLA-B*15 alleles. Allotypes that display higher expression in TAP-deficient cells are more resistant to viral TAP inhibitor-induced HLA-I down-modulation, and HLA-I down-modulation-induced NK cell activation. Conversely, the same allotypes are expected to mediate stronger CD8+ T cell responses under TAP-inhibited conditions. Thus, the degree of resistance to TAP inhibition functionally separates specific HLA-B allotypes.
| Human leukocyte antigen (HLA) class I molecules present pathogen-derived components (peptides) to cytotoxic T cells, thereby inducing the T cells to kill virus-infected cells. A complex cellular pathway involving the transporter associated with antigen processing (TAP) is typically required for the loading of peptides onto HLA class I molecules, and for effective anti-viral immunity mediated by cytotoxic T cells. Many viruses encode inhibitors of TAP as a means to evade anti-viral immunity by cytotoxic T cells. In humans, there are three sets of genes encoding HLA class I molecules, which are the HLA-A, HLA-B and HLA-C genes. These genes are highly variable, with thousands of allelic variants in human populations. Most individuals typically express two variants of each gene, one inherited from each parent. We demonstrate that about 15% of tested HLA-B allotypes have higher resistance to viral inhibitors of TAP or deficiency of TAP, compared to other HLA-B variants. HLA-B allotypes that are more resistant to TAP inhibition are expected to induce stronger CD8+ T cell responses against pathogens that inhibit TAP. Thus, unconventional TAP-independent assembly pathways are broadly prevalent among HLA-B variants. Such pathways provide mechanisms to effectively combat viruses that evade the conventional TAP-dependent HLA-B assembly pathway.
| MHC-I molecules play a pivotal role in immune surveillance of intracellular pathogens by presenting antigenic peptides to cytotoxic T cells (CTL). They also function to regulate natural killer (NK) cell activity by engaging NK cell receptors including KIR3DL1 [1], KIR2DL1/2/3 [2], CD94-NKG2A [3] and KIR3DS1 [4, 5]. MHC-I molecules have strong influences on disease progression in a number of infectious diseases and cancers [6, 7]. In many cases, the peptide-binding characteristics of individual MHC-I proteins are the major factor that determines immune control of diseases, but other characteristics of the MHC-I molecules, such as those relating to variations in the assembly and stability of individual MHC-I molecules, may also have an influence on disease outcomes.
Intracellular proteins are generally degraded into peptide fragments by the ubiquitin-proteasome system [8]. Peptides that bind MHC-I molecules are typically translocated into the ER lumen by the transporter associated with antigen processing (TAP) and then loaded onto MHC-I molecules with the help of other components of the peptide-loading complex (PLC), including tapasin, calreticulin and ERp57 [9]. Empty forms of MHC-I molecules are less thermostable than peptide-filled versions of MHC-I molecules [10–12]. ER quality control, including interactions with the PLC and calreticulin-mediated retrieval [13], contributes to the intracellular retention of empty forms of MHC-I molecules. Additionally, tapasin and the tapasin-related protein (TAPBPR) edit and proofread the MHC-I peptide repertoire by replacing suboptimal low affinity peptides with optimal high affinity peptides [14–20] that can mediate more durable CD8+ T cell responses. In general, an intact PLC is essential for efficient peptide assembly with MHC-I molecules and successful ER quality control. However, individual MHC-I allotypes are known to have different requirements for each component of the PLC. For example, high cell surface expression of some human MHC-I (HLA-I) allotypes is observed in tapasin-deficient cells, whereas other allotypes are poorly expressed [16, 21, 22]. There are known differences in steady state binding of HLA-I molecules to TAP [23]. There are also known allomorph-specific differences in proteasome-dependence [24].
TAP is thought to be the major cellular source of peptide for assembly of most MHC-I molecules. In TAP-deficient cells, MHC-I cell surface expression is generally severely compromised [25–27]. Many viruses down-regulate or inhibit TAP to evade CTL responses [28, 29]. In previous in vitro studies, we found that HLA-B allotypes display a hierarchy of refolding efficiencies and thermostabilities of heavy chains with β2-microglobulin (β2m) in the absence of peptide [12, 22], suggesting distinct intrinsic stabilities of empty forms of HLA-B. Molecular dynamics stimulations have also indicated that empty forms of some HLA-B molecules are more disordered than others [30–32]. Hein et. al. have shown that increasing intrinsic stability of H2-Kb-β2m complex by connecting the α1 and α2 helices with a disulfide bond close to the F-pocket, allowed suboptimally loaded forms of H2-Kb to bypass all cellular quality control steps in TAP-deficient cells [33]. Thus, in the trafficking process, the stability of empty heavy chain-β2m complexes is a key factor that determines the fate of MHC-I molecules in TAP-deficient cells. These findings raised the question of whether empty and suboptimally loaded forms of the more thermostable HLA-B allotypes can bypass ER quality control, traffic to the cell surface and maintain an increased steady-state presence there. Additionally, there can be influences of MHC-I peptide-binding specificities upon HLA-I cell surface expression levels under different conditions. It is known that peptides containing proline at the P2 or P3 position are poorly transported by TAP [34, 35], making it possible that MHC-I allotypes with these binding preferences (for example, HLA-B allotypes of the B7 supertype [36]) are more reliant on additional/alternate sources of peptide, and will have reduced sensitivity to TAP inhibition. Based on these observations, we hypothesized therefore that cell surface expression of MHC-I molecules would be differently dependent on TAP (the major source of MHC-I peptides), based on the intrinsic stabilities of their empty forms and peptide-binding specificity differences. As described below, our studies revealed differential expression levels of HLA-B allotypes on the surface of TAP-deficient and TAP-inhibited cells. Intrinsic stability of the empty form as well as peptide-binding preferences determine cell surface expression levels under TAP-deficiency conditions. Furthermore, we showed that cells expressing HLA-B molecules with Bw4 epitopes that are resistant to inhibition of TAP are more resistant to the activation of KIR3DL1+ NK cells under TAP-inhibited conditions. Together, our findings indicate that HLA-I molecules have evolved to assemble via distinct pathways, which are allotype dependent, as a way to counter pathogen evasion strategies that target the conventional assembly pathway.
In TAP-deficient cells, where the majority of peptides are prevented from entering ER, most HLA-I molecules are empty or suboptimally loaded and HLA-I cell surface expression is generally significantly reduced [25–27]. We expected that when peptide supply is highly deficient in the ER, allotypes with higher intrinsic stabilities of their empty forms might have a better chance to bypass the quality control system as empty molecules or after being loaded with suboptimal peptides to become expressed on the cell surface. To examine whether HLA-B allotypes differ in their abilities to become expressed on the surface of TAP-deficient cells, several HLA-B allotypes that occur at the highest frequencies in United States populations were expressed in the TAP1-deficient human melanoma cell line SK-mel-19 (SK19) [37] or in a TAP2-deficient human fibroblast cell line STF1 [38] using the previously described retroviral infection method [22, 39]. Cell surface expression of HLA-B allotypes was analyzed by flow cytometry after staining with W6/32, which recognizes different HLA-I allotypes with similar affinities. HLA-B allotypes showed large variations in cell surface expression in SK19 cells and STF1 cells (Fig 1A and 1B). Cell surface expression of HLA-B*57:03, B*35:03, B*15:01, B*35:01, and B*15:10 was over 10-fold higher than the cell surface expression of the endogenous HLA-I of SK19 cells and over 5-fold higher than the cell surface expression of the endogenous HLA-I of STF1 cells (Fig 1A and 1B). Cell surface expression of B*44:03, B*58:02 and B*44:02 was very low or undetectable in STF1 cells, and less than two-fold above endogenous HLA-I cell surface expression in SK19 cells (Fig 1A and 1B). Other HLA-B allotypes showed intermediate phenotypes (Fig 1A and 1B).
In general, there was poor correlation between exogenous HLA-I cell surface expression assessed by flow cytometry (Fig 1A and 1B) and total cellular expression assessed by immunoblotting analyses for HLA-I heavy chains (S1A and S1B Fig). For SK19 cells or STF1 cells with HLA-B that were detectable at low or high levels on the cell surface, overexpression of exogenous HLA-B molecules did not induce any consistent unfolded protein response (UPR) compared with vector-infected cells, as assessed by immunoblots for BiP (S2 Fig), induction of which is an UPR indicator [40]. There was a strong correlation between HLA-B cell surface expression levels in STF1 cells and those in SK19 cells (Fig 1C), suggesting that the HLA-B cell surface expression differences were not cell dependent, but rather were TAP-deficiency dependent. Supporting the latter possibility, we have previously shown small differences in the cell surface expression of the HLA-B allotypes in TAP-expressing cells such as a CD4+ T cell line, CEM [22]. To verify that the measured W6/32 signals in SK91 and STF1 cells reflect the intended HLA-B signals rather than any other possible signals, HA-tagged versions of selected HLA-B that were detectable at high or low levels in SK19 and STF1 cells were constructed and expressed in SK19 cells by retroviral infection. An antibody against the HA epitope tag was used to test cell surface or total HA-tagged HLA-B (HA-HLA-B) expression. The HA-HLA-B versions maintained the same expression phenotypes as their untagged counterparts (Fig 1D, S1C Fig).
To confirm varying TAP-dependencies of HLA-B cell surface expression, we examined TAP1-mediated cell surface induction of HLA-B molecules following further infection of selected SK19-HLA-B cell lines with a TAP1-encoding retrovirus (S3A Fig). There was an inverse correlation between the extent of TAP1-mediated induction (+TAP1/-TAP1) and cell surface expression under TAP1-deficiency conditions (Fig 2A and 2B). TAP1 expression was also reconstituted in SK19 cells expressing the HA-HLA-B (S3B Fig). There was again an inverse correlation between the extent of TAP1-mediated induction (+TAP1/-TAP1) and cell surface expression under TAP1-deficiency conditions (Fig 2C). To validate the TAP-dependency results, TAP1 was knocked-down in a TAP-sufficient easily–transfectable cell line, Hela. TAP1-knock down (KD) or parent Hela cells were infected with retroviruses encoding selected HLA-B allotypes that were detectable at high, low or intermediate levels in TAP1 and TAP2-deficient cells (as shown in Fig 1A and 1B). The allotypes expressed at high levels in SK19 and STF1 cells were down-modulated to a lesser extent by TAP1 knockdown compared to the HLA-B allotypes expressed at low levels in SK19 and STF1 cells (Fig 2D and 2E), consistent with the conclusion from TAP induction experiments. Thus, HLA-B allotypes have differential resistance to inhibition of TAP (RIT) phenotypes.
Higher intrinsic stability of the empty form, measured for many tapasin-independent allotypes [12, 22], would also favor a higher efficiency of peptide loading and thus cell surface expression under TAP-deficiency conditions. Zernich et. al. [41] attributed the advantage of B*44:05 cell surface expression under conditions of limiting peptide supply to the high peptide loading efficiency of nascent B*44:05, which also causes its tapasin independency [16, 22, 41]. The structural similarities between the F-pockets of B*44:05 and B*57:03 (the presence of Y116) might confer efficient peptide loading to both allotypes, while residue 116 is a D in B*44:02 and S in B*57:01. Differences in peptide loading efficiencies between B*57:03 and B*57:01 could explain the differences in tapasin- and TAP-dependencies of these two closely-related allotypes, which differ only in the F-pocket regions, at positions 114 and 116.
While there is a partial positive correlation between TAP-dependence and tapasin-dependence of HLA-B cell surface expression (Fig 3A and 3B), some allotypes are clear outliers. Individual HLA-B allotypes have different dependencies on TAP and tapasin. Some highly tapasin-independent allotypes such as B*18:01 and B*40:01, both members of the B44 supertype (pink, favoring peptides containing glutamic acid at position 2 (P2)), are more TAP dependent. Some highly tapasin-dependent allotypes such as B*51:01, a member of the B7 supertype (blue, similar to B*35:01 and B*35:03, favoring peptides containing proline at P2), are less TAP dependent (Fig 3A and 3B). These findings indicate that, the underlying mechanisms of TAP-independence and tapasin-independence are not fully overlapping.
Recent mass spectrometric studies have identified large numbers of HLA-I peptidomes for different allotypes. Comparisons of the anchor residue preferences based on peptide sequences mined from two recent datasets [42, 43] revealed that RIT allotypes generally have higher P2 diversity than several other non-RIT HLA-B (Fig 3C and 3D), which would also favor selection of TAP-independent peptide from unconventional sources. It is noteworthy that there is a strict conservation of P2 among members of the B44 supertype (including B*44:02, B*44:03, B*18:01 and B*40:01 (pink; Fig 3C and 3D)) compared to members of the B7 supertype (including B*35:01, B*51:01 and B*07:02 (blue; Fig 3C and 3D)). Glutamic acid is stringently conserved as a P2 anchor among these members of the B44 supertype, whereas proline, alanine, and other residues occurring at lower frequencies, are found as P2 anchors among members of the B7 supertype (based on data from Ref. 42 (Fig 3C) and 43 (Fig 3D)). B*15:01, another allotype with high RIT, also displays high sequence diversity at the peptide P2 position (53% Q, 15% L, 9% V, 6%I, 5% S, 12% other) (based on data from Ref. 42; Fig 3C). Although a large peptidome dataset is not available for HLA-B*57:03, recent B*57:01 peptidome data indicate high diversity at the peptide P2 position (based on data from Ref. 43; Fig 3D). Structural similarities between the B pockets of B*57:01 and B*57:03 (the P2 binding pocket) predict a high P2 diversity for peptides that bind B*57:03, similar to B*57:01.
Based on prior studies [44–47], signal peptides and hydrophobic peptides are expected to be a TAP-independent source of MHC-I peptides. We first examined the prevalence of anchor residues for TAP-dependent and RIT allotypes within human signal sequence datasets. Within known human signal peptide sequences (www.signalpeptide.de), N-terminal prolines and alanines (excluding the last 6 residues at the C-terminus, which cannot be a P2 residue for any HLA-I epitope), preferred anchor residues for the B7 supertype, are significantly more prevalent than N-terminal glutamic acid, the preferred anchor residue for the B44 supertype (Fig 3E). The low prevalence of glutamic acid within signal sequences could explain why the TAP-dependence phenotypes of B*18:01 does not mirror its high tapasin-independence and stability [12, 22]. Conversely, the higher prevalence of proline/alanine within signal sequences could explain why the TAP-dependence phenotype of B*51:01 is less stringent than predicted by its strong tapasin-dependence and lower stability [12, 22]. Preferred P2 residues for other RIT allotypes, such as B*57:03 (A/S/T) and B*15:01 (Q/L), are also highly represented within the N-termini of signal peptide sequences (Fig 3E). Further, using the NetMHC algorithm [48, 49], epitope predictions were undertaken with the signal peptide sequences from the signal peptide database (www.signalpeptide.de), for epitope estimation for several allotypes (Fig 3F). Significantly more peptides with IC50 < 500 nM (weak binders) or < 50 nM (strong binders) were identified for B*35:01, B*57 and B*15:01 compared to several members of the B44 supertype. We also examined the prevalence of anchor residues (Fig 3G) and predicted weak and strong binders (Fig 3H) for TAP-dependent and RIT allotypes within human transmembrane sequence datasets (TMbase25, ftp://ftp.ncbi.nih.gov/repository/TMbase/). Similar trends were noted as with signal sequences. Thus, our data support the model that peptide loading in the ER contributes to ER exit of RIT allotypes, which is favored by the increased prevalence of peptides with an appropriate P2 residue within signal peptides or transmembrane domains. There is prior evidence for TAP-independent presentation of peptides derived from both of these sources [44–47].
Findings from Fig 3 suggest that signal peptides and protein transmembrane domain-derived peptides could contribute to cell surface HLA-B molecules of RIT allotypes. However, limitation in this pool could result in loading with suboptimal sequences or in partial escape of empty molecules to the cell surface. To test the extent of peptide-receptive cell surface HLA-B, brefeldin A (BFA) decay assays were further conducted in SK19-HLA-B cells that were pre-incubated in the presence or absence of relevant HLA-B-specific peptides. Since anterograde transport is blocked by BFA, and cell surface HLA-I internalization is expected to be more rapid for empty or suboptimally loaded HLA-I [50], the peptide-inducible fraction of the cell surface RIT HLA-B provides an estimate of the fraction of empty or suboptimally loaded cell-surface HLA-B. Based on these analyses, about 30–40% of cell surface RIT HLA-B including B*35:01 (Fig 4A), B*57:03 (Fig 4B), B*15:01 (Fig 4C) and B*44:05 (Fig 4D) are estimated to be expressed in an empty or suboptimally loaded form in TAP1-deficient SK19 cells after overnight culture at 26 oC. Under this condition, empty MHC-I was previously shown to be induced at the cell surface and stabilized by exogenous peptides [25, 50]. Interestingly, even after overnight culture at 37 oC, a condition under which empty MHC-I are generally labile, significant fractions (~20–30%) of the RIT HLA-B allotypes were peptide-inducible (Fig 4A–4D). In contrast, on the surface of TAP-sufficient cells, only a small percentage (~5%) of HLA-B molecules are peptide receptive (Fig 4E and 4F). Thus, TAP-deficiency induces expression of HLA-B that is partially peptide-receptive.
To confirm the presence of suboptimally loaded HLA-B on the cell surface of TAP-deficient cells at 37°C, SK19 cells expressing different RIT HLA-B allotypes were stained with HC10 [51], which detects empty or open HLA-I conformations [52]. Higher levels of HC10-reactive RIT HLA-B allotypes were detectable on the cell surface compared to other HLA-B allotypes (Fig 4G). TAP1 supplementation generally reduced HC10-reactive RIT HLA-B, while simultaneously enhancing the W6/32-reactive forms, contributing to a net decrease in the HC10 / W6/32 ratios (Fig 4H).
In the classical secretion pathway, HLA-I molecules are transported through the Golgi-network to the cell surface. In this pathway, the quality control machinery will prevent suboptimally loaded HLA-I from migration into the medial Golgi apparatus where proteins are modified and become Endoglycosidase H (Endo-H) resistant. Since a subset of RIT HLA-I molecules are suboptimally loaded under TAP-deficiency conditions (Fig 4A–4H), alternative non-classical secretion pathway might exist to transport suboptimally loaded HLA-I molecules to the cell surface [53]. To address this model, the Endo-H sensitivities of HLA-I molecules in TAP-sufficient CEM and TAP-deficient SK19 cells were assessed. As shown in Fig 4I, most of the HLA-I molecules from either cell surface or total lysate of CEM-B*35:01 cells are Endo-H resistant, indicating that, in the steady state, most HLA-I molecules in CEM cells are mature and they traffic to the cell surface largely through the conventional pathway (Fig 4I). In contrast, a greater fraction of HLA-I molecules from SK19 cells expressing exogenous HLA-B molecules are Endo-H sensitive, suggesting that a larger fraction is ER-retained in SK19 cells compared to CEM cells. Interestingly, following surface biotinylation, a detectable portion of RIT HLA-B molecules on the surface of SK19 cells were found to be Endo-H sensitive, in contrast to the predominantly Endo-H resistant HLA-I of CEM-B*35:01 cells. On the other hand, consistent with flow cytometry data (Fig 1A), cell surface expression of a highly TAP-dependent HLA-allotype B*44:02 was barely detectable following surface biotinylation and immunoblotting (Fig 4I, lanes 13 and 14). These findings suggest a non-Golgi route exists for the trafficking of a subset of HLA-I from the ER to the cell surface of SK19 cells. Taken together, the results reported above suggest that under TAP-deficiency conditions, although a fraction of HLA-B molecules are transported to the cell surface through the conventional pathway, a fraction of RIT-HLA-B molecules follow an alternative non-conventional secretory pathway to reach the cell surface.
As an important component of the PLC, TAP becomes a target of immune evasion in many virus-infected cells and tumor cells. For example, the Epstein-Barr virus (EBV)-encoded lytic phase protein BNLF2a acts as a TAP inhibitor by arresting TAP in a transport-incompetent conformation [54]. We examined the effects of BNLF2a on cell surface down-modulation of HLA-B allotypes. Although BNLF2a was transduced to similar levels into CEM cells expressing different HLA-B allotypes (Fig 5A), variable BNLF2a-induced HLA-B down-modulation was observed (Fig 5B), consistent with the prior expression results in TAP-deficient cells (Fig 1A and 1B). Similar results were obtained in K562 cells, which express no endogenous HLA-I (Fig 5C and 5D). Thus, TAP-inhibition has differential effects on cell-surface expression of HLA-B allotypes.
Cell surface HLA-I with Bw4 epitopes function as inhibitory ligands for NK receptor KIR3DL1 [1]. Down-modulation of HLA-I with Bw4 epitopes can induce NK cell activation via the disengagement of KIR3DL1. We expected that under infection conditions which inhibit TAP function, cells expressing RIT HLA-B would be more resistant to NK cell lysis. For comparisons, we chose K562 cells expressing a highly TAP-dependent allele B*44:03, and a RIT allele B*57:03 and cells subsequently infected with a retrovirus encoding BNLF2a. Cell surface expression of B*44:03 was more strongly decreased by BNLF2a than B*57:03 (Fig 6A and 6B). After co-incubation with K562 cells, NK cells from PBMCs of three donors, D136, D187 and D215, were activated, and expression of IFN-γ was measured (Fig 6C, Column 1). Expression of B*57:03 and B*44:03 in K562 cells strongly inhibits KIR3DL1+ NK cell activation (Fig 6C, Columns 2 and 4). In B*44:03 expressing cells (Fig 6C, Column 5) but not B*57:03 expressing cells (Fig 6C, Column 3), KIR3DL1+ NK cells activation was increased by BNLF2a expression, consistent with the reduced expression of B*44:03 compared to B*57:03 on the cell surface.
Although the specific epitopes presented by HLA-I allotypes are well studied, the influences of folding and assembly variations among HLA-I allotypes on immunity are poorly characterized. Under normal conditions that are suitable for peptide loading, the effect of folding and assembly variations might not be significant. However, their effects could be amplified under pathological conditions whereby the function of PLC is disrupted by viral infection or tumorigenesis. In support of our prediction, we found that HLA-B allotypes are expressed at different levels on the surface of TAP-deficient or TAP-inhibited cells.
Our previous findings indicated that, in the absence of peptide, the refolding efficiencies and thermostabilties of HLA-B allotypes are quite variable [12, 22]. Under a tapasin-deficient condition, the capacity for assembly was generally higher for allotypes that had high refolding efficiencies in the absence of a peptide ligand [22]. HLA-I molecules with higher intrinsic stabilities of their peptide-deficient forms were expected to breach ER quality control mechanisms and more readily survive unfavorable assembly conditions such as low peptide supply (TAP-deficiency condition). However, we found that high stability of the peptide-deficient form alone is insufficient to induce the highest level of expression, as exemplified by the intermediate expression level of B*18:01, for which the ER peptide supply is predicted to be highly limiting under TAP-deficiency conditions (Fig 3F and 3H). Based on the findings in this study, we propose the following model: in normal cells when peptide is not limited for most allotypes, cell surface HLA-I molecules are generally loaded with optimal peptides as a result of the abundant peptide pool (Fig 7A). Under a suboptimal condition where the assembly factor tapasin is deficient, the observed expression hierarchy is determined by intrinsic stabilities and peptide loading efficiencies (Fig 7B) [22]. Under a third condition where peptide is highly limited due to TAP inhibition or deficiency (Fig 7C), surface expression of the majority of HLA-B allotypes is strongly reduced. On the other hand, surface expression of RIT allotypes is less affected, because they have high intrinsic stabilities, high peptide loading efficiencies or broader specificities for peptides prevalent within signal sequences or other unconventional sources.
Despite the expected role for peptides from unconventional sources as a determinant of TAP-independent HLA-B expression, many cell surface RIT HLA molecules are suboptimally loaded (Figs 4 and 7C). Suboptimally loaded HLA molecules arise as a result of a limiting supply of peptides in the ER, an imperfect ER quality control system for the retrieval of suboptimally loaded molecules, and alternative (non-Golgi) pathways for transport to the cell surface (Figs 4I and 7C) [53]. The Endo-H sensitive pool of RIT HLA-B is particularly noteworthy (Fig 4), and suggestive of models of peptide loading within a non-conventional secretory pathway for nascent HLA-I molecules, previously described within professional antigen presenting cells (APC) [53]. Other cell types such as melanoma cell lines also appear to have such pathways (Fig 7C). Although components of the PLC are very important for peptide loading to MHC-I molecules, unconventional antigen processing and peptide loading pathways do appear to widely exist (Fig 7B and 7C). Among the tested HLA-B allotypes, B*35:01, B*35:03 and B*15:01 are noteworthy for their high expression when either TAP or tapasin are deficient. Since inhibition of TAP and tapasin is a common evasion strategy used by pathogens and tumors [28, 55] we propose that the folding and assembly characteristics of these allotypes have evolved to allow CD8+ T cell-mediated immune surveillance to persist in the face of pathogenic challenges to the conventional pathway.
The B7 supertype is particularly noteworthy for the higher propensity for TAP-independent expression (Fig 3). Allotypes belonging to this supertype bind peptides with proline at P2, which are highly disfavored for TAP-mediated transport [35]. In a recent study, we showed that, compared with other HLA-B, those belonging to the B7 supertype tend to be expressed at lower levels in normal human lymphocytes but not monocytes. Taken together with findings in this study, it appears that mismatch between TAP-transporter specificity and HLA-I peptide binding specificity causes suboptimal assembly and expression of allotypes belonging to the B7 supertype in some cell types, but confers an expression advantage under TAP-deficient or TAP-inhibited cells and possibly in professional antigen presenting cells that have specialized antigen acquisition pathways for HLA class I.
While previously it was found that empty MHC-I molecules move to the surface of TAP-deficient cells only at sub-physiological temperature [50], here we show that partially peptide-receptive forms of RIT HLA-B allotypes are expressed on the surface of TAP-deficient cells even at physiological temperature (Fig 4). Duration of HLA-I molecules on the cell surface is dependent on their stabilities [56]. HLA-I molecules with higher stability of their empty forms are also expected to be more stable on the cell surface in their empty forms. On the other hand, for many allotypes, the empty forms will be rapidly internalized and degraded at physiological temperature due to the relative instability. Empty or open MHC-I conformers have been drawing increasing attention in recent years. They are proposed to be ligands for many receptors, including KIR3DS1 [4, 5], KIR3DL2 [57], KIR2DS4 [57] and LILRB2 [58]. Many of the described interactions with open MHC-I involve in vitro studies with acid-treated classical HLA-I. The natural prevalence of empty forms of classical HLA-I in cells is thus far poorly characterized. Under normal and TAP-deficiency conditions, RIT allotypes provide a natural source of partially empty class I, and might thus also be more efficient in triggering signals through receptors specific for open HLA-I. Our recent studies indicate that empty HLA-B*35:01 molecules on the cell surface can augment CD8+ T cell activation through enhanced engagement with CD8 [12]. Based on those findings, we expect that, under TAP-inhibited conditions, empty forms of all RIT HLA-B can synergize with reduced levels of antigenic peptide-bound versions to facilitate and maintain some level of CD8+ T cell surveillance of infections. Thus, although RIT HLA-I molecules may not show specific advantages under optimal antigen presentation conditions, they are expected to be more efficient in presenting TAP-independent peptides to CD8+ T cells in infection or tumor conditions involving TAP blockade. Nonetheless, it is important to note that viruses and cancers have developed many other strategies to evade immune recognition, such as the direct down-regulation of HLA-I expression and interference with IFN-γ signaling (for example, [59]). Thus, cells expressing RIT HLA-B could still escape immune surveillance under other different pathogenic conditions.
T cell epitopes associated with impaired antigen presentation (TEIPP) [60, 61] are known to emerge under conditions of inhibited antigen presentation, including TAP-deficiency conditions. In fact, it is reported that CD8+ T cells responsive specifically to TAP-inhibited cells are widely prevalent in the human blood probably due to the prevalence of viruses that encode TAP inhibitors such as EBV, CMV and HSV [62]. Given the high expression levels and suboptimal peptides, RIT HLA-B molecules may contribute dominantly to the HLA-B-restricted CD8+ T cell repertoire against TEIPP (including both self-peptides and viral epitopes) under conditions where TAP expression is inhibited or TAP function is suppressed, an area for further assessment. Moreover, the prevalence of RIT HLA-B molecules might be a reason that there is only mild immunodeficiency in TAP deficient humans [63], and RIT-HLA-I may be the dominant antigen presenting alleles in these patients.
In conclusion, it is well recognized that pathogens have developed strategies to escape cytotoxic T cell surveillance by, for example, disrupting HLA-I assembly pathways [28, 29]. It is now apparent that HLA-I molecules have also evolved to assemble via distinct pathways, which are allotype dependent, as a way to counter pathogen evasion strategies that target the conventional assembly pathway (Fig 7). Thus, the textbook-defined HLA-I assembly pathways are not fully applicable to all allotypes. In this study, we demonstrate that 15% of tested HLA-B allotypes are resistant to inhibition or deficiency in TAP, which is considered a central source of peptides for HLA-I assembly. Cell surface expression of several HLA-B allotypes is readily observable under TAP-deficiency conditions, and relates to HLA-B intrinsic stabilities, peptide loading efficiencies, peptide binding preferences and unconventional secretory pathways. Thus, TAP-independent pathways of antigen acquisition are quite broadly prevalent. RIT HLA-B molecules are expected to confer immune recognition advantages for the CTL response under TAP-inhibited conditions, via the mechanisms outlined above. Conversely, when TAP function is blocked, HLA-B allotypes with Bw4 epitopes that are strongly down-modulated confer induced abilities to mediate NK activation, via reduced KIR3DL1+ NK cell binding (Fig 6). Overall, the findings in this study point to important functional distinctions within the HLA-B locus that relate back to intrinsic structural features of the proteins and their intracellular assembly characteristics.
Blood was collected from consented healthy donors for functional studies in accordance with a University of Michigan IRB approved protocol (HUM00071750). All donors provided informed written consent.
Human melanoma cell line SK-mel-19 (SK19) [37] (obtained from the laboratory of Dr. Pan Zheng), fibroblast cell line STF1 [38] (obtained from the laboratory of Dr. Henri de la Salle), cervical cancer cell line Hela (obtained from the laboratory of Dr. Oveta Fuller) and ecotropic virus packaging cell line BOSC (obtained from the laboratory of Dr. Kathleen Collins) were grown in DMEM (Life Technologies) supplemented with 10% (v/v) FBS (Life Technologies) and 1× Anti/Anti (Life Technologies) (D10). T4-lymphoblastoid cell line CEM-ss (CEM) cells (obtained from the laboratory of Dr. Kathleen Collins) and chronic myelogenous leukemia cell line K562 cells (obtained from ATCC; CCL-243) were grown in RPMI 1640 (Life Technologies) supplemented with 10% (v/v) FBS, 1× Anti/Anti, 2 mM glutamine (Life Technologies) and 10 mM HEPES (Life Technologies) (R10).
The following monoclonal antibodies were used in this study: Pacific Blue-conjugated anti-human CD3 (clone UCHT1; BioLegend), PE-Cy7-conjugated anti-human CD56 (clone CMSSB; eBioscience), FITC-conjugated anti-human KIR3DL1 (clone DX9; BioLegend), Alexa Fluor 700 conjugated anti-human IFN-γ (clone B27; BioLegend), purified anti-HA.11 (Clone 16B12; BioLegend), anti-BiP (Clone C50B12; Cell Signaling Technology), anti-GAPDH (Clone 14C10; Cell Signaling Technology) and anti-vinculin (Clone E1E9V; Cell Signaling Technology). Dead cells were excluded from flow cytometric analyses with 7-amino-actinomycin D (7-AAD; BD Biosciences) or the amine-reactive dye Aqua (405nm, Life Technologies). HLA-I antibodies W6/32, HC10 and 171.4 were produced in the University of Michigan Hybridoma Core. The TAP1 antibody 148.3 was kindly gifted by Dr. Robert Tampé.
All HLA-B alleles in the retroviral vector LIC pMSCVneo were prepared as described previously [22]. HA-tagged versions of HLA-B*35:01, B*35:03, B*57:01, B*44:02 and B*4405 were prepared as described previously [64, 65]. To prepare HA-tagged versions of HLA-B*15:01, B*44:03, B*57:03 and B*58:02, corresponding clones from pMSCVneo [22] were digested with NaeI and XhoI to prepare the 3′ regions of these HLA-B (encoding the portion of the protein downstream of the signal sequence). The B*35:01 signal sequence plus HA-tag was isolated by EcoRI and NaeI digestion of HA tagged B*35:01. Finally, the HLA-B*15:01, B*44:03, B*57:03 and B*58:02 NaeI–XhoI fragments and the EcoRI-NaeI fragment from HA-B*35:01 were ligated into pMSCVneo (cut with EcoRI and XhoI) in a three-way ligation. Retroviruses were generated using BOSC cells and used to infect SK19, STF1, Hela, CEM or K562 cells. Cells were infected with retroviruses encoding the HLA-B molecules, selected by treatment with 1 mg/ml G418 (Life Technologies), and maintained in 0.5 mg/ml G418. Exogenous HLA-I expression was verified by immunoblotting analyses of cell lysates using the mouse anti-human monoclonal antibody 171.4 or anti-HA and secondary antibodies GαM-HRP (Jackson ImmunoResearch Laboratories) or GαM-IRDye 800CW (LI-COR Biosciences). SK19 cells expressing exogenous HLA-B molecules were infected with the human TAP1-encoding retrovirus and selected by treatment with 1 μg/ml puromycin (Sigma-Aldrich), and cells were maintained in 0.5 μg/ml puromycin. TAP1 expression in SK19 cells was verified by immunoblotting analysis of cell lysates using mouse anti-human TAP1 monoclonal antibody 148.3 [66] and secondary antibodies GαM-HRP or GαM-IRDye 800CW. The Western blots were developed for chemiluminescence using the GE Healthcare ECL Plus kit or scanned for IRDye fluorescence using Odyssey System (LI-COR Biosciences). CEM and K562 cells expressing exogenous HLA-B molecules were infected with the BNLF2a-encoding retrovirus and selected by treatment with 1 μg/ml puromycin (Sigma-Aldrich), and cells were maintained in 0.5 μg/ml puromycin. MSCV-N BNLF2a was a gift from Dr. Karl Munger [67] (Addgene plasmid # 37941). BNLF2a expression was verified by intracellular staining with primary antibody anti-HA and secondary antibody PE-conjugated goat anti-mouse IgG (GαM-PE, Jackson ImmunoResearch Laboratories).
TAP1 was knocked-down in Hela cells by using the CRISPR/Cas9 system based TAP1 Double Nickase Plasmid from Santa Cruz Biotechnology according to manufacturer’s protocol. Puromycin selection and limiting dilution was subsequently undertaken to obtain monoclonal TAP1-KD cell lines. TAP1 knockdown was verified by immunoblotting analysis of cell lysates using anti-TAP1 antibody 148.3 [66] and secondary antibodies GαM-HRP (goat anti-mouse horse radish peroxidase) and by intracellular staining with 148.3 [66] and secondary antibody GαM-PE. HLA-B alleles were expressed in Hela or Hela-TAP1-KD cells using the method described above.
A total of 1×105−1×106 cells were washed with FACS buffer (phosphate-buffered saline (PBS), pH 7.4, 1% FBS) and then incubated with W6/32 or HC10 antibodies at 1:250 dilutions or anti-HA at 1:50 dilution for 30–60 min on ice. Following incubation, the cells were washed three times with FACS buffer and incubated with GαM-PE or GαM-PE-Cy7 at 1:250 dilutions for 30–60 min on ice. The cells were then washed three times with FACS buffer and analyzed using a BD FACSCanto II cytometer. The FACS data were analyzed with FlowJo software version 10.0.8 (Tree Star, San Carlos, CA). Data are deposited in the Dryad repository: http://dx.doi.org10.5061/dryad.m4862mk [68].
The night before the experiment, cells were moved to 26°C or kept at 37°C. The next day, cells were washed with PBS, and the medium (containing 100 μM peptide where indicated) was added and cells were incubated at 26°C for 2h. Cells were then incubated at 37°C in the presence of 20 μg/ml brefeldin A (BFA) for an additional 2h and then harvested. The HLA-B signals were quantified by flow cytometry after staining with W6/32 and subtracting signals obtained from cells infected with a retrovirus lacking HLA-B. Peptide receptive HLA-I was quantified as (MFI HLA-I(+peptide)–MFI HLA-I(-peptide)) / MFI HLA-I(+peptide)*100 and averaged across 3–4 independent measurements for each condition. Peptides used (S1 Table) were B*57:03-restricted epitopes TSTLQEQIGW (TW10) and KAFSPEVIPMF (KF11), B*44:05-restricted epitopes VEITPYKPTW (VW10) and EEFGRAFSF (EF10), B*15:01-restricted epitopes LEKARGSTY (LY9) and ILKEPVHGVY (IY10) and B*35:01-restricted epitopes FPVRPQVPL (FL9) and LPSSADVEF (LF9) [64]. All peptides were purchased from peptide 2.0 (Chantilly, VA, USA). All peptides are in the IEDB database except self-peptide LF9.
Cell surface proteins were biotinylated by incubating cells with 2mM EZ-Link NHS-PEG4-Biotin (Thermo Scientific) in PBS for 10 min at room temperature followed by three washes in PBS. After washing, labeled cells were lysed in lysis buffer (1× PBS, 1 mM phenylmethylsulfonyl fluoride, and 1% Triton X-100) for 1h on ice. The lysates were centrifuged at 13,000 g to remove cell debris. Biotinylated proteins were bound to streptavidin conjugated beads for 2 h at 4°C. Beads were washed three times with lysis buffer, and boiled for 10 min in the presence of denaturing buffer. As controls, total cell lysates were directly boiled for 10 min in denaturing buffer. The materials obtained from the beads and total cell lysates were split into two equal aliquots and one of the aliquots was digested with Endo-H (New England Biolabs) according to the manufacturer’s protocol. HLA-I molecules were separated by SDS-PAGE and then immunoblotted using the mouse anti-human monoclonal antibody 171.4.
Fresh blood collected from donors was subjected to centrifugation over a Ficoll-Paque Plus (GE Healthcare Life Sciences) density gradient, washed twice with PBS + 2% FBS and resuspended in R10. Isolated PBMCs were cryopreserved in Recovery Cell Culture Freezing Medium (Life Technologies). IFN-γ expression in NK cells was detected by intracellular cytokine flow cytometry. Briefly, frozen PBMCs (2 × 105 cells/well) were incubated with K562 cells expressing or lacking HLA-B molecules at 1:1 (PBMC:K562) ratio in 200 μL complete media in 96-well U-bottom plates. GolgiPlug (containing brefeldin A, BD Biosciences) was added at 1:1000 1h later. After incubation for an additional five hours, cells were stained with Pacific Blue-conjugated anti-CD3, PE-Cy7-conjugated anti-CD56 and FITC-conjugated anti-KIR3DL1 mAbs for 30 minutes at 4°C, fixed in 4% paraformaldehyde for 10 minutes at room temperature, and permeabilized with 0.2% saponin for 10 minutes. Cells were then stained with Alexa Fluor 700-conjugated anti-IFN-γ for 30 minutes at 4°C and analyzed by flow cytometry.
Statistical analyses (ordinary one-way ANOVA analysis with Fisher’s LSD test) were performed using GraphPad Prism version 7.
|
10.1371/journal.pbio.0060208 | Evolution of Genomic Imprinting with Biparental Care: Implications for Prader-Willi and Angelman Syndromes | The term “imprinted gene” refers to genes whose expression is conditioned by their parental origin. Among theories to unravel the evolution of genomic imprinting, the kinship theory prevails as the most widely accepted, because it sheds light on many aspects of the biology of imprinted genes. While most assumptions underlying this theory have not escaped scrutiny, one remains overlooked: mothers are the only source of parental investment in mammals. But, is it reasonable to assume that fathers' contribution of resources is negligible? It is not in some key mammalian orders including humans. In this research, I generalize the kinship theory of genomic imprinting beyond maternal contribution only. In addition to deriving new conditions for the evolution of imprinting, I have found that the same gene may show the opposite pattern of expression when the investment of one parent relative to the investment of the other changes; the reversion, interestingly, does not require that fathers contribute more resources than mothers. This exciting outcome underscores the intimate connection between the kinship theory and the social structure of the organism considered. Finally, the insight gained from my model enabled me to explain the clinical phenotype of Prader-Willi syndrome. This syndrome is caused by the paternal inheritance of a deletion of the PWS/AS cluster of imprinted genes in human Chromosome 15. As such, children suffering from this syndrome exhibit a striking biphasic phenotype characterized by poor sucking and reduced weight before weaning but by voracious appetite and obesity after weaning. Interest in providing an evolutionary explanation to such phenotype is 2-fold. On the one hand, the kinship theory has been doubted as being able to explain the symptoms of patients with Prader-Willi. On the other hand, the post-weaning symptoms remain as one of the primary concern of pediatricians treating children with Prader-Willi. In this research, I reconcile the clinical phenotype of Prader-Willi syndrome with the kinship theory, contending that paternal investment relative to maternal investment increases after weaning. I also propose a genetic composition of the PWS/AS cluster, discuss the effects of new types of mutations, and contemplate the potential side effects of reactivating silent genes for medical purposes.
| Genomic imprinting refers to genes that are silent when maternally inherited but expressed when paternally inherited, or vice versa. Hailed as the most successful evolutionary explanation for genomic imprinting, the kinship theory contends that the paternally inherited copy of a gene, which determines the allocation of maternal resources to her offspring, is selected to extract more resources than its maternally inherited counterpart. The conflict between genes of different parental origin leads to the silencing of one copy but the expression of the other. As originally formulated, the kinship theory assumes that mothers contribute all resources to the raising of offspring. Yet this is not entirely true, as biparental care is common in some mammals, particularly humans. By positing that fathers contribute some resources and analyzing the effect on the kinship theory, I derived new conditions for the evolution of genomic imprinting and discovered that biparental care does not necessarily increase the opportunities for intragenomic conflict. Interestingly, biparental care allows for the evolution of patterns of expression opposite to the ones originally predicted in the kinship theory. Insight gained from my model explains the clinical phenotype of Prader-Willi syndrome, which has so far been a challenge to the kinship theory. Children suffering from this syndrome exhibit a striking biphasic phenotype characterized by poor sucking and reduced weight before weaning but voracious appetite and obesity afterwards. I argue that, in humans, the paternal contribution increases after weaning. This would explain the evolution of genes with the opposite imprinting pattern before and after weaning, and thus sheds light on the bi-phasic phenotype of Prader-Willi syndrome.
| Imprinted genes violate Mendel's laws by exhibiting an expression conditioned by their parental origin [1]. Either they are silent when maternally inherited (MI) and expressed when paternally inherited (PI), or vice versa [1].
This form of genetic memory captivated the interest of biologists early on. Since its discovery, many theories on the evolution of imprinted genes have been proposed. One of the first theories presents imprinting as an adaptation against ovarian trophoblastic disease [2]. Varmuza and Mann [2] contend that the inactivation of maternally derived genes in oocytes evolved to prevent the development of unfertilized oocytes into ovarian cancer (see [3] for a review of early theories). More recently, Day and Bonduriansky [4] posit that imprinting results from a conflict between genes selected in the opposite sense in each sex. Silencing of the MI copy of a gene is expected to evolve if the selective advantage of that gene through sons is greater than its selective disadvantage through daughters and vice versa. Wolf and Hager [5] claim that imprinting results from the coevolution between genes expressed in the mother and genes expressed in the offspring. Silencing of the PI copy of a gene expressed in the offspring allows the co-adaptation of maternal and offspring traits.
The kinship theory of genomic imprinting (henceforth the kinship theory)—one of the earliest theories on the evolution of imprinting—is currently the most widely accepted [6,7]. The theory's strength lies in its capacity to explain many empirical aspects of the biology of imprinted genes [6–9]. Consider the set of genes expressed in an offspring that influence the allocation of maternal resources. The kinship theory argues that the PI copy of these genes is selected to extract more resources than the MI copy [10]. This is true as long as there is certain asymmetry between matrilines and patrilines, an asymmetry that can be caused by a change in reproductive partners or a male-biased dispersal among others [10]. The kinship theory differentiates two types of gene: those that enhance the allocation of maternal resources when up-regulated, resource enhancers (REs); and those that inhibit the allocation of maternal resources when up-regulated, resource inhibitors (RIs). If the PI copy of an RE is selected for a greater expression than the MI copy, the MI copy will be silenced. If the PI copy of an RI is selected for a lesser expression than the MI copy, the PI copy will be silenced.
The assumptions behind the kinship theory have been intensely scrutinized in the biological literature. One assumption, however, has been largely ignored: mothers are the only source of parental care. In the case of mammals, assuming that fathers contribute little or no resources to their offspring is not always a realistic assumption. Although paternal provision is uncommon among mammals in general (less than 10% of all genera), it is not uncommon in important orders (Perissodactyla, Carnivora, Rodentia, and Primates) (almost 40% of genera within each order) [11]. More importantly, paternal contribution is common in humans [12–17]. Therefore, to understand the evolution of genomic imprinting in mammals, the kinship theory should be extended beyond maternal contribution of resources. I generalized the kinship theory by formulating a model in which both parents can contribute any amount of resources and maternal contribution only is a special case.
I start by discussing the evidence on paternal provision of resources in mammals. Then I elaborate the first model that considers the role of paternal resources in the evolution of genomic imprinting. This model allows me to address two questions: Given biparental care, does genomic imprinting evolve? And, which one of the patterns of expression evolves? In the original formulation of the kinship theory, lifetime monogamy is the only exception to the evolution of imprinting. I find a new condition such that imprinting does not evolve even when there is polygamy. Furthermore, in the original formulation of the kinship theory, for each type of gene, only one pattern of expression is expected to evolve. I conclude that, for some distributions of parental costs, the MI copy of an RE gene becomes silent but, for the rest of the distributions, the PI copy is the one that evolves to be silent. Interestingly, such reversion in the pattern of expression does not require that fathers contribute more resources than mothers. This exciting result illustrates the intimate connection between the expectations of the kinship theory and the social structure of the organism considered.
In the second part of this research, I apply the insight gained from considering paternal contribution of resources to solve one of the challenges to the kinship theory in its original formulation, namely explaining the clinical phenotype of Prader-Willi syndrome (PWS) and Angelman syndrome (AS) [3,18]. Deletion of the PWS/AS cluster of imprinted genes in human Chromosome 15 results in PWS when paternally inherited, but in AS when maternally inherited [1,19]. Children suffering from these syndromes experience feeding, weight, and growth problems, abnormal activity levels, and mental disabilities. The clinical phenotype of PWS children changes dramatically from poor sucking and reduced weight before weaning, to insatiable appetite and obesity after weaning [19–21]. The change is so dramatic that the medical literature describes this syndrome as biphasic [22]. Patients suffering from AS exhibit prolonged sucking—although poorly coordinated—before weaning [20,23] but appetite and weight problems are not constantly present after weaning [19,24].
None of the theories on the evolution of genomic imprinting can explain the clinical phenotype of PWS and AS Syndromes. The kinship theory, however, has shed light on the syndromes' clinical phenotype before weaning [20,25,26]. According to this theory, loss of the PI copy of the PWS/AS cluster implies the loss of all active REs and that PWS children are expected to present a lower than normal acquisition of maternal resources [21]. Loss of the MI copy implies the loss of all active RIs and that AS children are expected to present a greater than normal acquisition of maternal resources [21]. In the original formulation of the kinship theory, no room exists neither for the inversion of the clinical phenotype nor for the symptoms exhibited after weaning by PWS children. Furthermore, it has been pointed out that, after weaning, the clinical phenotype of PWS patients is the opposite to the one predicted by the kinship theory [3,18], which challenges the validity of this theory [3,18].
Haig and Wharton [21] proposed an alternative interpretation of the clinical phenotype that is consistent with the kinship theory. They assume that the more solid food an offspring gets, the less the breast milk it consumes. The substitution of breast milk for solid food would result in a cost to the offspring, because breast milk is of superior quality—both from a nutritional and an immunological perspective—and would result in a benefit to the mother because providing solid food is less costly to her—either because direct provisioning is less costly or because the offspring itself, or other members of the group, contribute to the provision of solid food [21]. If this were the case, a locus controlling appetite for solid food after weaning would be an RE and thus expressed when paternally inherited. Loss of the PI copy would result in greater than normal appetite for solid food and obesity would ensue. It is not clear what type of limitation would result in genes in the offspring being selected to substitute breast milk for solid food as oppose to using solid food as a supplement to breast milk.
In the second part of this research, I contend that the consideration of paternal investment provides a unique insight into the evolution of PWS and AS Syndromes. I reconcile the kinship theory with the clinical phenotype of PWS and AS. I also predict the composition of the PWS/AS cluster of imprinted genes and discuss the evolution of a new type of imprinted gene that exhibits a different pattern of imprint at different moments during development. These results are important not only from a theoretical but also from an applied perspective, as they may contribute to a better understanding of these syndromes. This is important because the frequency of PWS and AS has increased in recent years with the use of assisted reproductive technology [27]. Furthermore, the voracious appetite and obesity exhibited in PWS children after weaning are main concerns of caretakers and pediatricians treating them [19]. The results presented here imply that any attempt to treat these diseases by reactivating silent genes would require considering not only the type of gene reactivated but also when, within developmental time, this gene is expressed. If the time factor is ignored, reactivation could achieve the opposite result to the one expected. These results illustrate how evolutionary theory can have an impact on medicine [28].
In this section, I discuss the evidence of paternal care in mammals. I will concentrate on two aspects of paternal care relevant to the arguments discussed in this paper: whether fathers provide food to their nuclear families and whether this contribution increases after weaning.
The term paternal care is used in the biological literature either in a broad or a narrow sense. In a broad sense, paternal care refers to any action of a male that increases the reproductive success of related or unrelated offspring [11]. This is the definition adopted in most studies of natural history of mammals. In a narrow sense, paternal care refers to any action of a father that increases the reproductive success of his offspring at a cost to the reproductive success of his other offspring [29]. This is the definition adopted in studies of parent–offspring conflict and genomic imprinting. In a narrow sense, the male who acts to increase the reproductive success of an offspring must be the father —not any male— and the action must reduce his ability to invest in his other offspring. This distinction will be relevant when discussing the evidence on paternal care in relation to the evolution of genomic imprinting.
Parental care comes in many forms. When the parental action does not benefit the offspring directly, the term “indirect parental care” is used. Examples of indirect parental care in mammals are resource acquisition and defense, shelter construction and maintenance, and food provision to mothers. When the parental action does benefit the offspring directly, the term “direct parental care” is used. Examples of direct parental care before weaning include gestation and lactation, huddling, and cleaning and transportation of the young; after weaning, they include food provision and defense and socialization of young [11]. In mammals, a marked distinction exists between resources that each parent can provide before and after weaning. While mammalian mothers can provide nutrients directly during the gestation and lactation periods, fathers can only provide food directly after weaning.
Within all mammalian orders, direct paternal care has been observed in 9% of genera [11]. Although this figure suggests that the role of paternal care in the evolution of mammals is negligible, a closer look at its taxonomic distribution provides a very different picture. Direct paternal care has been observed in 33%–34% of genera within the Perissodactyla and Carnivora orders and, more interestingly, in 39% of genera within the Primates [11,30]. Furthermore, it has been suggested that, within the Rodents, the percentage of genera exhibiting paternal care might be greater than any other order; that is at least 40% [11] (Figure 1).
These values should be interpreted cautiously. On the one hand, they are conservative in the sense that the natural history of many mammals is unknown, and the lack of observations on particular genera was recorded as absence of paternal care. On the other hand, these observations correspond to paternal care in the broad sense, and some of them may not qualify as paternal care in the narrow sense. I will further elaborate on the existing evidence on paternal care by discussing the provision of food in two species of non-human mammals—chosen on the basis of data availability— and in human hunter-gatherer societies.
In this section, I generalize the kinship theory [40] by allowing the expression of a gene in an offspring to affect both maternal and paternal investment (as opposed to maternal investment only). The model here presented (see the Methods section for a detailed formulation) relates to a class of models formulated to study either parent-offspring conflict [29,41,42] or the kinship theory [40]. All these models assume that parental care is limited to maternal care, except a model for parent-offspring conflict formulated by Parker [43].
Consider a nuclear family formed by a mother, a father, and their offspring. Both parents have a limited amount of resources to invest either in the current or future offspring. Consider a gene expressed in the current offspring that affects the amount of resources provided by the mother and the father to the current offspring. In particular, a greater expression of the locus containing this gene results in a greater investment of both parents—as opposed to a reduced provision of resources.
Consider a mutant that modifies the expression of either the MI or the PI copy of an RE. The change in expression at this locus results in a benefit BO to the current offspring at a cost to its parents, CM, CP. The parental cost can be subdivided into the cost derived from a greater investment, CMi, CPi, and the investment elicited by a greater expression, IMx, IPx, that is CM = CMiIMx, CP = CPiIPx. Hence, I will talk about maternal CMi and paternal CPi cost of investment and about maternal CM and paternal CP cost of expression. Benefits and costs refer to changes in the number of offspring of the current offspring, and in the number of grand offspring of the current offspring's parents respectively.
The extent to which the costs that the father incurs translate into maternal costs, κM, and the costs that the mother incurs translate into paternal costs, κP, are determined by the mating system of the population. For example, if there is lifetime true monogamy κM = κP = 1. If mothers die before fathers and fathers replace their partners κM ≤ 1 and κP = 1 with κM being smaller as the cost of finding a new partner becomes smaller. If parents change partners for the production of every offspring, κM = κP = 0. In theory, it is possible that κM, κP > 1 . However, the biological conditions for this happening are rare (see discussion in [44]) and will be ignored in this research, that is, 0 ≤ κM, κP ≤ 1.
Consider a population in which the resident gene shows the same level of expression when maternally and paternally inherited. The fate of a rare mutant that, when maternally inherited, increments the expression of the RE locus considered is determined by its inclusive fitness effect:
The inclusive fitness effect of a mutant acting on the MI copy is given by the difference between the benefit experienced by the current offspring and half of the cost that the mother incurs. Such cost has two components: the maternal cost of expression, CM, and the paternal cost of expression, CP, multiplied by the extent in which costs borne by the father translate into maternal costs, κM. While benefits are weighed by the probability that the allele expressed in the current offspring is present in itself, namely one, costs are weighed by the probability that the allele expressed in the current offspring is present in its siblings, namely half.
The fate of a rare mutant that, when paternally inherited, increments the expression of the RE locus considered is determined by its inclusive fitness effect:
The inclusive fitness effect of a mutant acting on the PI copy is given by the difference between the benefit experienced by the current offspring and half of the cost that the father incurs Such cost has two components: the cost the father incurs due to enhanced paternal investment elicited by up-regulated expression, CP, and the cost the mother incurs due to enhanced maternal investment elicited by up-regulated expression, CM, multiplied by the extent in which costs borne by the mother translate into paternal costs, κP.
Let the ordered pair
correspond to the level of expression of the MI and PI copies of the allele fixed in the population and x* correspond to the combined level of expression of both copies. The optimal level of expression from the perspective of the MI copy
can be derived from imposing condition Wm = 0 in Equation 1. The optimal level of expression from the perspective of the PI copy
can be derived from imposing condition Wp = 0 in Equation 2.
When
, there is no conflict between the MI and PI copies and imprinting does not evolve .When
, there is conflict and imprinting does evolve. If
, the intralocus conflict results in the evolutionarily stable pattern of expression
. If
, the intralocus conflict results in the evolutionarily stable pattern of expression
. The direction of the imprint—either MI copy silent or PI copy silent—is neatly summarized by the “loudest voice prevails” principle [10]: natural selection favors silencing of the copy whose optimal level of expression is lower, and expression, at its optimal level, of the allele whose optimal level of expression is greater [10].
The difference between the inclusive fitness effect of the maternally and paternally inherited copies
characterizes not only whether there is intralocus conflict but also, in case of conflict, what will be the evolutionarily stable pattern of expression. If Wm − Wp = 0, then
and there is intralocus conflict. If Wm − Wp > 0, then
and silencing of the paternally inherited copy evolves. If Wm − Wp < 0, then
and silencing of the maternally inherited copy evolves.
When paternal investment is negligible, CP ≈ 0 , silencing of the maternally inherited copy
evolves unless all maternal offspring are sired by the same father κP = 1. This is the condition derived by Haig [10]. When maternal investment is negligible, CM ≈ 0, silencing of the paternally inherited copy
evolves unless all paternal offspring are born to the same mother κM = 1.
When both paternal and maternal investment are non-negligible, genomic imprinting does not evolve when there is lifetime true monogamy κM = κP = 1, and when the paternal cost of expression that is exclusive to the father (not shared by the mother) is equal to the maternal cost of expression that is exclusive to the mother (not shared by the father),
(Figure 2). More interestingly, both patterns of imprint—MI copy silent and PI copy silent—can evolve. If the paternal cost of expression that is exclusive to the father is greater than the maternal cost of expression that is exclusive to the mother, (1 − κM)CP > (1 − κP)CM, then silencing of the PI copy evolves. However, if the maternal cost of expression that is exclusive to the mother is greater than the paternal cost of expression that is exclusive to the father, (1 − κM)CP < (1 − κP)CM, then silencing of the MI copy evolves.
Assuming that the fraction of nutrients provided by the mother to the current offspring is a constant σ, it is possible to determine the critical fraction of resources that contributed by the mother would result in the extinction of the intralocus conflict
where CMi and CPi are the maternal and paternal costs of investment in the current offspring. Using the language of the “loudest voice prevails” for values of σ lower than σ̂)
σ̂, the maternal voice within the offspring speaks louder than the paternal voice, and silencing of the PI copy evolves. For values of σ greater than
σ̂), the maternal voice within the offspring speaks softer than the paternal voice, and silencing of the MI copy evolves (Figure 3).
It is worth noticing that the extinction of the intralocus conflict does not require that both parents contribute the same amount of resources. More interestingly, the intralocus conflict becomes extinct when the mother contributes more resources than father,
σ̂ > ½, if the paternal cost of investment that is exclusive to the father is greater than the maternal cost of investment that is exclusive to the mother,
(Figure 3). This might be of particular interest, because in nature, it might be found that mothers contribute more resources than fathers in many cases.
If a greater expression of the gene considered results in a lower provision of resources, symmetric results can be derived. The only difference being that the current offspring incurs in a cost (CO = −BO) , while its parents experience a benefit (BM = −CM, BP = −CP). The expected patterns of expression are opposite to the ones derived for an RE. If the paternal benefit of expression that is exclusive to the father is greater than the maternal cost of expression that is exclusive to the mother, (1 − κM)BP > (1 − κP)BM, then silencing of the MI copy evolves. However, if the maternal cost of expression that is exclusive to the mother is greater than the paternal cost of expression that is exclusive to the father, (1 − κM)BP < (1 − κP)BM, then silencing of the PI copy evolves.
The kinship theory of genomic imprinting was formulated under the assumption that only mammalian mothers invest in their offspring. Although biparental care is not common among mammals (less than 10% of genera), it is common in certain orders (close to 40% of genera in Perissodactyla, Carnivora, Rodentia, and Primates) [11] and more notably, observed in humans. Therefore, to understand the evolution of genomic imprinting as a result of intralocus conflict, the kinship theory should be extended beyond maternal contribution of resources only. In this research, I generalize the kinship theory by elaborating a model in which both parents can contribute any amount of resources, and maternal contribution only is a special case.
In the original formulation of the kinship theory, genomic imprinting evolves unless there is lifetime monogamy. The introduction of paternal care allows an alternative scenario when genomic imprinting does not evolve. If the cost experienced exclusively by the father due to the expression of a PI gene in the current offspring is equal to the cost experienced exclusively by the mother due to the expression of an MI gene in the current offspring, then deviations from an unimprinted expression are not favored by natural selection. An example to illustrate this point may shed light on what this condition entails. Consider a family in which the mother contributes most of the food, 75% (Figure 2). Assume that fathers (reproducing males) produce more offspring than mothers (reproducing females); notice that the average reproductive success of males does not have to be equal to the average reproductive success of females. While mothers conceive most of those offspring with the same father, 75% of them, the fraction of offspring that fathers conceive with the same mother represent 50% of the fathers' total production. Assume that the mother contributes food sources such as roots and fruits, and the father contributes meat. Hunting is more risky, 1.5 times more, than gathering food.
In this scenario, an additional contribution of resources to the current offspring reduces the amount of resources available to future offspring. The father provides 1/4 of the additional resources and suffers a cost four times the additional investment in the offspring. Half of the time, the cost experienced by the current father is not shared by the current mother. Hence the cost that the father incurs—due to the expression of a PI gene in the current offspring—that is not shared by the mother is (1 − κM)CPiIPx = (1/2) 4 (1/4) = (1/2). On the other hand, the mother provides 3/4 of the additional resources and suffers a cost that is 8/3 times the additional investment in the current offspring. One-fourth of the time, the cost experienced by the current mother is not shared by the current father. The cost the mother incurs—due to the expression of an MI gene in the current offspring—that is not shared by the father is (1 − κP)CMiIMx = (1/4) (8/3) (3/4) = 1/2. Given that, in this case, (1 − κM)CP = (1 − κP)CM, there is no intralocus conflict between the MI and the PI and genomic imprinting would not evolve.
In the original formulation of the kinship theory, there is only one direction of imprint for each kind of gene. If the gene considered is an RE, it is expected to be maternally silent. However, if the gene considered is an RI, it is expected to be paternally silent. The introduction of paternal care allows that the direction of the imprint is reversed for each kind of gene. Consider an RE, a gene is expected to be paternally silent if (1 − κM)CP > (1 − κP)CM, but maternally silent if (1 − κM)CP < (1 − κP)CM. Consider an RI, a gene is expected to be maternally silent if (1 − κM)BP > (1 − κP)BM, but paternally silent if (1 − κM)BP < (1 − κP)BM. Interestingly, I found that it is not necessary that fathers contribute more resources than mothers for this change in the direction of the imprint to happen.
If the cost experienced exclusively by the father due to his own investment is greater than the cost experienced exclusively by the mother due to her own investment, then an RE evolves to be paternally silent, even if the mother is contributing more resources than the father. I will return to the previous example to illustrate this condition. The cost the current father incurs—due to his own investment—that is not shared by the current mother is (1 − κM)CPi = (1/2) 4 = 2. The cost the current mother incurs—due to her own investment—that is not shared by the father is (1 − κP)CMi = (1/4) (8/3) = 2/3. Thus (1 − κM)CPi > (1 − κP)CMi and even if the mother contributes more than half the resources, an RE may still evolve to be paternally silent. In particular, if the mother contributes less than 75% of the resources, an RE evolves to be paternally silent. But if the mother contributes more than 75%, an RE evolves to be maternally silent.
The kinship theory requires genes expressed in the offspring that can affect the allocation of parental resources. During gestation, the interaction between mother and offspring is chemical—mediated by hormones and proteins, some of them secreted into the maternal bloodstream. During lactation, however, the interaction is mostly behavioral—mediated by actions performed by the offspring. In both cases, the relation between mother and offspring is intimate [48]. After weaning, the interaction between mother and offspring continues to be behavioral but becomes less intimate [48]. In general, genes expressed in the offspring can affect the allocation of maternal resources before birth by eliciting the segregation of hormones into the maternal bloodstream and after birth by eliciting behavioral acts that produce a maternal response. It is difficult to think how genes expressed in the offspring may influence the allocation of paternal resources during gestation. The only possibility would be that hormones secreted into the maternal blood stream elicit a maternal behavior that produces a paternal response. Before weaning, the contribution of food by the father is given indirectly through the mother. If a gene expressed in the offspring were to influence the allocation of paternal resources, it would have to elicit a behavior that produces a paternal provision of food to the mother who would be responsible for transferring such resources to the offspring. Furthermore the intimacy of the relation between mother and offspring would make mothers more likely to respond to any behavioral act of the offspring. After weaning, the food contribution by both parents is direct. Genes expressed in the offspring will have the same means and opportunities to influence each parent provision of nutrients.
From a taxonomic point of view, genomic imprinting was initially expected to be circumscribed to mammals and flowering plants. Such taxonomic distribution responds to the possibility of interaction between offspring and mother during gestation in eutherians and angiosperms. There is no room for interaction between mother and offspring during gestation in birds, most reptiles, and amphibians. Thus genomic imprinting was not expected to evolve in these classes. However it was soon noticed that after birth, the offspring can influence the allocation of maternal resources in a much more broad variety of organisms. The only condition is that there is extended postnatal parental care. Therefore imprinting is expected to be found in genes expressed after birth in mammals and after hatching in birds and fishes [48,49]. While biparental care is important in some mammalian orders, it is much more abundant in birds. The consideration of biparental care does change the pattern of imprint expected to evolve; in principle, it should not affect the expected taxonomic distribution of imprinting before and after birth. The only exception would occur in the particular case in which the life history and qualities of both parents as resource providers are equivalent, that is (1 − κM)CP = (1 − κP)CM. In such case, genomic imprinting would not evolve.
The kinship theory, when extended to allow biparental care, provides a unique insight on the evolution of PWS and AS. I argue that the relative contribution of resources made by mammalian fathers increases after weaning, in particular that σb >
σ̂ > σa. If this is the case, an RE expressed before weaning is expected to be maternally silent, but another RE expressed after weaning is expected to be paternally silent. An RE expressed before and after weaning that can adjust its expression over time is expected to be maternally silent before weaning and paternally silent afterward (Figure 7). Similarly an RI expressed before weaning is expected to be paternally silent, but another RI expressed after weaning is expected to be maternally silent. An RI that is expressed before and after weaning is expected to be paternally silent before weaning and maternally silent afterwards (Figure 7). This explains that the loss of the PI copy of the PWS-AS cluster will result in a deficit of food intake and weight before weaning and an excess afterwards. The loss of the MI copy of this cluster will result in an excess of food intake before weaning and a normal food intake afterwards—if the genes expressed after weaning are RIs only. This matches the clinical phenotype of children with PWS and AS concerning appetite and weight. While the kinship theory can account for the aspects of PWS and AS concerning appetite and weight, there are other aspects of these disorders, mostly mental problems, that are difficult to interpret straight away.
Our understanding of the evolutionary forces and types of genes involved in the post-weaning phenotype of children with PWS might be important for the treatment of this disease. If the post-weaning phenotype is caused by an RI that is maternally silent, as I suggest, it could be possible to re-activate the maternal copy. However, it should be re-activated after weaning only and not before, because this would exacerbate the condition of children with PWS before weaning.
The phenomenon I am describing does not need genes that are expressed before and after weaning and thus change the direction of the imprint. It is enough to have a gene expressed before weaning and another one expressed afterwards. However my model does suggest the possibility that genes evolve to change the direction of the imprint from maternally silent to paternally silent or vice versa. I refer to these genes as flip-flop genes. While there is no evidence indicating that flip-flop genes may exist, I do not know of any attempt to find them. Interestingly, it is known that some genes can change the direction of the imprint depending on the tissue in which they end up being expressed. This is the case of gene Grb10 in mouse, which is expressed from two promoters with opposite pattern of imprint, namely the paternal allele remains silent in most tissues but the maternal allele remains silent in the brain [50,51]. In both cases, the protein product is the same. If it is possible that a gene uses promoters that are imprinted in the opposite direction in different tissues, my prediction of imprinted genes using promoters that show the opposite pattern of imprint at different moments in developmental time does not seem too far fetched. If this were the case, it would be possible to observe a new type of mutation, one that fails to use the correct promoter in the correct developmental time. Such a mutation would result in complicated clinical phenotypes.
This analysis can be extended to other diseases in which imprinted genes are involved. Genes that affect the contribution of resources after weaning may not follow the standard pattern of expression (maternally silent REs and paternally silent RIs). PWS is unique because it is one of the few diseases caused by imprinted genes that affect post weaning growth. While Beckwith-Wiedemann and Silver-Russell syndromes are caused by imprinted genes, both involve growth disorders that manifest themselves during gestation and before weaning but not afterward. In this sense, PWS provides a unique insight into the forces acting on imprinted genes. However the characterization of the function of imprinted genes will allow us to find more genes that affect provision of parental resources after weaning. In this case, the role of fathers will become relevant.
Consider a family formed by a mother, a father, and an offspring. Both parents' limited amount of resources can either be invested in the current offspring or kept for future ones. All individuals are diploid, and the population is at equilibrium. Let xm and xp be the level of expression of the maternally and paternally inherited copies of an allele in the current offspring, where xm, xp ≥0. The aggregate level of expression x = xm + xp determines the amount of maternal iM and paternal iP investment in the current offspring, iM = gM(x) and iP = gP(x) , and the residual investment of each parent in future offspring. For simplicity, I assume that the aggregate level of expression x determines the total amount of parental investment i, where i = iM + iP and i = g(x), and the investment of each parent is a constant fraction of the total investment, iM = σg(x) and iP = (1 − σ)g(x). Such simplifying assumption ignores the scenario in which a gene expressed in the offspring can affect the contribution of maternal and paternal resources in the opposite sense; it is no longer possible that diM/dx > 0, diP/dx < 0 or diM/dx < 0, diP/dx > 0. This assumption is more plausible from a biological perspective and does not influence the model in any other way. Following the literature on genomic imprinting, I use the term resource enhancer (RE) to refer to a gene whose greater expression results in an enhanced allocation of parental resources, di/dx > 0; and resource inhibitor (RI) to refer to a gene whose greater expression results in a reduced allocation of parental resources, di/dx < 0.
Parental investment in the current offspring determines not only the fitness of the offspring vO (defined as the number of offspring left by the current offspring), vO = hO(iM, iP), but also the residual fitness of the mother vM and father vP (defined as the number of grand offspring, other than those produced by the current offspring, left by each of the parents of the current offspring), vM = hM(iM, iP) and vP = hP(iM, iP) . Consider a population fixed for a particular allele such that the aggregate level of expression at a loci homozygous for the resident allele is x*. Consider the fate of a rare epimutation that modifies the expression of the resident allele either when maternally inherited or when paternally inherited. The aggregate level of expression at a loci heterozygous for any of these mutants is x. Any change in expression can produce a change in fitness of the same dvO/dx > 0 or the opposite sign dvO/dx < 0. I use the term benefit B to refer to the first one and cost C to refer to the second one. If considering a RE the offspring derives a benefit at a cost to its parents. The offspring benefit is defined as the change in offspring fitness caused by a change in expression evaluated at x*, BO = dvO/dx|x=x*. The parental cost is the change in parental fitness caused by a change in expression evaluated at x*, CΩ = −dvΩ/dx|x=x* where Ω ∈{M,P}. The parental cost derived from a change in expression can be subdivided into two components: the cost derived from a greater investment, CΩi = ∂vΩ/∂iΩ, and the change in investment caused by a greater expression, −IΩx = ιΩ(di/dx)|x=x*, that is CΩ = CΩiIΩx. Notice that if considering an RI, it is the parents that derive a benefit, BM and BP, at a cost to their offspring CO.
Define κP as a measurement of how change in maternal fitness affects paternal fitness, κP = dvP/dvM, and κM as a measurement of how change in paternal fitness affects maternal fitness, κM = dvM/dvP where 0 ≤ κM, κP ≤ 1.
The inclusive fitness of the MI copy in the current offspring vm is given by the number of offspring produced by the current offspring, plus the number of offspring produced by future maternal siblings weighed by the probability that the MI allele in the current offspring is present in future maternal siblings, vm = vO + ½vM. The inclusive fitness of the PI copy in the current offspring vp is given by the number of offspring produced by the current offspring, plus the number of offspring produced by future paternal siblings weighed by the probability that the PI allele in the current offspring is present in future maternal siblings, vp = vO + ½vP.
The inclusive fitness effect of the MI copy, Wm = dvm/dxm|x=x* = dvm/dx|x=x*, is:
and the inclusive fitness effect of the PI copy, Wp = dvp/dxp|x=x* = dvp/dx|x=x*, is:
An expression pattern of the MI and PI copies of an allele is evolutionarily stable
when adopted by most individuals in the population is resistant to invasion by any rare alternative strategy
or
. In mathematical terms, this occurs when all the components of vector
satisfy one of the following criteria [52]: (1)
is a local fitness maximum, that is Wξ = 0 and ∂Wξ/∂x|x=x*<0, (2)
is a corner solution, that is
= 0 and Wξ < 0 (where ξ∈{m,p}).
Imposing condition Wm = 0 on Equation 7 yields the aggregate level of expression that is optimal from the perspective of the MI copy,
,
and imposing condition Wp = 0 on Equation 8 yields the aggregate level of expression that is optimal from the perspective of the PI copy,
,
Consider that the initial condition is an equal expression of the MI and PI alleles
(unimprinted pattern of expression). When the optimal level of expression for the MI and PI alleles coincide,
, there is no conflict. Selection does not favor deviations from the initial conditions, and the unimprinted pattern of expression is evolutionarily stable. When the optimal level of expression for the MI and the PI alleles diverge,
, there is conflict. Selection favors deviations from the unimprinted pattern of expression. If
, the only pattern of expression that satisfies the conditions for evolutionary stability is
. If
, however, this pattern of expression is
.
The difference
characterizes not only whether there is conflict between the MI and the PI copies but also, in case of conflict, what is the evolutionarily stable pattern of expression. Assume that for each function Wξ there is only one value
that satisfies condition Wξ = 0. If Wm − Wp = 0 then
and there is no conflict. A positive sign, Wm − Wp > 0, implies that
but a negative sign, Wm − Wp < 0, implies that
.
No intragenomic conflict: By imposing condition Wm = Wp on Equation 11, it is possible to conclude that there will be no conflict when:
This condition is satisfied when one of the following is true: (1) κM = κP = 1; that is, changes in maternal fitness perfectly affect paternal fitness while changes in paternal fitness perfectly affect maternal fitness, for example in the case of lifetime true monogamy; (2) CM = CP = 0 where CM = CMiIMx and CP = CPiIPx; (2a) IMx = IPx= 0, that is, the expression of the gene considered does not influence parental investment; (2b) IMx = 0 and CPi = 0; (2c) CMi = 0 and IPx = 0, that is, the expression of the gene considered does not influence the investment of one of the parents while the investment of the other father is cheap, meaning that its investment does not translate in any cost for the provider; (2d) CMi = CPx = 0 investment is cheap for both parents; (3) (1 − κM)CP = (1 − κP)CM, that is, the fraction of paternal cost caused by paternal investment that does not translate into maternal cost, has to be equal to the fraction of maternal cost caused by maternal investment that does not translate into paternal cost.
Substituting CM = CMiIMx and CP = CPiIPx in Equation 12 and keeping in mind that IMx = σ (∂g/∂x) and IPx = (1 − σ) (∂g/∂x), it is possible to determine the fraction of the total resources contributed by the mother, σ, which extinguishes the conflict:
Such a point exists as long as the fitness functions are continuously differentiable. If σ =
σ̂, genomic imprinting does not evolve.
It is worth noticing that the conflict's extinction does not require both parents to contribute the same amount of resources. In particular, if the cost of paternal investment experienced by the father but not shared with the mother is greater than the cost of maternal investment experienced by the mother but not shared with the father, (1 − κM)CPi > (1 − κP)CMi, then the extinction of the conflict happens for a value σ such that mothers contribute more resources than fathers,
σ̂ > ½, and vice versa. This might be relevant in nature because mothers contribute more resources than fathers in many cases.
Intragenomic conflict: If Wm ≠ Wp there is conflict between the MI and PI copies of an RE and genomic imprinting evolves. If (1 − κM)CP > (1 − κP)CM then Wm > Wp and the pattern of expression
is the only evolutionarily stable stategy (ESS). If (1 − κM)CP < (1 − κP)CM, however, then Wm < Wp and the pattern of expression
is the only ESS.
Similarly, if σ ≠
σ̂, there is conflict between the MI and PI copies of an RE and genomic imprinting evolves. If σ <
σ̂, then
, and the pattern of expression
is the only ESS. If σ >
σ̂, then
, and the pattern of expression
is the only ESS.
Even if the fitness functions are not continuous in
σ̂, this value still marks the boundary between the two patterns of expression,
and
. This can be concluded from Equation 11 by noticing that for an RE, if there were maternal care only σ = 1, then
. However, if there were paternal care only σ = 0 then
.
If the gene considered were an RI, then Wm − Wp= ½ [(1 − κP)BM − (1 − κM)BP] and the expected patterns of expression are the opposite to the RE ones. If (1 − κM)BP > (1 − κP)BM, then Wm < Wp and the pattern of expression
is the only ESS. If (1 − κM)BP < (1 − κP)BM, however, then Wm > Wp and the pattern of expression
is the only ESS. Similarly, if σ <
σ̂ then
and the pattern of expression
is the only ESS. If σ >
σ̂ then
and the pattern of expression
is the only ESS.
All previous results apply when, given a level of expression, parental investment does not change over the window of time when the gene is expressed. This might be the case either because parental investment does not change over developmental time altogether or because parental investment does change, but not within the window of time when the gene is expressed. In this section, I will discuss the case when, for the same level of expression, each parent invest more or less at different developmental stages.
Inclusive fitness: Let t represent time in the life of the current offspring, and [0,T]g be the window of time when gene g is expressed. The relative investment of each parent is a function of time, σ = fM(t); therefore iΩ = fΩ(t)g(x). The inclusive fitness effect of the MI allele when the gene is expressed within the window of time [0,T] is:
and the inclusive fitness effect of the PI allele is:
For simplicity, I consider that parental investment may differ between two periods of time, namely, before weaning t < tw and after weaning t > tw. Given a certain level of expression x, the amount of resources contributed by each parent remains constant within each period but may change between periods, that is iM = σbg(x) for t ≤ tw but iM = σag(x) for t > tw.
Evolutionary stable level of expression: The conditions for evolutionary stability are the conditions provided in the previous section; however, now the vector considered contains two vectors—one for the pattern of expression before weaning and another one for the pattern of expression afterwards. Which pattern of expression evolves depends on whether the expression of the gene considered can be modified during each period of time.
Fixed expression: Assume that the gene's level of expression is unable to change through developmental time. Let μ = tw/T be the fraction of time before weaning within the window of time in which the gene considered is expressed. There is no conflict between the MI and PI alleles whenever σba =
σ̂, where σba = μσb + (1−μ)σa is the mean maternal investment over the window of time when the gene considered is expressed. Intragenomic conflict occurs when σba ≠
σ̂. Consider an RE. If σba<
σ̂ then
and the pattern of expression
is the only ESS. If σba >
σ̂ then
and the pattern of expression
is the only ESS. Therefore, whether a gene evolves pattern
or
depends on whether the window of time when the gene considered is expressed occurs mostly before or after weaning.
Variable expression: Assume that the level of expression of the gene considered can be adjusted over developmental time. No conflict exists between the MI and PI alleles whenever σb = σa =
σ̂. Let the two ordered pairs
represent the level of expression of the MI and PI copies before and after weaning respectively. There is intragenomic conflict when either σb ≠
σ̂ or σa ≠
σ̂ or both. Consider an RE. If σb >
σ̂ and σb <
σ̂ then before weaning
but after weaning
, and the pattern of expression
is the only ESS. If σb <
σ̂and σa >
σ̂ then before weaning
but after weaning
, and the pattern of expression
is the only ESS. These kinds of genes will show two different patterns of expression before and after weaning.
|
10.1371/journal.pcbi.0030192 | Metabolic Reconstruction and Modeling of Nitrogen Fixation in Rhizobium etli | Rhizobiaceas are bacteria that fix nitrogen during symbiosis with plants. This symbiotic relationship is crucial for the nitrogen cycle, and understanding symbiotic mechanisms is a scientific challenge with direct applications in agronomy and plant development. Rhizobium etli is a bacteria which provides legumes with ammonia (among other chemical compounds), thereby stimulating plant growth. A genome-scale approach, integrating the biochemical information available for R. etli, constitutes an important step toward understanding the symbiotic relationship and its possible improvement. In this work we present a genome-scale metabolic reconstruction (iOR363) for R. etli CFN42, which includes 387 metabolic and transport reactions across 26 metabolic pathways. This model was used to analyze the physiological capabilities of R. etli during stages of nitrogen fixation. To study the physiological capacities in silico, an objective function was formulated to simulate symbiotic nitrogen fixation. Flux balance analysis (FBA) was performed, and the predicted active metabolic pathways agreed qualitatively with experimental observations. In addition, predictions for the effects of gene deletions during nitrogen fixation in Rhizobia in silico also agreed with reported experimental data. Overall, we present some evidence supporting that FBA of the reconstructed metabolic network for R. etli provides results that are in agreement with physiological observations. Thus, as for other organisms, the reconstructed genome-scale metabolic network provides an important framework which allows us to compare model predictions with experimental measurements and eventually generate hypotheses on ways to improve nitrogen fixation.
| Nitrogen fixation is an important process for improving plant development in crops. Overall, it constitutes a central role in the nitrogen cycle which is essential to life. In this work we were interested in understanding nitrogen fixation in Rhizobium etli from a genome-scale perspective. Using the genome annotation and scientific literature, we reconstructed the metabolic network for R. etli, a bacterium that fixes nitrogen. The reconstructed metabolic network was used to analyze how this network is utilized during nitrogen fixation. From this metabolic network, we built a model that was found to be in agreement with the general biochemical properties of Rhizobia, when it fixes nitrogen. Additionally, we have included an analysis of how gene deletions affect symbiotic nitrogen fixation. We propose that the metabolic reconstruction presented here can be useful as a theoretical template to understand and suggest a hypothesis for improving nitrogen fixation and its biochemical interaction with plants.
| Free-living bacteria belonging to the Rhizobium genera are often symbiots associated with plants of the family leguminosae. These bacteria differentiate and have the ability to fix atmospheric nitrogen into ammonium when some compounds are exchanged between the bacteroid and its plant host [1–3]. Establishment of effective symbiotic nitrogen fixation between plant and bacteria is a complex process whose understanding constitutes an interesting scientific challenge with clear implications in plant development sciences and agriculture [4]. Nitrogen fixation in Rhizobium involves a complex plant–bacteria symbiotic relationship orchestrated by the genetic and metabolic networks of both organisms [3]. In general, the plant supplies carbon sources and glutamate to the bacteroid, while the bacteroid in turn provides the plant with ammonium, aspartate, and alanine [5,6] (see Figure 1). The exchange of these compounds creates a dependent symbiotic relationship between these two organisms whose effectiveness is essential to improving plant growth and bacterium survival.
Rhizobium etli CFN42 is a nitrogen-fixing bacterium whose genome annotation has been reported recently [7]. Nitrogen fixation in R. etli occurs in the last of three developmental stages. The first developmental stage is related to the infection process of plant roots by Rhizobium. This begins when the plant excretes flavonoids which signal nodule formation [2]. The second stage is characterized by bacterial growth inside the plant, and the construction of a compartmentalized globular structure, called a nodule. This specialized structure protects nitrogenase, a key enzyme in nitrogen fixation, against irreversible oxidative damage by oxygen [5]. Finally, the last stage involves differentiation of the bacteria into a bacteroid able to reduce atmospheric nitrogen inside the nodule [8]. The computational analysis of this last stage is the main focus of this work.
Systemic understanding of nitrogen fixation in R. etli requires the construction of a model able to integrate genomic and high-throughput data in a hierarchical and coherent fashion [9]. Integrative models of this sort constitute a powerful and elegant strategy to study the mechanism of cell behavior. In particular, constraint-based models constitute such an approach, with a capacity to predict organism phenotypes operating at steady state [10–13]. Here, we present a reconstruction of the metabolic network in R. etli, the first reconstruction made for a Rhizobium organism. A constraint-based approach [12], including flux balance analysis (FBA) [14], is used to analyze the physiological capability of the bacterium when it fixes nitrogen. To show the utility of this analysis, the consistency between model predictions with experimental observations in some metabolic pathways is evaluated. Then we analyze the effects that some gene deletions have on symbiotic nitrogen fixation and compare them with available experimental observations. Experimental evidence on how these gene deletions affect nitrogen fixation activity are available for most cases investigated computationally, and it provides important information to validate our in silico modeling.
The metabolic reconstruction was generated from the KEGG annotated genome sequence for R. etli [7], journal publications, automated reconstruction databases [15], and information found in biochemical textbooks on nitrogen fixation [1]. Thus, our metabolic reconstruction includes reactions with evidence from the genome annotation or with clear experimental evidence for Rhizobia. The resulting reconstructed metabolic network for R. etli includes 387 reactions involving 371 metabolites and 363 genes. This reconstruction spans 26 metabolic pathways involving central metabolism (44 reactions), amino acids metabolism (136 reactions), purine and pyrimidine metabolism (89 reactions), PHB synthesis (8 reactions), and nitrogen metabolism (19 reactions). The properties of the network and the complete set of metabolic reactions with their corresponding gene–protein reaction associations are available in Dataset S1, Table S1, and Dataset S2, respectively. Nomenclature used for metabolites is included in Dataset S3. Figure 2 shows a metabolic map of the pathways present in the reconstruction. The journal publications supporting our metabolic reconstruction are reported in Dataset S4. The gap analysis of the metabolic reconstruction is reported in Dataset S5.
The characterization of the physiological capability of R. etli and prediction of ways to improve nitrogen fixation in the bacteroid are central themes in this study. For our analysis, we assume that symbiotic nitrogen fixation between the plant and bacterium has reached the steady state. To simulate nitrogen fixation during symbiosis, we have constructed an objective function (OF) representing a set of chemical compounds whose production in the bacteroid is essential to make this symbiotic process efficient. Instead of incorporating biomass components (phospholipids, proteins, DNA, RNA) for the organism into the OF, as has been done for other organisms such as Escherichia coli, only compounds which are known or thought to be produced during symbiotic nitrogen fixation are included. The main reason for this is that the bacteroid does not grow during the stage where nitrogen fixation occurs, which is the life stage of interest in this study.
After reviewing the available literature, we can postulate an OF for use in FBA, which represents symbiotic nitrogen fixation in R. etli. This OF is based on the following physiological information.
1) Plant–bacteroid exchange of some amino acids during nitrogen fixation has been suggested to be a general mechanism in Rhizobia. Aspartate and alanine play a central role in the development of nodules and in efficient nitrogen fixation [6]. According to a recent hypothesis, aspartate and alanine are provided to the plant from the bacteroid, while glutamate is supplied by the plant to the bacteroid [6,16]. In this context, we postulate that symbiotic nitrogen fixation is closely related to the efficient supply of alanine and aspartate from the bacteroid to the plant. L-lysine has been reported in Bradyrhizobium japonicum bacteroids, a related organism to R. etli, and this amino acid was also included in the OF [17].
2) Nitrogen fixation is energetically costly requiring 16 ATP molecules to reduce a di-nitrogen molecule into two molecules of ammonium, which are then exported to the plant [2,5]. The transport of ammonium from the bacteroid to the plant is crucial for establishing symbiosis with the plant.
3) During nitrogen fixation, some Rhizobia accumulate poly-ß-hydroxybutyrate (PHB) and glycogen [18,19]. These compounds serve as storage for carbon that can be used when others are not present [20,21] . The presence of PHB has been verified in R. etli in the free-living and bacteroid states [22]. Even though the synthesis of PHB requires energy, PHB production is important because it reduces the levels of NAD(P)H which can repress the activity of some enzymes in the citric acid (TCA) cycle [21]. Nodules in R. etli are characterized as determinate nodules, and there is experimental evidence that glycogen and PHB molecules are produced during nitrogen fixation [5]. Thus, we consider the production of PHB and glycogen to be important components of the symbiotic nitrogen fixation OF.
The identified essential components produced by the bacteroid and exchanged with plant during nitrogen fixation are depicted in Figure 1. The OF used to represent symbiotic nitrogen fixation is a linear combination of all the components noted above.
Thus, by maximizing this OF we suggest that during nitrogen fixation steady-state fluxes maximize the equal molar production of these compounds. In the rest of this work we will refer to the OF as the function that maximizes symbiotic nitrogen fixation. FBA consists of finding a flux distribution subject to steady-state mass balance and thermodynamic constraints, such that a linear OF is maximized [14]. FBA was performed (see Methods section), and the resulting flux distribution was compared with the known physiology of the organism. To verify the robustness of our results, we also ran the analysis using 1,000 different randomly weighted OFs (with weights selected from a uniform distribution) that did not assume equimolar contributions of the components shown in Equation 1 (see Methods section). We observed that the identified inactive metabolic pathways were independent of the stoichiometric coefficients of the OF (see Figure S1). The exchange reactions and the corresponding constraints used in simulations can be found in Dataset S2 and Table S2.
R. etli, like most other Rhizobia, fixes nitrogen in a microaerobic environment [23,24]. We simulated microaerobic conditions by constraining the upper bound of oxygen uptake rate to be 1 mmol/gDW/hr, a range estimated from experimental reports [25]. The simulations suggest that under microaerobic conditions, symbiotic nitrogen fixation activity is possible and some patterns for using the central metabolic pathways can be observed. During nitrogen fixation, dicarboxylates (succinate) provided by plant are metabolized via the TCA cycle. Efficient nitrogen fixation requires a high degree of coordination between the TCA cycle and oxidative phosphorylation [5].
Bacterial respiration is important for nitrogen fixation for two reasons: it (1) reduces oxidative damage to nitrogenase by consuming oxygen, and (2) simultaneously generates ATP molecules needed for nitrogen fixation. R. etli is characterized by a branched respiratory chain which involves at least four terminal cytochrome oxidases [26,27] and whose activity is regulated by the oxygen concentration in the environment. Nonzero fluxes through oxidative phosphorylation were observed in the simulation results, indicating oxidative phosphorylation is needed to maximize symbiotic nitrogen fixation. In fact, removal of all cytochrome oxidase reactions results in a complete loss of symbiotic nitrogen fixation. This result is in agreement with the reported essentiality of respiration for nitrogen fixation [5,24] (see Figure 3).
The model predicts incomplete use of the TCA cycle. Experiments have reported the activity of all enzymes in the TCA cycle under microaerobic conditions in some Rhizobium bacteroids [5]; however, all TCA cycle enzymes are not always detected experimentally [20]. In addition, mutants in TCA cycle enzymes in B. japonicum are still able to fix nitrogen, suggesting that a complete set of TCA cycle enzymes is not strictly required to fix nitrogen [5,28]. Here, FBA supports the notion that incomplete use of the TCA cycle can still result in nitrogen fixation at low oxygen uptake rates in R. etli. The model predicts that citrate synthase, isocitrate dehydrogenase, and 2-oxoglutarate dehydrogenase are not used, which agrees with the enzymes suggested by experimental results to be inactive under microaerobic conditions [23]. This result is supported by flux variability analysis across alternate optimal solutions [29], and it was insensitive to changes in the weightings in the OF coefficients of metabolites in Equation 1. The flux through these enzymes was always zero in all optimal solutions, independent of the coefficients defined in the OF.
The model also predicts that there is no flux through the Entner-Doudoroff and pentose phosphate pathways under microaerobic conditions. As above, no flux through Entner-Doudoroff and pentose phosphate pathways was observed across all optimal solutions for 1,000 randomly weighted OF. Proteomic analysis of the B. japonicum bacteroid did not detect the presence of enzymes involved in the Entner-Doudoroff pathway in agreement with the model predictions [20]. However, the proteome data shows the presence of several of the pentose phosphate pathway enzymes, which contradicts the in silico predictions. One possible explanation for this discrepancy is that the use of this pathway could be strain-dependent. Thus, experimental measurements in R. etli bacteroids are needed to support or refute this computational result.
It has been shown experimentally that the glutamine synthetase and glutamate synthase pathway (GS-GOGAT) constitutes the central mechanism of ammonium assimilation in free-living Rhizobia [30–32]. FBA predicts that there is no activity in the ammonium assimilation pathway during symbiosis (a non–free-living state), which agrees with recent reported measurements where ammonium assimilation was not observed during nitrogen fixation [20,32]. As before, flux variability analysis and random selection of the OF coefficients showed the inactivity of this metabolic pathway. The ammonium assimilatory pathways in the bacteroid are not active since most of the ammonium produced is transported to the plant. In fact, there is experimental evidence that an increase in ammonium assimilation negatively affects the nodulation process [32]. It also suggests that the inactivity of ammonium assimilation establishes optimal conditions for symbiotic nitrogen fixation, where complete transport of ammonium to the plant favors symbiotic nitrogen fixation.
Myo-inositol is an abundant compound in nodules and bacteroids, and it has an important influence on the efficiency of nitrogen fixation [33]. Although its origin is not well-understood, it constitutes an essential precursor for the synthesis of rhizopines. Rhizopines may function as osmotic protectants and in some Rhizobium strains confer a competitive advantage at early stages of nodulation [33,34]. The metabolic pathway of myo-inositol catabolism has been included in the metabolic reconstruction, and simulations with high uptake rates of myo-inositol suggest that it can be used as an energy source for nitrogen fixation. In silico analysis suggests that the activity of the myo-inositol dehydrogenase enzyme, encoded by idhA, increases symbiotic nitrogen fixation. Conversely, elimination of myo-inositol dehydrogenase enzyme from the network decreases the predicted nitrogen fixation activity. These model results have been observed experimentally in an idhA mutant in Sinorhizobium fredii [33]. The results obtained from simulations support the idea that myo-inositol catabolism plays an important role during nitrogen fixation, not only at early stages of nodulation but during nitrogen fixation as well. From a physiological point of view, there are some other possible explanations for reduced nitrogen fixation when the myo-inositol dehydrogenase enzyme is inactive. It may be a consequence of toxic levels of myo-inositol or a consequence of insufficient concentrations of myo-inositol for growth and maturation of the bacteroid [33]. Experimental data suggest that the essentiality of idhA is strain-dependent, and in this reconstruction we have included this metabolic pathway in agreement with the KEGG database [15]. From our in silico analysis, we observe that a deletion of myo-inositol dehydrogenase enzyme decreases nitrogen fixation activity, although it is not essential for symbiotic nitrogen fixation in R. etli. According to our in silico results, idhA is a key gene with potential implications for symbiotic nitrogen fixation.
In this section, we evaluate the capacity of the model to predict the physiological behavior of the bacteroid when it suffers gene deletions. In silico analysis was conducted for deletions in PHB synthase, glycogen synthase, arginine deiminase, myo-inositol dehydrogenase, and pyruvate carboxylase. The first three enzymes are involved in PHB and glycogen synthesis and in arginine degradation, respectively. We have selected these enzymes because experimental results are available for Rhizobium on how these gene deletions affect symbiotic nitrogen fixation with respect to the wild-type strain.
To simulate gene deletions, the fluxes through reactions associated with the corresponding enzyme are constrained to zero, and the OF for FBA is constructed as before, excluding PHB or glycogen for the PHB and glycogen synthase deletion simulations, respectively (see Dataset S2). Gene deletion studies for E. coli have shown that the method of minimization of metabolic adjustment (MOMA) makes better mutant predictions than FBA [35]. Unlike FBA, MOMA identifies the flux distribution for the mutant strain that is closest to the wild-type flux distribution (measured as the Euclidean distance) [35]. Predicted changes in fluxes through some key reactions in the mutant strains versus the wild-type are reported in Figure 3, where mutant predictions were done using both FBA and MOMA. Qualitatively, FBA and MOMA made similar predictions for most mutants with respect to the wild-type symbiotic nitrogen fixation flux.
For the double gene deletion (PHB and glycogen synthase), FBA and MOMA predict different results. FBA predicts an increase in symbiotic nitrogen fixation, while MOMA predicts a decrease in symbiotic nitrogen fixation; unfortunately, no experimental results are available for the double deletion mutant. For this case, we observed slight variations in exchange fluxes when using the two analysis methods. There was also a set of biochemical reactions that have no flux in the FBA solution for the double mutant but are used in the MOMA solution: aconitase (ACONT), 2 oxoglutare dehydrogenase (AKGDH), citrate synthase (CS), cystathionine b-lyase (CYSTL), cystathionine beta-synthase (CYSTS), 2-dehydro-3-deoxy-phosphogluconate aldolase (EDA), 6-phosphogluconate dehydratase (EDD), and fructose bisphosphate aldolase (FBA). Conversely, we identified three reactions, methylmalonate (MMSAD3), inositol dehydrogenase (INS2D), and inositol catabolic reaction (INSCR), whose fluxes were active in FBA, but inactive in the MOMA solution. The complete flux distributions for the double mutant predictions are included in Dataset S6.
Simulations for deletions of PHB synthase predict that symbiotic nitrogen fixation increases, in agreement with the experimental observations in R. etli [22]. Experimentally, deleting PHB synthase increases NADH levels and consequently reduces the NAD+ / NADH ratio. A possible physiological explanation for the increase in nitrogen fixation rate could be that the increase of reductive power (lower NAD/NADH ratio) can be channeled to nitrogen fixation [22]. A similar effect on symbiotic nitrogen fixation is predicted for the glycogen synthase deletion, which also agrees with the observed physiological response reported for Rhizobium tropici [36]. FBA suggests that the same effect would occur for R. etli. The simulations predict that the flux through PHB synthase increases in the glycogen synthase mutant, and, similarly, the flux through glycogen synthase increases in the PHB synthase mutant. These results are in agreement with experimental reports suggesting that the quantities of both polymers are relatively flexible such that inhibition of one results in an accumulation of the other [22], a property that is qualitatively observed in our in silico modeling.
It is unclear whether arginine is supplied to the bacteroid by the plant, and as a result the metabolic network reconstruction lacks an arginine transport reaction and ornithine degradation reactions. Recent evidence has found that the arginine deiminase pathway is active in R. etli bacteroids which convert arginine into ornithine while generating ATP and CO2. By allowing the ornithine carbamoyl transferase reaction to be reversible and including an arginine–ornithine antiporter in the network, the symbiotic nitrogen fixation flux increased by more than 40% compared with the condition where arginine is not metabolized by the bacteroid (using iOR363). Deletion simulations removing arginine deiminase from this modified network predict a decrease in symbiotic nitrogen fixation (Figure 3), which is in agreement with experimental mutations of the arcA gene, which encodes the arginine deiminase enzyme in R. etli [37].
Pyruvate carboxylase is not used in the wild-type solution identified using FBA, so mutant predictions for pyruvate carboxylase are identical to the wild-type predictions. It is known experimentally for R. etli that deletion of this enzyme does not affect nitrogen fixation [38]; however, it does not discard the possibility that this enzyme may be important in other life stages or in rhizosphere competition [39,40].
To verify that these conclusions obtained from the gene deletion analysis are not sensitive to the coefficients in the OF, we randomly assigned OF coefficients for the five metabolites from a uniform distribution of values between 0 and 1. A total of 1,000 different random OF were subsequently used individually in the gene deletion analysis. The qualitative effect of the gene deletions on symbiotic nitrogen fixation was independent of the OFs used; for example, in all 1,000 cases the deletion of PHB synthase increased symbiotic nitrogen fixation. However, quantitative prediction of gene deletion effects was dependent on the chosen OF coefficients. To improve quantitative model predictions, a better approximation of the physiological levels of the OF components is needed.
Phenotypic phase plane (PhPP) analysis is a useful method to characterize the steady-state solution space projected in two dimensions [41,42]. Through this analysis, the steady-state flux distributions can be divided into a finite number of regions, each with similar metabolic flux patterns and characterized by equivalent shadow prices [41]. The shadow prices give us information on how the OF would change if the metabolites were additionally supplied to the network. These shadow prices can be used to classify the phenotypic phase plane into regions where the availability of different metabolites limits symbiotic nitrogen fixation.
The metabolic network for R. etli was analyzed with respect to two parameters, succinate and oxygen uptake rates. The low oxygen uptake rate inside the nodules is such that it creates a microaerobic environment optimal for nitrogen fixation [43,44]. Constraining our analysis to low uptake rates of oxygen (1 mmol/gDW/hr ), we found three regions in the PhPP, each one characterized by a qualitatively different optimal use of metabolic pathways [41] (see Figure 4). As shown in Figure 4, it is possible to identify a finite number of regions, each one characterized by the effect that oxygen and succinate uptake rates have on the OF. Region I is characterized by a single limiting substrate, succinate, which establishes an independent relationship between the OF and the oxygen uptake rate for a fixed succinate uptake rate. Region II has dual substrate limitations. In this region an increase in the uptake rate of either succinate or oxygen will increase symbiotic nitrogen fixation. Finally, in region III, an increase in succinate uptake rate produces a decrease in the symbiotic nitrogen fixation, while an increase in oxygen uptake rate increases the OF in region III. Region III is defined as a futile region [41]. The line of optimality represents the optimal relation between the substrate uptake rates and the OF. In this case, it lies on the boundary between regions I and II. Points on the line of optimality represent the optimal oxygen uptake required for the oxidation of succinate to maximize symbiotic nitrogen fixation [45].
In silico analysis shows that an increase in the succinate uptake rate, at a fixed oxygen uptake rate (1mmol/gDW/hr), increases symbiotic nitrogen fixation until a threshold in succinate uptake rate is reached (see Figure 4). At succinate uptake rates higher than this value, an inhibitory effect on symbiotic nitrogen fixation is observed. Here, the available oxygen is not enough to oxidize the excess succinate and it reduces symbiotic nitrogen fixation. This indicates that oxygen is a limiting compound in nitrogen fixation.
The contributions of this work are the metabolic reconstruction of R. etli (the first reconstruction made for Rhizobia) and the FBA of the resulting model during nitrogen fixation stages. Even with a lack of detailed experimental information, like kinetics constants and flux limitations for most reactions, we have been able to show that constraint-based methods can describe the capabilities of the metabolic network consistent with available experimental information. The construction of an OF that mimics symbiotic nitrogen fixation and qualitatively agrees with bacteroid physiology constitutes a substantial contribution in this work. Our OF was sufficient to obtain some of the main qualitative physiological characteristics for R. etli during nitrogen fixation. Thus, our model consistently reproduces, in agreement with the literature, the utilization of pathways like oxidative phosphorylation, gluconeogenesis, and PHB biosynthesis during nitrogen fixation. Additionally, the analyzed gene deletion set was in qualitative agreement with the response of the experimentally observed symbiotic nitrogen fixation activity (see Figure 3). These particular gene deletions were selected because their effects on nitrogen fixation activity have been measured experimentally, allowing us to evaluate our computational results. This reconstruction can be used to suggest gene deletions that could enhance symbiotic nitrogen fixation. Here we show one example where a double gene deletion in PHB synthase and glycogen synthase could potentially increase symbiotic nitrogen fixation.
The advantages of this in silico framework have been shown in other organisms [10,12,45], and we expect that for R. etli such computational analysis will be useful to design and improve nitrogen fixation in plant development and agriculture. This last point constitutes a valuable scientific objective which requires integration of experimental data to improve and update the metabolic reconstruction. For instance, aspartate aminotransferase in B. japonicus is essential, and its deletion is detrimental to nitrogen fixation [17], while a deletion of glutamine synthetase increases nitrogen fixation [17]. We do not observe this behavior in silico for R. etli. Experimental evaluation of these mutants in R. etli will provide a means to validate and improve the model.
Similarly, recent studies have shown the ability of Rhizobia to produce amino acids in well-defined environments and the effects that these amino acids have on symbiotic nitrogen fixation [16,17,46]. Although the production of the complete spectrum of amino acids has not been well-characterized in R. etli during nitrogen fixation, we have predicted their production with the model, and compared them with the amino acids experimentally observed in B. japonicum bacteroids [17]. FBA predicts the production of amino acids included in the OF (aspartate, alanine, and lysine). Experimental evidence shows the production of these amino acids in B. japonicum bacteroid [17]. Conversely, the experimental data for the B. japonicum bacteroid disagrees with the in silico analysis for R. etli, where asparagine, methionine, leucine, isoleucine, glycine, glutamine, and serine are not predicted by the model to be produced. Measurements of amino acid production in R. etli bacteroids are needed so a more accurate OF can be constructed [46]. Taken together, the reconstruction and analysis presented here provides an initial template for studying symbiotic nitrogen-fixing bacteria, and it can be used to generate hypotheses, design experiments, and to test predictable control principles for the metabolic network of R. etli.
Metabolomic flux prediction through the R. etli metabolic network was done using FBA [12]. We assume that all the chemical compound concentrations and fluxes are at steady state. To constrain the space of all the possible steady-state flux distributions, we impose stoichiometric mass balance constraints, thermodynamic constraints pertaining to reaction reversibility, and some enzyme capacity flux constraints [12]. Optimization of the OF was solved using the SimPheny software (Genomatica). MOMA calculations were carried out as described previously [35] using GAMS with only the metabolic and transport fluxes used in calculating the Euclidean distance; the exchange fluxes and OF values were omitted from the distance metric (see Datasets S1–S6). For the arginine deiminase mutant predictions, the following network modifications were made: ornithine carbamoyl transferase became reversible, an arginine–ornithine antiport reaction was included, and exchange fluxes for arginine and ornithine were added (a maximum arginine uptake rate of 5 mmol / gDW / hr was used).
The bacteroid is an open thermodynamic system that exchanges components with the plant through its peribacteroid membrane (see Figure 1). In our in silico analysis, we have classified chemical reactions into three categories: internal, exchange, and sink. The first contains most of the reactions in the metabolic reconstruction (all metabolic and transport reactions occurring inside the bacteroid). The second includes reactions which represent an interchange with the plant host (i.e., they allow metabolites to cross the system boundary). Finally, the third classification includes the entry or exit of metabolites by an unidentified source into the bacteroid. The only sink included is for myo-inositol, since the compound is observed in bacteroids, but it is unknown whether it is supplied by the plant or synthesized inside the bacteroid. In the case of the myo-inositol sink, we have limited the flux through the reaction so that myo-inositol present in the bacteroid can only be consumed. Most of the exchange reactions were defined as freely taken up or secreted, with the exception of arginine and oxygen. The oxygen uptake rate was limited by reported experimental measurements made in bacteroids [25]. The complete set of reactions and their flux constraints are available in Dataset S2 and Table S2.
In silico analysis of gene deletions for arginine deiminase, myo-inositol dehydrogenase, cytochrome oxidase, and pyruvate carboxylase was simulated by removal of the corresponding enzymatic reactions. In all these cases, the OF used was the same as the one presented in Equation 1. However, for glycogen synthase, PHB synthase, and the double mutant (PHB + glycogen synthase), new reduced OFs were used, where the corresponding component(s) (glycogen, PHB, or both) were omitted. For example, in the PHB synthase mutant the new reduced OF was:
The optimization problems (FBA and MOMA) were then solved with these new reduced OFs. The OF for each gene deletion is presented in Dataset S2.
To verify that the inactive metabolic pathways identified in this in silico analysis are not dependent on the stoichiometric coefficients defined in the OF, we performed flux variability analysis [29] combined with random assignment on the coefficient of the OF. We verified the robustness of our reported results for some enzymes in the following metabolic pathways: TCA cycle (aconitase A and B, isocitrate dehydrogenase, 2-oxoglutarate dehydrogenase, citrate synthase), Entner-Doudoroff Pathway (2-dehydro-3-deoxy-phosphogluconate aldolase, 6-phosphogluconate dehydratase), pentose phosphate pathway (glucose-6-phosphate dehydrogenase, 6-phosphogluconolactonase, phosphogluconate dehydrogenase, and ribulose 5-phosphate 3-epimerase), and ammonium assimilation (glutamate synthase, glutamine synthetase, and glutaminase). We assigned random numbers (from a uniform distribution between 0 and 1) for the stoichiometric coefficients defining the OF. After generating a randomly weighted OF, flux variability analysis was used to calculate the maximum and minimum flux values for each enzyme across all alternate optimal solution which maximizes this random OF. This procedure was repeated 1,000 times. This analysis was done using Matlab and LINDO. |
10.1371/journal.ppat.1003562 | Francisella tularensis Harvests Nutrients Derived via ATG5-Independent Autophagy to Support Intracellular Growth | Francisella tularensis is a highly virulent intracellular pathogen that invades and replicates within numerous host cell types including macrophages, hepatocytes and pneumocytes. By 24 hours post invasion, F. tularensis replicates up to 1000-fold in the cytoplasm of infected cells. To achieve such rapid intracellular proliferation, F. tularensis must scavenge large quantities of essential carbon and energy sources from the host cell while evading anti-microbial immune responses. We found that macroautophagy, a eukaryotic cell process that primarily degrades host cell proteins and organelles as well as intracellular pathogens, was induced in F. tularensis infected cells. F. tularensis not only survived macroautophagy, but optimal intracellular bacterial growth was found to require macroautophagy. Intracellular growth upon macroautophagy inhibition was rescued by supplying excess nonessential amino acids or pyruvate, demonstrating that autophagy derived nutrients provide carbon and energy sources that support F. tularensis proliferation. Furthermore, F. tularensis did not require canonical, ATG5-dependent autophagy pathway induction but instead induced an ATG5-independent autophagy pathway. ATG5-independent autophagy induction caused the degradation of cellular constituents resulting in the release of nutrients that the bacteria harvested to support bacterial replication. Canonical macroautophagy limits the growth of several different bacterial species. However, our data demonstrate that ATG5-independent macroautophagy may be beneficial to some cytoplasmic bacteria by supplying nutrients to support bacterial growth.
| Francisella tularensis is a highly virulent bacterial pathogen that infects hundreds of different animal species including humans. During infection, F. tularensis bacteria invade and rapidly multiply inside host cells. Within the host cell environment, basic nutrients that bacteria require for growth are in limited supply, and the majority of nutrients are tied up in complex molecules that are not readily available in forms that can be used by bacteria. In this study we asked and answered a very simple question; how does F. tularensis harvest sufficient carbon and energy sources from the host cell to support rapid intracellular growth? We found that F. tularensis induces a host recycling pathway in infected cells. Thus the host cell degrades nonessential proteins and releases amino acids. F. tularensis harvests the host-derived amino acids to generate energy and build its own more complex molecules. When we inhibited the host recycling pathway, growth of the intracellular bacteria was limited. Therefore, manipulation of host cell metabolism may be a means by which we can control the growth of intracellular bacterial pathogens during infection.
| When intracellular bacterial pathogens invade host cells, the bacteria must scavenge energy sources and anabolic substrates from the nutrient-limited intracellular environment. Most of the potential nutrient sources inside a host cell are stored within complex structures such as lipid droplets, glycogen and proteins, which are not immediately available to intracellular pathogens. To obtain nutrients for proliferation, intracellular bacteria must degrade these complex structures into their constituents (fatty acids, carbohydrates and amino acids respectively) or increase nutrient import. The strategies that bacteria use to acquire nutrients could potentially have widespread effects on the host cell. For example, pathogens that import amino acids from the host cell cytoplasm may starve the cell. Host cell amino acid starvation leads to mammalian target of rapamycin (mTOR) inhibition, thereby inhibiting mRNA transcription and other critical cellular homeostatic processes [1].Thus, nutrient acquisition is an important step in the pathogenesis of intracellular bacteria and is critical to understand how a pathogen interacts with the host.
Autophagy is a highly conserved eukaryotic cell process that can be initiated by a variety of factors such as amino acid starvation, energy depletion, mTOR inhibition and immune signaling [2], [3]. Autophagy is a process by which multi-membranous vesicles called autophagosomes surround and degrade cellular constituents (during starvation) or cytoplasmic bacteria (during infection through a related innate immune response termed xenophagy [4]). The autophagosomes fuse with lysosomes to become autolysosomes, which then degrade the engulfed material. During starvation, autophagy can degrade nonessential proteins, thereby releasing free amino acids that are recycled into new proteins. Current studies of the interactions between host autophagy and intracellular bacterial pathogens are primarily focused on xenophagy [5]–[7]. However, a few intracellular pathogens are known to benefit from autophagy [8]–[10]. Autophagosome formation is induced during infection with Anaplasma phagocytophilum and the autophagy derived nutrients are harvested and used by A. phagocytophilum to enhance intracellular replication [9]. Likewise, dengue virus uses autophagic byproducts to acquire lipids for viral replication [10]. Pathogens such as Listeria monocytogenes express active mechanisms that prevent bacterial degradation via xenophagy, yet autophagy still occurs in the infected cell and has the potential to provide nutrient sources for the bacteria [11]. These and other recent studies highlight the potential role of autophagy in providing nutrients or other benefits for intracellular pathogens.
Francisella tularensis is a facultative intracellular bacterium that infects over 200 different species (from amoeba to humans) [12]. The highly virulent F. tularensis subsp. tularensis Schu S4 strain has an infectious dose of fewer than 25 bacteria and a mortality rate of 30–60% in untreated pneumonic infections [13], [14]. F. tularensis infects a diverse range of cell types including macrophages, which are a key replicative niche for F. tularensis in humans and other susceptible mammals. F. tularensis also invades and replicates within several other cell types including epithelial cells and endothelial cells [12], [15]. F. tularensis enters the host cell through phagocytosis and proceeds to escape the phagosome and replicate in the host cell cytoplasm. By 24 hours post inoculation, F. tularensis replicates up to 1000-fold inside host cells. This rapid intracellular replication plays a major role in F. tularensis pathogenesis but the mechanisms by which this organism acquires nutrients are not well characterized. Therefore, we sought to determine how these nutrients become available to support efficient F. tularensis intracellular replication.
In primary murine macrophages, F. tularensis induces the formation of a multi-membranous, autophagosome-like structure termed the Francisella containing vacuole (FCV) through an autophagy related process [16]. FCV formation occurs between 20 and 36 hours post inoculation, after the majority of F. tularensis replication has taken place. Blocking FCV formation late during infection does not increase F. tularensis proliferation, suggesting that FCV formation does not play a role in controlling intracellular F. tularensis replication [16]. However, the formation of FCVs hints that autophagy may be induced during F. tularensis infection. Additionally, replication deficient and chloramphenicol treated F. tularensis bacteria, but not wild type F. tularensis bacteria, are degraded via canonical autophagy [17]. This observation implies that F. tularensis avoids xenophagy. Lastly, treating F. tularensis infected macrophages 2 hours post inoculation with chloroquine or autophagy- inhibiting levels of ammonium chloride impairs F. tularensis intracellular replication [18]–[20]. Although chloroquine and ammonium chloride inhibit acidification of cellular compartments and have broad effects on the host cell, these data raise the intriguing possibility that autophagy may contribute to F. tularensis intracellular replication.
Taken together these observations suggest that intracellular F. tularensis avoids xenophagy yet induces autophagy or an autophagy-like process that contributes to F. tularensis proliferation. We therefore examined the potential role of autophagy in aiding F. tularensis intracellular growth.
F. tularensis replicates efficiently and rapidly in host cells. Indeed, transmission electron microscopy analysis showed that F. tularensis consumed over half of the area of the cell cytoplasm of infected mouse embryonic fibroblasts (MEFs) by 16 hours post inoculation (Figure S1). F. tularensis cannot make all of the nutrients it needs de novo and must interact with the host to acquire certain metabolites to support rapid proliferation. In particular, F. tularensis is auxotrophic for 13 amino acids, some of which mammalian cells also do not synthesize. Thus, for sustained proliferation within infected cells, the bacteria must either take up amino acids imported by the host cell or degrade host proteins and reuse the resulting amino acids. To distinguish between these possibilities, we determined if decreasing the availability of free amino acids limited F. tularensis intracellular growth. We replaced the media on infected MEFs with media lacking amino acids at 3 hours post inoculation. F. tularensis replicated to similar numbers with or without amino acids present in the tissue culture media (Figure 1A). This result demonstrates that F. tularensis can acquire the amino acids it needs to sustain growth directly from the host cell. Since the majority of host amino acids are typically sequestered in proteins inside the cell, protein degradation likely occurs to provide sufficient amino acids to support F. tularensis intracellular growth. Additionally, amino acid depletion results in starvation induced autophagy [21]. Starvation induced autophagy will degrade proteins to produce amino acids. Thus, F. tularensis may take advantage of host cell autophagy to acquire free amino acids.
To determine if autophagy had any impact on F. tularensis intracellular growth we measured bacterial replication inside cells treated with several different autophagy inhibitors. MEFs were treated with 3-methyladenine (3MA), which inhibits autophagosome formation, thereby blocking autophagy. F. tularensis replication inside 3MA treated MEFs was significantly reduced (Figure 1B), suggesting that intracellular F. tularensis benefit from host cell autophagy. Since autophagy is both a starvation response and a process by which damaged organelles and non-essential proteins are degraded we considered the possibility that F. tularensis may scavenge and utilize amino acids released by this process. We therefore wanted to determine if exogenous amino acid supplementation would rescue F. tularensis growth in MEFs that have impaired autophagy function. Indeed, F. tularensis intracellular growth in the presence of 3MA was restored by the addition of excess amino acids to the culture media (Figure 1B). These results, which were corroborated using confocal fluorescence microscopy of cells infected with GFP-expressing F. tularensis Schu S4 (Figure 1C) indicate that autophagy provides a source of nutrients that support F. tularensis replication.
To determine if degradative autophagy was responsible for optimal bacterial growth, we quantified F. tularensis intracellular growth in the presence of Bafimoycin A(1) (Baf) or chloroquine (CQ), each of which inhibits autophagy by blocking functional autolysosome formation. We tested the effect of these drugs on F. tularensis replication kinetics by infecting MEFs with F. tularensis containing a bioluminescence reporter plasmid (Schu S4-LUX) [22] and measuring luminescence every 30 minutes to determine the bacterial growth kinetics. The limit of detection for this assay was approximately 50 relative light units (RLUs) or approximately 105 bacteria in a 96 well format (data not shown). We verified this technique by treating F. tularensis infected cells with 3MA or 3MA supplemented with amino acids and observed similar results to the standard intracellular proliferation assays (Figure S2A, S2B). Additionally, CQ significantly reduced F. tularensis growth and amino acid supplementation rescued bacterial growth in CQ treated cells (Figure S2C, S2D). Similar to 3MA and CQ, treating MEFs with Baf also significantly reduced F. tularensis intracellular growth and growth was rescued with amino acid supplementation (Figure S2E, S2F). None of the inhibitors affected F. tularensis growth in broth culture (Figure S3B). Although 3MA, CQ, and Baf were each cytotoxic to MEFs, viability was comparable between treatments with and without amino acid supplementation (Figure S3A). Thus, the observed rescue was not due to increased eukaryotic cell viability upon amino acid supplementation.
Since all chemical inhibitors have the potential to confer off-target or non-specific effects on host cell processes we wanted to confirm the inhibitor results using genetic approaches. Beclin-1 is required for autophagosome formation in most autophagy pathways [23]. We therefore reasoned that depletion of Beclin-1 should limit bacterial replication if autophagy is required for F. tularensis growth. We created two Beclin-1 knock down MEF cell lines, Beclin-1 KD-1 and KD-2 that expressed 63.8% (+/−14.4%) and 59.2% (+/−12.9%) of the scrambled shRNA control Beclin-1 mRNA, respectively (Figure S4). Despite the relatively modest reduction of Beclin-1 mRNA F. tularensis replication was significantly reduced in the knockdown cell lines compared to the scrambled control (Figure 1D); supporting the conclusion that autophagy may have a pro-bacterial role in F. tularensis infected cells. Interestingly, the infection frequency of the knock down cell lines was approximately 2-fold higher than the scrambled control (data not shown) suggesting that Beclin-1 activity may modestly impair F. tularensis infection of host cells.
During the course of infection F. tularensis invade and replicate within many different cell lineages and types. Intracellular growth properties of F. tularensis vary depending on host cell type. For example, F. tularensis infects monocytes at a significantly higher frequency than epithelial cells or fibroblasts. On the other hand, F. tularensis intracellular growth is more prolonged, and achieves nearly 10-fold higher peak numbers in MEFs as compared to monocytes (data not shown). Growth within monocytes is a property that is fundamental to F. tularensis virulence. F. tularensis is also a human pathogen; we therefore wanted to determine the relevance of autophagy in supporting F. tularensis growth within human macrophages. Inhibition of autophagy with 3MA significantly decreased F. tularensis growth in hMDMs, and growth was rescued in 3MA treated hMDMs by supplementing the media with excess amino acids (Figure 1E). Therefore, autophagy provides amino acids that support F. tularensis intracellular growth in primary human monocytes, a property that is crucial to F. tularensis pathogenesis.
We compared the rate of degradation of long-lived proteins in uninfected and infected cells to determine if F. tularensis infection impacted autophagic flux. Since we were attempting to quantify a specific infected host cell response we performed this analysis in the J774A.1 monocyte cell line (J774) where the F. tularensis infection frequency is much greater than the infection frequency in MEFs (data not shown). We first labeled cellular proteins by incubating J774 cells in media containing 35S methionine and cysteine for 18 hours and chased for 2 hours to remove any remaining labeled free amino acids. The labeled cells were inoculated with F. tularensis and incubated for 16 hours. Following infection, infected cells had a 49.5%+/−7.9% (Average +/− SEM) decrease of 35S label in the TCA insoluble fraction of the cytoplasm (which will primarily contain proteins) compared to uninfected J774 cells (Figure 2A). Thus, infected cells had increased turnover of long lived proteins than uninfected cells. This result is consistent with autophagy induction in F. tularensis infected J774 cells. The decrease of total 35S label in both host and bacterial proteins in infected cells may indicate that the transfer of amino acids from the host to the bacteria is inefficient or that the majority of amino acids are used by F. tularensis for energy rather than protein synthesis. Uninfected and infected J774 cells had similar levels of cytotoxicity at 16 hours post inoculation, indicating that the loss of label in infected compared to uninfected cells was not due to cell lysis (Figure S3D).
To confirm that F. tularensis imports amino acids derived from host proteins, we monitored transfer of radiolabelled amino acids from host proteins into bacterial proteins. MEFs were first metabolically labeled with 35S-labeled methionine and cysteine for 18 hours to fully label all host proteins. Then the radiolabel was removed and the cells were incubated in unlabeled media for two hours prior to infection with F. tularensis to remove 35S that was not incorporated into protein. At 16 hours post infection (18 hours after the radiolabel was removed) we lysed the MEFs and purified F. tularensis by mixing cell lysate from either uninfected or infected cells with magnetic beads linked to an anti- F. tularensis lipopolysaccharide (LPS) antibody. We then determined if F. tularensis proteins contained radiolabeled amino acids by examining the trichloroacetic acid (TCA) insoluble fraction of purified F. tularensis. There was a significant increase of radiolabel in the TCA insoluble, F. tularensis bead purified fraction from infected MEFs as compared to uninfected control samples (Figure 2B). Indeed, 6.22%+/−4.15% (average +/− SEM, n = 5 samples) of the TCA insoluble radiolabel present prior to infection transferred to the bacteria during the 16 hour infection. To control for possible direct transfer of labeled amino acids that were not incorporated into host proteins we analyzed infected MEFs that were treated with cycloheximide during 35S labeling prior to infection. There were negligible amounts of radiolabel present in the bead purified fraction of cycloheximide treated cells (Figure 2B). F. tularensis survived and replicated within cycloheximide pre-treated cells and F. tularensis was present in the bacterial purified fraction (data not shown). Thus, host cell lysis due to the cycloheximide treatment was not solely responsible for the lack of radiolabel in the bacterial fraction. 35S radiolabel was primarily incorporated into host proteins, rather than as free 35S labeled amino acids. Taken together, these data demonstrate that F. tularensis synthesized proteins using amino acids derived from host cell proteins.
Treating the radiolabeled cells with either Baf or 3-MA resulted in significantly decreased incorporation of the radiolabel by F. tularensis (Figure 2C). Since F. tularensis proliferation is reduced in 3MA and Baf treated MEFs, several fold fewer bacteria were present in the bacteria purified fraction of the treated MEFs (data not shown). Nevertheless, the median 35S counts per bacteria were significantly lower in the 3MA or Baf treated samples compared to untreated samples (untreated: 0.016 CPM/bacteria, 3MA: 0.000 CPM/bacteria, Baf: 0.000 CPM/bacteria [n = 3 experiments done in duplicate]). Therefore, transfer of radiolabeled amino acids to bacterial proteins was reduced by both 3MA and Baf treatment, indicating that under normal culture conditions, amino acids derived by the degradation of host cell proteins via autophagy were used by F. tularensis.
F. tularensis is capable of using amino acids as an energy source when simple carbohydrates such as glucose are not available (Figure 3A). Thus, autophagy derived amino acids could conceivably be used by intracellular F. tularensis for either the synthesis of new proteins or to provide energy for other bacterial processes. Although we found that F. tularensis uses host-derived amino acids for protein synthesis (Figure 2B), the proportion of amino acids used for protein synthesis as opposed to energy is unknown. To determine if F. tularensis uses autophagy-derived amino acids primarily as anabolic precursors or as an energy source, we supplemented autophagy inhibited, F. tularensis infected MEFs with either serine or the metabolite pyruvate. Annotation of the F. tularensis genome indicates that F. tularensis encodes the protein L-serine dehydratase, which degrades serine directly into pyruvate. The addition of either pyruvate or serine alone rescued F. tularensis intracellular growth in Baf treated cells (Figure 3B). Fibroblasts cannot convert serine or pyruvate into all of 13 of the amino acids required to fulfill F. tularensis auxotrophies. Thus, host autophagy-derived nutrients are used by F. tularensis primarily as a source of energy. Although F. tularensis can incorporate autophagy derived amino acids into bacterial proteins (Figure 2B), these data indicate that energy, rather than amino acids for protein synthesis, was the limiting factor for F. tularensis proliferation in autophagy-deficient cells cultured in tissue culture media.
Canonical autophagy is typically induced by the inhibition of mammalian target of rapamycin (mTOR). Thus, monitoring mTOR activity through downstream substrates such as S6 kinase is likely to correlate well with canonical autophagy induction. To determine if F. tularensis infection activates the autophagy signaling cascade, we assessed mTOR activity in infected J774 cells by measuring phosphorylation of the mTOR substrate S6 ribosomal protein. The ratio of phospho- S6 ribosomal protein to unphosporylated S6 ribosomal protein decreased progressively over the course of infection, which is consistent with mTOR inhibition and thus autophagy induction (Figure 4A, 4B) [24]. However, loss of phospho - S6 ribosomal protein was not evident before 8 hours post inoculation suggesting that mTOR inhibition occurred after some bacterial replication had already taken place.
In the canonical autophagy pathway the protein ATG5 is essential for autophagosome formation. Thus, we would predict that ATG5 expression would be required for autophagic degradation of host proteins to amino acids that support F. tularensis intracellular growth. However, it was recently shown that F. tularensis replicates efficiently within ATG5−/− macrophages [17]. We also found that F. tularensis replication was not impaired in ATG5−/− MEFs (Figure 5A). In fact, there was a slight but statistically significant increase in bacterial replication in ATG5−/− MEFs compared to wild type MEFs (Figure 5A). Therefore, ATG5 is not required for efficient F. tularensis intracellular proliferation. Treatment of ATG5−/− MEFs with 3MA resulted in decreased bacterial proliferation and bacterial growth was rescued by supplementing treated cells with amino acids (Figure 5B). Taken together, these data suggest that F. tularensis intracellular growth is supported by nutrients generated by an ATG5-independent autophagy pathway.
Unlike canonical autophagy, ATG5-independent autophagy generates autophagosomes from the trans-Golgi apparatus [25]. Brefeldin A (Bref A) inhibits ATG5-independent autophagosome formation but does not affect canonical autophagosome formation [24]. To determine if ATG5-independent autophagy provides metabolites for F. tularensis in macrophages, we measured F. tularensis replication in J774 cells in the presence and absence of Bref A. Cells were infected with Schu S4-LUX and growth was monitored by measuring luminescence every 30 minutes. We found that F. tularensis replication was significantly reduced in Bref A-treated J774 cells (Figure 5C, 5D), and growth was significantly rescued in Bref A treated cells by the addition of amino acids (Figure 5C, 5D). Bref A cytotoxicity was comparable regardless of amino acid supplementation, indicating that the increase in bacterial replication was not due to decreased eukaryotic cell cytotoxicity in amino acid treated cells (Figure S3C). The ability of amino acids to rescue bacterial replication in Bref A-treated cultures indicates that Bref A affects F. tularensis nutrient availability. This result is consistent with the conclusion that ATG5-independent autophagy provides nutrients that support F. tularensis growth in macrophages as well as in MEFs.
We wanted to determine the extent to which autophagosomes are formed during F. tularensis infection, and the spatial relationship between the bacteria and autolysosomes in ATG5−/− cells. Analysis of transmission electron microscopy (TEM) micrographs revealed that autophagic vacuoles constituted a greater percentage of the cytoplasm in F. tularensis infected as compared to uninfected ATG5−/− MEFs (Figure 6A–D) confirming that autophagy is induced in ATG5−/− MEFs.
Since morphological analysis of autophagic structures by TEM is inexact, we used fluorescence confocal microscopy as a secondary means to identify acidified autophagic vacuoles in infected MEFs. We stained and quantified the number of LysoTracker Red positive acidic vacuoles in infected and uninfected ATG5−/− MEFs. There were significantly more acidic vacuoles in the infected ATG5−/− MEFs as compared to uninfected ATG5−/− MEFs (Figure 6E). LysoTracker Red can also stain other acidic vacuoles including lysosomes and phagosomes. However, the increased number of acidic vacuoles found in infected wild type and ATG5−/− MEFs as compared to uninfected and 3MA treated infected control cells strongly argues that the increase in acidic vacuoles correlate with an increase in autophagic vacuoles. Combined with the morphological analysis of the infected-cell vacuoles by TEM this data demonstrates that F. tularensis induced ATG5-independent autophagy in infected cells.
The slight but statistically significant increase in F. tularensis growth observed in ATG5−/− MEFs suggested that canonical autophagy may be induced in infected cells and exert some control over bacterial growth. It is also possible that in addition to destroying the bacteria, canonical autophagy could serve as a redundant mechanism for nutrient acquisition. To determine if canonical autophagy was induced in addition to ATG5-independent autophagy during infection with F. tularensis, we analyzed infected MEFs that were transiently transfected with a GFP-LC3 plasmid for an increase in GFP-LC3 puncta. LC3 puncta formation is stimulated by canonical autophagy; however, ATG5-independent autophagy does not induce LC3 puncta formation [24], [26]. LC3 puncta levels were unchanged in infected compared to uninfected MEFs at 16 hours post inoculation, whereas both the amino acid starvation and Torin1 controls conferred an increase in LC3 puncta (Figure 7A, B). Thus, it appears that canonical autophagy remained at basal levels in F. tularensis infected cells during late stages of infection.
To determine if induction of canonical autophagy would either increase bacterial clearance or generate additional nutrients that support bacterial replication, we artificially induced autophagy throughout infection with the mTOR inhibitor Torin1. Torin1 treatment throughout infection had no impact on F. tularensis intracellular survival or growth in MEFs (Figure 7C). Thus, F. tularensis evades destruction by canonical autophagy and increased canonical autophagy did not benefit F. tularensis intracellular replication.
F. tularensis induces ATG5-independent autophagy while canonical autophagy remains at basal levels during infection. Little is known about the functional differences between canonical and ATG5-independent autophagy. However, xenophagy is known to occur via canonical autophagy whereas xenophagy via ATG5-independent autophagy has not been addressed. In canonical autophagy, cytosolic pathogens including chloramphenicol treated F. tularensis are targeted for xenophagy when bound to p62/SQSTM1 and polyubiquitin [17], [27]–[29]. We therefore investigated the role of polyubiquitin and p62/SQSTM1 in ATG5-independent autophagy induction in F. tularensis infected cells.
There was a significant decrease in the number of polyubiquitin puncta in the cytoplasm of infected wild type and ATG5−/− MEFs as compared to uninfected MEFs (Figure 8A). If polyubiquitin was degraded upon ATG5-independent autophagy induction, we would expect a corresponding increase in co-localization between polyubiquitin and acidic vacuoles in infected cells. However, the number of acidic vacuoles co-localizing with polyubiquitin in uninfected cells (15.2%+/−2.2%) and infected cells (20.0%+/−3.5%) was not significantly different (n>25 cells, mean +/− SEM) (Figure 8B). These data indicate that the decrease in polyubiquitin aggregates in infected cells was independent of autophagy.
In addition, there were similar numbers of p62/SQSTM1 puncta in infected MEFs compared to uninfected MEFs (Figure 8C, S5C–S5E). Interestingly, although there were similar total numbers of p62/SQSTM1 puncta, there was increased co-localization of p62/SQSTM1 with acidic vacuoles in infected wild type MEFs. However, there was no difference in p62/SQSTM1 co-localization between uninfected and infected ATG5−/− MEFs (Figure 8D). The increased co-localization of p62/SQSTM1 with acidic vacuoles may indicate that some basal level of xenophagy is occurring in an ATG5-dependent manner, which is consistent with the increase in bacterial replication that we observed in ATG5−/− MEFs. Taken together, these data indicate that F. tularensis induced ATG5-independent autophagy is not associated with polyubiquitin, LC3B, or p62/SQSTM1.
A recent study demonstrated that Salmonella enterica associates with ubiquitinated aggregates that are degraded by autophagy [30]. Although these aggregates likely target S. enterica for degradation rather than supplying nutrients, these data suggest that mechanisms exist which target autophagosomes to bacteria or vice versa. We hypothesized that F. tularensis may recruit autophagic vacuoles, resulting in bacteria localizing in close proximity to autophagosomes to facilitate bacterial nutrient acquisition. Indeed, F. tularensis was frequently found within 250 nm of autophagic vacuoles in both ATG5−/− MEFs and J774 cells as determined by TEM (Figure S6A, S6B). Indeed, 25.8+/−4.0% (average +/− SEM) of the autophagic vacuoles in ATG5−/− MEFs were also within 250 nm of a bacterium.
We confirmed the TEM results using confocal microscopy. Since ATG5-independent autophagy does not appear to require ubiquitination or any known target marker, we were limited to examining the relationship between bacteria and acidified vacuoles. Infected cells were stained with LysoTracker Red and Z-stacks from infected cells were analyzed by confocal microscopy. 28.0%+/−3.7% of bacteria in wild type MEFs and 35.1%+/−5.1% of bacteria in ATG5−/− MEFs were within 250 nm of an acidic vacuole (Average +/− SEM, n>10 cells) (Figure S6 C–H). At least 1 bacterium was within 250 nm of an acidic vacuole in every cell. The number of bacteria within 250 nm of an acidic vacuole was significantly lower in 3MA treated MEFs compared to the untreated MEFs (p = .01) (Figure S3 H). These data suggest that F. tularensis may recruit or traffic to autophagic vacuoles. Further investigation may reveal that not only autophagy induction, but also proximity to an autophagic vacuole contributes to F. tularensis nutrient acquisition.
Intracellular pathogens have evolved to thrive within the hostile nutrient-limited host cell environment. Successful pathogens disarm or avoid innate and adaptive immune responses while simultaneously extracting carbon and energy sources to support their proliferation. Autophagy is a highly conserved degradation process that serves a multitude of functions including cell development, stress response and resistance to cytoplasmic pathogens. Herein we investigated the interaction between F. tularensis and the host cell autophagy response. Our results demonstrate that ATG5-independent autophagy is triggered in F. tularensis infected cells and that intracellular bacterial replication was enhanced by this process. Furthermore, F. tularensis can replicate in cells when there are no amino acids present in the media, indicating that F. tularensis obtains all of the amino acids necessary to fulfill its 13 amino acid auxotrophies from the host cell through processes such as autophagy. F. tularensis acquires amino acids, and possibly other nutrients, via autophagy. These nutrients are then used for both energy and protein synthesis, although decreased bacterial replication in ATG5-independent autophagy deficient cells is primarily due to a lack of available energy. Autophagy derived nutrients are necessary for optimal F. tularensis replication, but F. tularensis still replicated in cells with decreased ATG5-independent autophagy. This indicates that F. tularensis uses other nutrient acquisition strategies in conjunction with ATG5-independent autophagy to supply nutrients for rapid and efficient proliferation.
Rapid bacterial proliferation requires readily available and abundant carbon and energy sources, commodities that are typically limited in the eukaryotic cell environment. Intracellular pathogens must acquire all required nutrients from the host cell, but the strategies that these pathogens employ to accomplish this task are only beginning to be characterized and vary widely between pathogens [9], [10], [31]–[33]. For example, Legionella pneumophila uses the byproducts of host proteosomal degradation rather than autophagy to obtain amino acids for energy [31]. Dengue virus growth is supported by autophagy mediated release of lipids while autophagosome formation increases nutrient availability for Anaplasma phagocytophilum [9], [10]. It is likely that other intracellular pathogens that successfully avoid autophagic destruction benefit from the nutrients that are released by this process. Thus, autophagy subversion through various means may be a more common strategy for pathogens to acquire nutrients from the host than previously thought.
The conclusion that autophagy derived amino acids were sufficient to rescue intracellular growth was supported by the fact that the absence of amino acids in tissue culture media did not appreciably affect F. tularensis intracellular replication. Thus, host cell amino acid import was not required to support bacterial growth. This result would seem to contradict the recent observation that knocking down expression of the amino acid transporter SLC1A5 decreases F. tularensis LVS growth approximately 2-fold [32]. LVS is an attenuated F. tularensis vaccine strain that, like fully virulent F. tularensis, grows within macrophages and other cell types, but is significantly less virulent than F. tularensis and other wild type F. tularensis strains in humans and animal models of infection. Unlike F. tularensis Schu S4, we found that LVS intracellular growth was significantly impaired in ATG5−/− MEFs and growth in these cells was restored by supplying excess amino acids, implying that LVS harvests nutrients via ATG5-dependent autophagy or another ATG5-dependent mechanism (data not shown). It is therefore likely that LVS is less reliant on ATG5-independent autophagy to support efficient intracellular growth. It is also possible that SLC1A5 contributes to the export of free amino acids out of autolysosomes thereby making autophagy derived amino acids available to the cytoplasmic bacteria. Amino acid transporters export amino acids from autolysosomes to the cytosol in Saccharomyces cerevisiae, and a similar system likely exists in mammalian cells [34]. This latter possibility highlights the fact that currently little is known about how free amino acids derived from autophagic degradation of host proteins are transported within eukaryotic cells.
Canonical autophagy destroys several different pathogens, including replication deficient and chloramphenicol treated F. tularensis [17]. The slight increase in bacterial replication in ATG5 −/− MEFs compared to wild type MEFs supports the notion that canonical autophagy can degrade wild type bacteria in MEFs, although this may be cell type specific as there is no difference in F. tularensis replication between wild type and ATG5−/− bone marrow derived macrophages [17]. Also, induction of autophagy by starvation or Torin1 treatment did not reduce bacterial replication. Surprisingly, although we observed mTOR inhibition in J774 cells and autophagy induction in ATG5−/− MEFs, our results suggest that canonical autophagy is either at or close to basal levels 16 hours post inoculation. Our results suggest that F. tularensis suppresses canonical autophagy downstream of mTOR or that mTOR is inhibited in ATG5-independent autophagy and other signals help determine which autophagy pathway is induced.
In contrast to xenophagy via canonical autophagy, ATG5-independent autophagy is involved in the lifecycle of two other intracellular bacterial pathogens. Mycobacterium marinum and Brucella abortus are each sequestered into an autophagosome-like structure via an ATG5-independent pathway as part of their intracellular lifecycles [8], [35]. It is unclear why M. marinum is sequestered, but bacterial sequestration by autophagy appears to be part of the B. abortus intracellular lifecycle and may benefit the bacteria by increasing cell to cell spread rather than providing nutrients [8], [35]. Both of these interactions with ATG5-independent autophagy are different from that of F. tularensis. What remains to be determined is if this difference is due to bacterial manipulation, if there are multiple ATG5-independent autophagy pathways, or if there are different functions for the same ATG5-independent autophagy pathway. Unfortunately, there is little information about how the various autophagy pathways are functionally different. We found that ATG5-indepdendent autophagy, unlike canonical autophagy, does not appear to use two proteins associated with xenophagy during infection. Further characterization of how xenophagy and ATG5-independent autophagy are associated may reveal why certain pathogens induce ATG5-independent autophagy.
Little is known about how ATG5-independent autophagy is induced or the role that it plays in a healthy eukaryotic cell, let alone during pathogenesis. However, there appears to be distinct benefits for certain pathogens to induce ATG5-independent autophagy over the canonical autophagy pathway. Determining how this pathway is induced in F. tularensis infected cells may give us insight as to how different autophagy pathways are initiated and how these pathways differentially impact intracellular pathogen survival and growth.
Francisella tularensis subsp. tularensis Schu S4 was obtained from Biodefense and Emerging Infections Research Resources Repository. For inoculation of eukaryotic cells Schu S4, Schu DSred, Schu S4-GFP [15] and Schu S4 – LUX (plasmid from [22]) were each grown initially on chocolate agar supplemented with 1% isovitalex then overnight in Chamberlain's defined broth media (CDM).
J774A.1 macrophage-like cells (J774) cells were maintained in 4.5 g/L glucose Dulbecco's minimal essential media (DMEM) with 10% FBS and supplemented with L-glutamine and sodium pyruvate. Mouse embryonic fibroblasts (MEFs) were maintained in 4.5 g/L glucose DMEM with 10% FBS. For treatment of MEFs without amino acids, DMEM with 4.5 g/L glucose was made following the ATCC DMEM protocol without adding amino acids and supplemented with 10% dialyzed FBS.
Human monocyte derived macrophages (hMDMs) were obtained by isolating peripheral blood mononuclear cells (PBMCs) from blood via ficoll gradient centrifugation. PBMCs were cultured for 2 hours in RPMI with 10% FBS and then washed to remove non-adherent cells. The adherent cells were cultured for 2 weeks in RPMI containing 10% FBS and 3 ng/ml GM-CSF (Biolegend). The media was replaced every 2 days. Experiments were performed using PBMCs isolated from peripheral blood from 2 healthy volunteers who gave informed, written consent following a protocol approved by the Institutional Review Board for human volunteers at University of North Carolina at Chapel Hill. Peripheral blood was obtained specifically for these experiments.
Stable Beclin-1 knockdown (TRCN0000087289 or TRCN0000087291) and scramble cell lines were generated by transducing MEFs with lentivirus encoding each shRNA. Cells were propagated in media containing 1 µg/ml puromycin for 2 weeks prior to the first experiment to select for transduced cells. Concurrent with the first experiment and last intracellular bacterial proliferation assay in the knockdown cell lines, mRNA was harvested from the transduced cells, subjected to reverse transcription, and was analyzed by quantitative RT-PCR to determine the amount of Beclin-1 mRNA present in each sample. The results were normalized to a GAPDH control. Primer sequences in are in Table S1.
3-methyladenine (10 mM) (Sigma), bafilomycin A(1) (200 nM) (Sigma), and chloroquine (160 µM) (Sigma) were each added with 25 µg/ml of gentamicin to the MEFs 3 hours post bacterial inoculation. Brefeldin A (17 µM) (Sigma) was added to J774 cells 3 hours post inoculation. Torin1 (250 nM) (Tocris Biosciences) was added overnight prior to inoculation and maintained throughout the infection. The excess amino acid mixture (12 mM L-amino acids containing aspartic acid, arginine, cysteine, histidine, isoleucine, leucine, lysine, methionine, proline, serine, threonine, tyrosine, and valine), L- serine (15 mM) or pyruvate (18 mM) were added at the same time as the inhibitors. All media was brought to a pH of 7.5.
Inhibitor cytotoxicity in MEFs was determined using a Live/Dead Fixable Green Dead Cell Stain kit (Invitrogen) following the manufacturer's instructions. Drugs were placed on cells for the same duration they would be on cells during infection (21 hours for Baf and CQ, 29 hours for 3MA). Percent cytotoxicity by flow cytometry was determined by gating. Cytotoxicity of F. tularensis in J774 cells 16 hours post inoculation was determined by testing the amount of lactate dehydrogenase (LDH) in the supernatant with a CytoTox-Glo cytotoxicity kit (Promega) following the manufacturer's instructions. Percent cytotoxicity was determined based on media and digitonin treated controls. Brefeldin A cytotoxicity was determined 21 hours post treatment using an In vitro Toxicology Assay Kit (Sigma) to measure LDH release from J774 cells.
MEFs were plated at 2×105 cells per well in 24 well tissue culture treated plates and grown overnight. MEFs were inoculated with a multiplicity of infection (MOI) of 100 with wild type Schu S4. The media was removed 3 hours post inoculation and replaced with media containing 25 µg/ml of gentamicin to inhibit the growth of any remaining extracellular bacteria. MEFs were lysed by vortexing for 1 minute and the lysates were serially diluted and plated on chocolate agar to calculate the number of intracellular bacterial cells at the indicated times.
hMDM cells were inoculated with an MOI of 100 wild type Schu S4 in RPMI containing 10% FBS. At 2 hours post inoculation, the media was replaced with media containing 10 µg/ml of gentamicin. At 4 hours post inoculation, the media was replaced with media that did not contain gentamicin. Intracellular bacteria were quantified as described previously.
Bacterial intracellular growth kinetics was calculated by measuring luminescence of Schu S4 – LUX infected MEFs or J774 cells. MEFs and J774 cells were plated at 5×104 cells per well in 96 well black wall clear bottom polystyrene plates (Corning) the night before infection. Each well was inoculated at an MOI of 100 with Schu S4- LUX and treated with gentamicin and inhibitors as described above. Luminescence was measured every 30 minutes using an Infinite M200 Pro plate reader (Tecan) maintaining constant 37°C temperature and 5% carbon dioxide.
All intracellular growth assays were performed in triplicate for each independent experiment. All of the inhibitors were added 3 hours post inoculation to reduce the impact of the inhibitors on F. tularensis phagosomal escape.
Bacterial growth curves of broth cultures were generated by measuring the optical density at 600 nm (OD600 every 15 minutes) using an Infinite M200 Pro plate reader (Tecan) maintaining constant temperature (37°C). To test toxicity of each drug on Schu S4, the bacteria were grown in CDM overnight, and then diluted to an OD600 of 0.05 in CDM containing the indicated inhibitors. CDM glucose substitution media were made without added glucose and 30 mM of the defined amino acid or carbon source. 50 mM MES buffer was added to all CDM media in the glucose substitution experiments.
For confocal fluorescent microscopy images depicting the number of bacteria in drug treated cells, MEFs were plated at 1×104 cells per well in an 8 well chamber slide (Nunc) and grown overnight. MEFs were inoculated at a MOI of 100 with Schu S4-GFP or Schu S4- DSred and treated with 25 µg/ml of gentamicin as described above. At the indicated time post inoculation, the MEFs were washed and fixed with 4% paraformaldehyde for 15 minutes and then washed again in PBS. To stain the plasma membrane, 10 µg/ml of AF647 conjugated wheat germ agglutinin (Invitrogen) was added to the fixed cells for 5 minutes and then washed away. DAPI containing mounting media (Vector Shield) was added to the slides to identify the nucleus.
Infection frequency was determined by fixing GFP infected MEFs 5 or 6 hours post inoculation and comparing the number of cells containing green puncta to the total number of cells completely within the field of view.
To quantify LC3B puncta, GFP-LC3 MEFs were generated by transfecting MEFs attached to an 8 well chamber slide (Nunc) with an eGFP-LC3 plasmid (Addgene plasmid 21073) [26]. 18 hours after transfection, the media was replaced with fresh media for one hour. After one hour, the cells were either infected with Schu-DSred or placed in fresh media. 3 hours post inoculation, the media in all wells was replaced with media containing 25 µg/ml gentamicin. 14 hours post inoculation, Torin1 or media lacking amino acids was added to the appropriate wells. The cells were fixed as above and stained with a mouse anti-GFP antibody (1∶250 dilution, Millipore) followed by an AF488 anti-mouse secondary antibody (Invitrogen) as previously described.
To quantify acidic vacuoles and determine co-localization with polyubiquitin and p62, MEFs were initially prepared as described above but were incubated for 2 hours in the presence of 150 ng/ml of LysoTracker red (Invitrogen) beginning at 14 hours post inoculation. The cells were washed and MEF media was added for an additional 10 minutes at 16 hours post inoculation. The cells were fixed in 4% paraformaldehyde and treated with 10 mM ammonium chloride following fixation. The MEFs were incubated with a polyubiquitin antibody (1∶1000 dilution, Enzo Life Sciences) or a p62/SQSTM1 primary antibody (1∶250 dilution, Abnova) followed by an AF647 conjugated anti-mouse secondary antibody (Invitrogen). DAPI containing mounting media (Vector Shield) was added to the slides to identify the nucleus. Images were acquired using a Zeiss 700 confocal laser scanning microscope (Carl Zeiss SMT, Inc.). Image acquisition, contrast adjustments, and cropping were all performed using Zen 2011 (Carl Zeiss SMT, Inc.).
Acidic vacuoles, p62, and polyubiquitin puncta were quantified by setting thresholds using ImageJ [36]. Only polyubiquitin puncta outside of the nucleus were counted. Co-localization of p62 or ubiquitin puncta with acidic vacuoles was determined by manual counting overlap. Any acidic vacuole or bacteria that overlapped any portion of the puncta was considered to co-localize.
To determine the distance between acidic vacuoles and F. tularensis, Z-stacks of LysoTracker red stained cells were taken using a Flow View 500 confocal laser scanning microscope (Olympus America). The distance between the bacteria and the acidic vacuoles was determined using ImageJ [36]and Corsen [37], following the protocols described in Jourdren et al. Additional protocol information and ImageJ plug-ins were available at http://transcriptome.ens.fr/corsen. The distance between objects was measured from the surface of the bacteria to the closest surface of the nearest acidic vacuole. To decrease the impact of noise, acidic vacuoles and bacteria with a volume of less than 0.05 µm (as determined by the Corsen program) were not included in the analysis.
To monitor transfer of amino acids from the host cell to F. tularensis, 4×105 MEFs were incubated in cysteine and methionine free DMEM containing 10% dialyzed FBS and 0.125 mC of S35 radiolabelled cysteine and methionine (EasyTag Express 35S, Perkin-Elmer) for 18 hours. 10 µg/ml of cycloheximide was added with the radiolabel in the indicated sample. The MEFs were then washed once and then incubated with DMEM containing 10% FBS for 2 hours. DMEM contains in excess of 100,000 times more cysteine and methionine than the initial radiolabel. The MEFs were then inoculated with F. tularensis Schu S4 at an MOI of 100 for 3 hours in fresh media. At 3 hours post inoculation, the media was replaced with media containing 25 µg/ml of gentamicin and either Baf or 3MA, as indicated, and supplemented with either a 12 mM amino acid mixture or 18 mM serine. The cells were washed in PBS, scraped from the plate, and lysed by vortexing the in PBS 16 hours post inoculation. The cell lysates were mixed with streptavidin coated magnetic beads (Solulink) that were pre-bound to biotinylated anti-F. tularensis lipopolysaccharide antibody (US biological). The anti-F. tularensis LPS antibody was biotinylated using a Biotin-xx protein labeling kit following the manufacturer's instructions (Invitrogen). The bead lysate mixture was incubated at room temperature for 20 minutes and then washed three times on a magnet. After the final wash, an equal volume of beads was added to 20% trichloroacetic acid (TCA) to make a final concentration of 10% TCA. The TCA mixture was mixed with an equal volume of 5% BSA and spun to pellet the TCA insoluble fraction. The TCA soluble fraction was removed and the TCA insoluble fraction was resuspended in PBS, added to scintillation fluid, and the number of counts was measured. An aliquot of the sample after the final wash was plated on chocolate agar to determine the number of bacteria present. The percent of radiolabel that was incorporated into F. tularensis was calculated by dividing the radiolabel counts from samples taken immediately prior to infection by the difference between the infected and uninfected samples.
To evaluate host protein degradation, J774 cells were radiolabeled for 24 hours, chased with non-radioactive media, inoculated and treated with gentamicin as described above. At 16 hours post inoculation, the cells were washed in PBS and lysed in RIPA buffer. The lysate was spun immediately to pellet the insoluble fraction. The soluble fraction was harvested and added to an equal volume of 20% TCA. The TCA insoluble fraction was then prepared and quantified as above.
Uninfected and Schu S4 infected J774 cells or ATG5−/− MEFs were maintained on small plastic tissue culture dishes. 25 µg/ml of gentamicin was added 2 hours post inoculation for J774 cells and 3 hours post inoculation for MEFs. 16 hours post inoculation the cells were fixed for 1 hour at room temperature in 2% paraformaldehyde, 0.5% glutaraldehyde in 0.15 M sodium phosphate buffer at pH 7.4. The cells were then rinsed in buffer and post-fixed with 0.5% osmium tetroxide/0.15 M sodium phosphate buffer, pH 7.4, for 10 minutes.
TEM samples for J774 cells were prepared similarly, although the cells were post-fixed for 1 hour in 1% osmium tetroxidein 0.15 M sodium phosphate buffer at pH 7.4 and then stained en bloc with 2% aqueous uranyl acetate for 20 minutes.
Both fixed samples were dehydrated in ethanol (30%, 50%, 75%, 100%, 5 minutes each step) and infiltrated and embedded in L.R. White Resin (Electron Microscopy Sciences). The dehydrated samples were sectioned en face (parallel to the substrate) at 70 nm, mounted on 200 mesh nickel grids, and post-stained with 4% uranyl acetate followed by Reynolds' lead citrate. Samples were observed with a LEO EM910 transmission electron microscope operating at 80 kV (Carl Zeiss SMT, Inc.) and digital images were acquired using a Gatan Orius SC1000 CCD Digital Camera with Digital Micrograph 3.11.0 (Gatan).
For the phospho- S6 ribosomal protein western blots, J774 cells were inoculated with Schu S4 at an MOI of 100 and treated with 25 µg/ml gentamicin 2 hours post inoculation. The uninfected sample had media replaced and media containing gentamicin added at the same times as infected samples. The uninfected samples were harvested 24 hours post inoculation. At the indicated times, cells were lysed by adding water containing phosphatase (Roche) and protease inhibitor cocktails (Pierce) and vortexing. The lysates were filtered through two 0.22 µm filters, separated on an SDS-PAGE gel under reducing conditions and then transferred to a nitrocellulose membrane. The membranes were probed with rabbit anti- S6 ribosomal protein or rabbit anti- phospho S6 ribosomal protein (Ser 235/236). All primary antibodies were obtained from Cell Signaling Technologies. Membranes were then probed with a horse radish peroxidase conjugated goat anti-Rabbit IgG (KPL) and bands were detected using an ECL Western Blotting Detection Kit (GE Life Sciences). Densitometry analysis was performed using ImageJ and comparing the amount of phosphor S6 ribosomal protein to the total amount of S6 ribosomal protein at the same time point [36]. The densities were then normalized to the uninfected sample.
Fold change was determined by subtracting each sample from the average of 3 samples taken at 5 hours post inoculation and a Mann-Whitney test was used to determine significance. The rest of the bacterial proliferation assays were pooled across experiments, log10 transformed, and then analyzed by a two-tailed Student's t-test were used to measure statistical significance. Significance for bacterial kinetic experiments was performed by pooling the maximum luminescence of each replicate for each experiment and performing a Mann-Whitney test. All luminescence and bacterial proliferation experiments were performed in triplicate in each experiment unless otherwise stated. Statistical significance for the distance measurement between F. tularensis and acidic vacuoles was performed using a two tailed Student's t-test on the pooled distance measurements across all 3 experiments for each sample. Significance for radiolabel incorporation into F. tularensis was determined by a Mann-Whitney test.
Morphology analysis was performed on the transmission electron micrographs by outlining the whole cell, nucleus, and each bacteria or autophagic vacuole in ImageJ to determine the area of each [36]. Morphology was determined with the aid of the following references [38]–[40]. Any rips in the slice were excluded from this analysis. Each micrograph depicted the nucleus and all infected cells had at least one bacteria present in the slice. The area of cytoplasm was determined by subtracting the area of the nucleus and bacteria from the area of the whole cell. At least 20 cells of each sample were examined and significance was determined by a two tailed Student's t-test.
|
10.1371/journal.pbio.1001381 | Clade Age and Species Richness Are Decoupled Across the Eukaryotic Tree of Life | Explaining the dramatic variation in species richness across the tree of life remains a key challenge in evolutionary biology. At the largest phylogenetic scales, the extreme heterogeneity in species richness observed among different groups of organisms is almost certainly a function of many complex and interdependent factors. However, the most fundamental expectation in macroevolutionary studies is simply that species richness in extant clades should be correlated with clade age: all things being equal, older clades will have had more time for diversity to accumulate than younger clades. Here, we test the relationship between stem clade age and species richness across 1,397 major clades of multicellular eukaryotes that collectively account for more than 1.2 million described species. We find no evidence that clade age predicts species richness at this scale. We demonstrate that this decoupling of age and richness is unlikely to result from variation in net diversification rates among clades. At the largest phylogenetic scales, contemporary patterns of species richness are inconsistent with unbounded diversity increase through time. These results imply that a fundamentally different interpretative paradigm may be needed in the study of phylogenetic diversity patterns in many groups of organisms.
| Species richness varies by many orders of magnitude across the evolutionary "tree of life." Some groups, like beetles and flowering plants, contain nearly incomprehensible species diversity, but the overwhelming majority of groups contain far fewer species. Many processes presumably contribute to this variation in diversity, but the most general explanatory variable is the evolutionary age of each group: older groups will simply have had more time for diversity to accumulate than younger groups. We tested whether evolutionary age explains differences in species richness by compiling diversity and age estimates for nearly 1,400 groups of multicellular organisms. Surprisingly, we find no evidence that old groups have more species than young groups. This result appears to hold across the entire tree of life, for taxa as diverse as ferns, fungi, and flies. We demonstrate that this pattern is highly unlikely under simple but widely used evolutionary models that allow diversity to increase through time without bounds. Paleontologists have long contended that diversity-dependent processes have regulated species richness through time, and our results suggest that such processes have left a footprint on the living biota that can even be seen without data from the fossil record.
| One of the most striking large-scale patterns in biology is the uneven distribution of species richness across the tree of life. Some groups are characterized by nearly incomprehensible species diversity (beetles, grasses), yet many other groups are species-poor (tuataras, ginkgoes). Evolutionary biologists have long been preoccupied with identifying the causal mechanisms underlying these differences in species richness [1]–[3]. These mechanisms include a vast range of biological, historical, and geographic factors. For example, lineage-specific molecular evolutionary traits (e.g., rates of molecular evolution or genome duplication) might be associated with net rates of species diversification [4],[5]. Likewise, species diversification rates might be a function of ecological traits, including those associated with the use of novel resources or defense from natural enemies [6],[7]. The list of factors that have been linked to differential diversification rates is substantial and continues to increase [8]–[11].
The most general explanatory variable of all is clade age [12]: clades vary in age, and this age variation should lead to differences in clade diversity, particularly if all clades have identical net rates of species diversification through time. If clade diversity is generally increasing through time, there is a strong theoretical expectation that species richness should be associated with their age (Figure S1). Even if individual clades are characterized by a “balanced” random walk in diversity, such that speciation and extinction rates are exactly equal, we may still observe a positive relationship between age and richness through time if clade diversity is conditioned on survival to the present day (Figure S1). Stochastic models of clade diversification through time consistently suggest that species richness and clade age should be correlated [13],[14]. These expectations differ from patterns observed for extinct clades [15],[16], presumably because living clades have survived to the present to be observed. The expectation that age and diversity should be correlated does not minimize the importance of evolutionary “key innovations” [7],[17],[18] and other factors as determinants of clade richness. In fact, to the extent that such factors influence net diversification rates, their effects should further accentuate differences in richness attributable to age variation alone.
Surprisingly, previous analyses have reached contrasting conclusions regarding the importance of clade age as a determinant of species richness [12],[13],[19],[20]. For some groups, clade age does not appear to predict species richness, suggesting that clade richness is regulated by diversity-dependence of speciation and extinction rates [14],[21],[22]. Some have suggested that this pattern lacks generality and that that is merely to be expected when clades vary in net diversification rates [20],[23]. The nature of the age-diversity relationship critically influences how we analyze and compare patterns of species richness among clades and between geographic regions. If age and richness truly are decoupled, then species richness in clades should not be modeled as the outcome of a simple time-constant diversification process, as is done in the overwhelming majority of evolutionary and biogeographic studies.
In this study, we evaluate the relationship between clade age and species richness across 1,397 clades of multicellular eukaryotes, including fungi, plants, arthropods, and vertebrates. We explicitly incorporate phylogeny into our analyses to ask the following questions: (i) What is the overall relationship between clade age and species richness across major clades of eukaryotes? (ii) Can simple models of among-clade variation in diversification rates account for the observed relationship between age and richness? (iii) How does the nature of this relationship vary across major subclades of eukaryotes?
We tested the relationship between clade age and species richness using a recent time-calibrated super-phylogeny [24] that spans virtually the entire tree of life and that contains a record of the phylogenetic relationships and stem clade ages of 1,592 higher taxonomic groups (e.g., families of beetles). We surveyed the literature for data on the extant species richness of all multicellular eukaryotic clades contained within this timetree, including fungi, plants, arthropods, and vertebrates. We obtained richness estimates for a total of 1,397 clades, totaling more than 1.2 million species (Figure 1).
Using phylogenetic generalized least-squares (PGLS) regression [25], we find no relationship between clade age and log-transformed species richness across the full set of 1,397 major clades of multicellular eukaryotes (Figure 2; t = 0.438; p = 0.66; df = 1395; β = 0.0008, where the regression coefficient β is the change in log-transformed diversity per million years). Use of non-phylogenetic regression models to analyze the age-richness relationship is inappropriate for these data, due to significant phylogenetic signal in clade size across the timetree (variance in independent contrasts test: p<10−20). We found that high phylogenetic signal in clade size can result in extremely high Type I error rates when the data are analyzed with OLS regression models, even when there is no true relationship between age and diversity (see Materials and Methods; Figure S2). Our results do not break down for younger clades: we found no relationship between age and log-transformed richness for the 307 clades younger than 50 Ma (β = −0.0251; p = 0.122; df = 305). Similar results were found for other subsets of the data (e.g., subsets of all clades less than 50, 100, 150, 200, and 250 Ma; Table S1; β≤0 for all analyses). Thus, there is no evidence that diversity increases asymptotically with respect to clade age.
We then examined the relationships between age and richness for the most densely sampled higher taxonomic groups within the timetree (Figure 3). Within this set of 12 major groups (1,133 clades total), only beetles show a significant relationship between age and richness (PGLS β = 0.017, p = 0.004). We repeated this analysis across all 352 subtrees within the timetree that contained at least 10 terminal clades and found no evidence that these patterns are simply an artifact of looking at “major” taxonomic groups (Figure S3). Moreover, the significant age-diversity correlation within beetles (Figure 3) is almost entirely attributable to a single subtree containing just 22 terminal clades (Figure S3). Because beetles represent the sole group showing a positive age-diversity correlation, we repeated our analyses on a comprehensive time-calibrated tree of 327 beetle subfamilies from a previous study [26], with the prediction that patterns observed at the family level should hold for more comprehensive subfamily-level sampling. We find no relationship between clade age and species richness at this scale (Figure S4; PGLS β = −0.002, t = −0.54; p = 0.59; df = 325), raising the possibility that the results we observe for beetles are a consequence of the large number of statistical tests we performed. We note that our analyses should have been biased in favor of detecting a significant age-diversity relationship as we did not correct any tests for multiple comparisons.
Substantial variation among clades in net rates of species diversification should weaken the expected relationship between clade age and species richness [14], and previous studies have found that diversification rates show phylogenetic signal across the branches of phylogenetic trees [3],[27],[28]. To address among-clade rate variation, we used the MEDUSA model [3] to estimate the extent of diversification rate variation within each of the 12 major groups shown in Figure 3. MEDUSA analyses strongly supported the presence of multiple rate shifts within each group (Table 1).
The MEDUSA model assumes, but does not test, whether constant-rate diversification processes can account for observed patterns of species richness within higher taxa. To test whether the MEDUSA model of rate variation could result in the age-diversity relationships we report here, we performed a posteriori simulations under the fitted MEDUSA parameters and evaluated the model-predicted relationship between clade age and species richness. Performing simulations under the MEDUSA model is challenging, because it requires a stochastic model that can account for the origin of higher taxa as well as for the occurrence of diversification rate shifts on phylogenetic trees. Our implementation assumed a two-state birth-death process, where the units are (i) individual lineages and (ii) higher taxa (see Materials and Methods). We modeled the origin of higher taxa as point occurrence events on the branches of phylogenetic trees; the occurrence of these events can be viewed as analogous to the acquisition of a phenotypic or ecological feature that defines a particular named higher taxon. We further assumed that diversification rate shifts occur within individual lineages under a Poisson process defined by the fitted MEDUSA model. We computed the Spearman correlation between clade age and species richness for each age-diversity dataset generated by the MEDUSA process and compared these distributions to the observed rank-correlations.
Our results indicate that the MEDUSA model of rate variation cannot explain the observed lack of relationship between clade age and species richness (Figure 4). For 10 of the 12 groups, the observed correlation between clade age and species richness is significantly less than the model-predicted correlation (p<0.05). Even for beetles, the correlation between age and richness is much lower than expected under the MEDUSA model (p<0.002). The two groups for which the MEDUSA model could potentially explain the observed age-diversity correlation (actinopterygiians and gymnosperms) were characterized by the smallest number of subclades (N = 12 in each case). The mean age-diversity correlation for each null distribution (Figure 4) is highly correlated with the number of subclades in the dataset (r = 0.88; p<0.001; Figure S5), suggesting that the effects observed for actinopterygiians and gymnosperms may be manifestations of small sample sizes.
The MEDUSA-based simulations described above are explicitly phylogenetic, in that closely related lineages tend to share common diversification parameters. We also considered a non-phylogenetic model of rate variation whereby each clade diversifies under a constant-rate birth-death process but with individual clade rates drawn from some overall distribution of rates [13],[14]. We implemented this model in a Bayesian framework, assuming that clade rates were drawn from a lognormal distribution [14] but with no phylogenetic signal in the resulting distribution of rates. To test whether this “relaxed rate” model could explain the lack of relationship between age and richness, we conducted posterior predictive simulation by (i) sampling parameters from their joint posterior distributions under the model, (ii) using the sampled parameters to simulate clade species richness, and (iii) using PGLS to evaluate the relationship between clade age and (simulated) species richness. We then computed the standardized effect size (SES) for the observed PGLS slopes to determine whether the observed age-diversity correlation is less than expected if net diversification rates among clades follow a simple lognormal distribution.
As with the MEDUSA simulations (Figure 4), our results reject the hypothesis that among-clade variation in net diversification rates can explain the lack of relationship between age and richness (Table 2). For every combination of subclade and relative extinction rate, the observed slope of the age diversity relationship is lower than the corresponding model-predicted value.
Clade age and species richness are decoupled across major clades of multicellular eukaryotes. When considering the full set of 1,397 clades, we found no significant relationship between age and species richness. When the data are partitioned into major subgroups (Figure 3), only beetles are found to have a significant age-diversity relationship. However, a more comprehensive analysis of age-diversity relationships in beetles reveals no relationship between age and richness (Figure S4). We found little evidence for positive age-diversity relationships for individual subtrees containing at least 10 terminal lineages (Figure S3). We found that among-clade variation in net diversification rates is unlikely to explain the lack of relationship between age and richness in any subgroup using two general approaches to model heterogeneity in diversification rates (Figure 4; Table 2). A MEDUSA-type model where diversity in taxonomic groups is produced by rate shifts along a phylogenetic backbone predicts strong positive relationships between clade age and species richness (Figure 4) as do non-phylogenetic models of diversification rate variation (Table 2). Even for beetles, the observed correlation between age and richness is significantly lower than expected under all models of diversification rate heterogeneity.
Although error in the estimates of clade age could theoretically weaken an age-diversity relationship, we consider it unlikely that such error accounts for the patterns we report here. We performed simulations to evaluate the amount of error in clade age that would be required to eliminate a true positive relationship between clade age and species richness (Figure S6). Additional work is needed to fully address this problem, but our results suggest that even extreme error in divergence time estimation is unlikely to eliminate this relationship entirely. These results are consistent with analyses suggesting that inferences about diversification rates from higher taxa are relatively robust to uncertainty in divergence times [29].
Our finding that clade age does not predict species richness challenges a fundamental assumption in most phylogeny-based diversity studies. Previous analyses of limited taxonomic scope have reached different conclusions about the relationship between clade age and diversity [12],[13],[20],[30],[31]. Here, we have demonstrated that (i) the lack of relationship between age and richness is a ubiquitous feature of recognized higher taxa and (ii) this pattern cannot be explained by variation in net diversification rates across the tree of life.
A number of possible mechanisms can account for this general pattern: it may reflect diversity-dependence of speciation and extinction rates [1],[32]–[35]; it may reflect a mixture of expanding and declining diversity trajectories across clades; or it may be an artifact of the way we delimit some clades (but not others) as named higher taxonomic groups (e.g., families). It is also possible that a lack of comparability across clades contributes to the overall lack of relationship between age and richness, and it would be interesting to test whether these results hold at finer phylogenetic scales (e.g., genera within families). Regardless of the underlying causal mechanism, a general decoupling of age and diversity at this scale has profound implications for how we measure and compare diversification and species richness across higher taxa.
If diversity-dependent processes regulate species richness within clades [1], then clade age should be a poor predictor of species richness [21],[36]. Clade age will predict species richness only when clades are growing through time. This type of diversity-dependent control is fundamentally related to Simpson's notion of “adaptive zones” [18]: higher taxa, such as the clades we consider in this study, would thus represent monophyletic groups of species that have radiated into a set of related ecological niches. This line of reasoning also implies that diversity dynamics are governed by clade-specific carrying capacities.
Macroevolutionary carrying capacities represent an important component of adaptive radiation [37],[38] and are intrinsically linked to the notion that ecological opportunity influences the tempo and mode of species diversification through time [39]–[41]. We may not understand the ecological mechanisms underlying “carrying capacity” dynamics, but we must still wrestle with substantial neontological and paleontological evidence for their existence. These include patterns of lineage and phenotype diversification as inferred from molecular phylogenies [40],[42]–[44], diversity rebounds after mass extinction [45]–[47], diversity-dependence of speciation and/or extinction rates [33],[48], long periods of diversity-constancy through time [32],[49], and double-wedge patterns of clade turnover through time [50]. Explosive radiations into novel adaptive zones have also been suggested to underlie long-term patterns of phenotypic evolution in a broad range of taxa [51]. In some groups, morphological innovations appear to have promoted shifts in carrying capacities even within geographically restricted radiations [35].
The central challenge in ascribing diversity-dependent causality to the age-diversity relationship in higher taxa is to explain why carrying capacity dynamics would pertain to sets of named higher taxa. The existence of a clade-specific carrying capacity implies that there is something special about named clades themselves, and there is no reason to accept this explanation if higher taxa are effectively random clades with no special meaning. However, higher taxa are clearly not random draws from the tree of life: major clades frequently comprise sets of taxa that are highly distinct in both phenotypic and ecological space (e.g., whales, bats, and carnivores within mammals). In a Simpsonian framework, recognized higher taxa are those clades that have acquired ecological innovations enabling them to radiate in new regions of ecological space, and there is nothing random about our recognizing them as such.
We note that a positive relationship between age and richness need not imply an absence of diversity-dependent regulation of speciation-extinction dynamics. Indeed, positive relationships between stem clade age and richness are expected even under strong diversity-dependence, at least during the initial phase of diversity expansion [36],[52]. However, once clades have reached carrying capacity, age and richness should become decoupled, as has been observed in analyses of several species-level molecular phylogenies [53],[54].
An alternative explanation for the lack of relationship between age and species richness is that the dataset contains clades undergoing both diversity increase and diversity decline. Paleobiologists have long noted that clades in the fossil record tend to wax and wane through time [1],[15],[50],[55]. At least intuitively, it seems reasonable that older clades are more likely to be on the “decline” phase of a diversity trajectory, as has been suggested for snakes [56]. This would provide an immediate explanation for the observed lack of relationship between age and diversity, and would link the patterns described here to the rise and fall of species richness in the fossil record [1],[15].
We find little evidence for a “hump-shaped” relationship between species richness and time (Figures 2–3), one possible pattern that may be consistent with declining diversity scenarios [15],[56]. However, we have only recently begun to explore the mechanisms by which diversity declines might shape age-diversity relationships in extant clades [56]. Recent studies suggest that it may be difficult to detect the signal of diversity declines even with complete species-level molecular phylogenies [57]. Fully addressing the role of diversity declines will presumably require the integration of neontological with paleontological data [58].
It is possible that the lack of relationship between clade age and richness is an artifact of the non-random manner by which higher taxa are recognized and which has nothing to do with the underlying process of diversity regulation [14]. Clearly, some property of clades causes us to recognize some as cohesive, named units (Aves, Squamata, Actinopterygii); we know very little about the consequences of such taxonomic ranking.
Perhaps the clades we recognize as higher taxa represent a subset of clades that have accumulated exceptional phenotypic distinctiveness relative to other clades. Such clades might, in turn, be those clades that have had lengthy and independent evolutionary histories during which to accumulate sufficient evolutionary change to merit recognition as a distinct higher taxonomic group. One prediction of this model is that named higher taxa would represent crown clades with exceptionally lengthy stem branches. Thus, higher taxa themselves might represent units delimited (albeit indirectly) by a property related to their age, and this could potentially compromise general conclusions about the relationship between clade age and species richness. Likewise, named higher taxa might correspond to clades that have undergone substantial shifts in the tempo and mode of phenotypic evolution [59]; this property itself might be associated with shifts in the dynamics of species diversification. We can at best acknowledge the possibility that the age-diversity relationship might be a statistical artifact attributable to yet-unknown perceptual biases that cause us to name a select subset of the total set of available clades across the tree of life.
Constant-rate estimators of “net diversification rate,” which assume a sustained increase in species richness through time, remain exceedingly popular for studying the dynamics of diversification from molecular phylogenetic data [3],[20],[60],[61]. This is undoubtedly due in part to the analytical tractability of these methods. Recent methods have been developed for accommodating temporal changes in rates of species diversification on complete species-level phylogenies [53],[62]–[66], but constant-rate estimates remain widely employed in the study of diversification patterns for higher taxonomic levels (but see [13],[14],[56]). At the phylogenetic scales we consider here, constant-rate diversification rate estimates may not be meaningful. This may also be true for the widely used MEDUSA model of rate variation [3], which appears to be incapable of recovering age-diversity relationships consistent with patterns observed in real datasets. If species richness is independent of stem clade age, time-constant models will misleadingly produce rate estimates that are negatively correlated with clade age. Our results suggest that, when age and diversity are not correlated, the significance of rate estimates in macroevolutionary studies should be interpreted with extreme caution since these estimates may offer little insight into the actual underlying processes that regulate species richness within clades [14],[36]. This is true regardless of the underlying causes of the observed age-diversity relationship: even if the absence of an age-diversity relationship is a statistical artifact of the manner by which we recognize higher taxa, our results imply that estimates of diversification rates for higher taxa may have little to do with the factors that influence clade species richness. We are unaware of any theoretical or empirical evidence demonstrating that “constant rate” estimators of net diversification, as applied to stem ages for extant clades, provide any useful insight into evolutionary processes in the absence of a positive relationship between clade age and species richness.
The relationship between clade age and species richness is fundamental to interpreting the effects of ecological, life-history, geographic, and other factors on clade diversity. A positive relationship between age and richness implies that species richness in clades is controlled by net rates of species proliferation. A decoupling between age and richness implies that other factors exert primary control on richness, or that clade diversity may be declining through time. The notion that species richness in clades can be decoupled from time seems counterintuitive, but is the expected outcome of diversity-dependent regulation of speciation-extinction dynamics. It is possible that species richness across the clades considered here is shaped by a mixture of processes, including diversity-dependence, declining rates, and rate heterogeneity. We are presently unable to determine the relative importance of these and other candidate processes, but integrating other data types (paleontological data; species-level molecular phylogenies) into studies such as this may provide a fruitful avenue for future research. In addition, further research is needed on the nature of higher taxa and the possibility that the results reported here might be a purely statistical consequence of the non-random process by which systematists have designated some clades as higher taxonomic groups. However, we are not presently aware of any non-biological mechanism that can account for this lack of relationship. Our results suggest that large-scale phylogenetic diversity patterns reflect constraints on species richness within clades rather than sustained diversity increases through time.
We used a recently published timetree for the tree of life in our analysis [24]. The timetree represents a synthesis of ∼70 time-calibrated, mostly interfamilial studies generated by experts on major taxonomic groups. Although diverse phylogenetic methods were used to generate and time-calibrate these topologies, high congruence in age estimates was observed between the most inclusive timetrees that linked major subsections of the tree of life together and the lower level timetrees contained within each subsection (see Chapter 3 in reference [24]). The combined timetree thus broadly summarizes our current understanding of the timing of major splits across the tree of life and provides a framework for investigating the tempo of diversification of extant lineages. We tabulated data on species richness of each terminal clade represented in the timetree using counts taken from the literature. We preferentially used data from published compendia of species or online checklists that formed parts of ongoing species databasing efforts. These resources were supplemented with richness estimates from other primary literature sources where no checklists were available. Many higher level clades in the timetree were incompletely sampled. In these instances (Table S2), we assigned richness of missing lineages to their closest sister lineage that was present in the time tree, collapsing clades if necessary. This resulted in a total of 1,226,871 species assigned to 1,397 clades.
We conducted simulations to test whether phylogenetic conservatism in clade size alone could generate significant age-richness correlations. Species richness is typically modeled as a geometric random variable, but incorporating covariance among clades due to shared evolutionary history is challenging. We assumed simply that the logarithm of species richness evolved across the phylogeny under a Brownian motion process. Strictly speaking, this is not a valid process-based model for the distribution of species richness across higher level phylogenetic trees. Specifically, this approach assumes that the “backbone structure” of the phylogeny is independent of the process that gives rise to richness at the tips of tree, as species richness is treated as a variable that can simply evolve across a pre-defined tree. This is unlikely to be valid in general, as both the phylogenetic backbone and the tip richness values presumably reflect common dynamic processes of speciation and extinction. However, our objective in these simulations was simply to test whether phylogenetic signal in clade size per se could lead to spurious relationships between clade age and species richness when no such relationship exists in the data, and we note that previous studies have analyzed this relationship in a non-phylogenetic framework [12],[67].
To loosely parameterize our simulations, we first estimated Pagel's lambda [68], which we denote by Λ, for the distribution of log-transformed species richness across the timetree. We found strong support for phylogenetic signal in log-transformed richness (ΔAIC = 372 in favor of model with Λ>0 versus non-phylogenetic model with Λ = 0; maximum likelihood estimate of Λ = 0.724). Using the maximum likelihood estimate of Λ and the corresponding Brownian motion parameters (root state and variance), we simulated 500 datasets under an unconstrained Brownian motion process with the fitted root state and variance parameters. Each simulation thus generated a distribution of log-transformed richness values, with a level of phylogenetic signal (Λ = 0.724) parameterized from the observed data, but with species richness values that are independent of clade age. Significant correlations between clade age and species richness were nonetheless observed in a majority of simulated datasets (Figure S2), despite no relationship between age and richness in the simulation model. This suggests that a simple tendency for closely related clades to be similar in size can lead to a highly misleading perspective on the relationship between age and richness and potentially explains positive age-diversity correlations reported in previous non-phylogenetic analyses [12],[67].
The MEDUSA algorithm [3] attempts to identify a mixture of constant-rate birth-death processes that can explain patterns of species richness across higher level phylogenetic trees. We fit the MEDUSA model to the 12 core “higher taxa” with substantial within-group sampling (see Figure 3). It was not feasible to fit a single model to the full dataset of 1,397 clades. Briefly, the algorithm uses a forward stepwise model selection procedure to incrementally add rate-shifts to a phylogenetic tree. The process ends when the addition of a new rate shift fails to improve the log-likelihood of the data beyond a pre-determined AICc (AICc, Akaike Information Criterion with finite sample size correction) threshold. These AICc thresholds for each subtree of N taxa were determined using the threshold selection function as implemented in the GEIGER package [69], where the threshold is computed as ΔAICc = A*(N−B)C+D. Default values for these parameters in GEIGER are A = −35.94105, B = 6.73726, C = −0.10062, and D = 27.51668. We modified the source code in the original MEDUSA implementation to allow extinction rates to exceed speciation rates, thus enhancing our ability to detect the signal of declining clade diversity through time.
We tested whether the MEDUSA model of rate variation could explain the observed lack of relationship between clade age and species richness by performing a posteriori simulations under the fitted models. We developed a simulation model for the MEDUSA process that enabled us to generate a phylogenetic backbone tree as well as higher taxonomic groups and associated species richness values. We assumed a two-state birth-death process, with units of (i) individual lineages and (ii) higher taxa. Our model adds two parameters to the speciation (λ) and extinction (μ) rates of the simple birth-death process. First, we assumed that higher taxa originate from individual lineages at a per-lineage rate Φ. These transitions are irreversible: individual lineages can transition to higher taxa, but the reverse transition is not permitted. Second, we assumed that lineages undergo transitions to new diversification rate classes with rate α.
Each simulation was initiated with n = 2 lineages, and simulations were run for a length of time equal to the crown age (Tc) of each major group shown in Figure 3. For each lineage, we sampled the waiting time to the next event from an exponential distribution with parameter β = λ+μ+Φ+α; the identity of the event was then sampled with probability proportionate to the event rate. For example, the probability of a higher taxon formation event would be Φ/β. Upon formation of a higher taxon at time T1, we assumed that the new taxon inherited the speciation and extinction parameters of the parent lineage; this is consistent with the MEDUSA model formulation, which allows rate shifts only along the internal branches of a phylogenetic tree. Given the remaining interval of time until the present day (t = Tc−T1), we then simulated clade richness (given λ, μ, and t) by sampling an integer-valued random variable from the expected distribution of progeny lineages under the birth-death process [70],[71]. We allowed higher taxa to become extinct before the present. The precise time of origin of a particular higher taxon (T1) cannot be inferred from the reconstructed phylogenetic trees generated by this simulation procedure; we can only know that the events that define higher taxa occurred at some time after the stem clade age of the group. Thus, phylogenetic trees generated by this algorithm are similar to the higher-level phylogenies analyzed in this and many other studies.
We constrained the per-lineage rate of higher taxon formation to be equal to the rate of speciation at any point in time. This decision was motivated by the observation that these rates must be roughly balanced under the model: for each phylogeny containing N higher taxa, we note that the interior “backbone phylogeny” necessarily contains N−1 speciation events (including the root node). Failing to allow approximate equality of these rates can lead to simulated trees consisting entirely of just a few higher taxa (if Φ>λ), or to trees consisting primarily of individual lineages that reached the end of the simulation without forming a higher taxon (if λ>Φ).
Each simulation was initiated by sampling a matched pair of speciation and extinction rates from the set of fitted rate classes inferred under the MEDUSA model. For the diptera, for example, we inferred nine rate shifts under MEDUSA, corresponding to a total of 10 rate classes (including the ancestral rates at the root). When a rate shift event occurred during the simulation, we sampled (with replacement) another matched pair of speciation-extinction rates from the set of fitted MEDUSA values. We set the shift rate equal to the maximum likelihood estimate under a Poisson process model of rate variation. This is obtained by noting simply that the observed number of rate shifts (e.g., nine for diptera) occurred on the internal branches of the phylogeny; an estimate of the event rate is thus given by the number of inferred events divided by the summed internal branch lengths of the phylogeny.
We automatically rejected any simulations that resulted in an exceptionally large or small number of terminals. We set the rejection threshold at 50% and 150% of the observed number of terminals for each dataset; for a dataset with 100 higher taxa, we would thus reject all simulated phylogenies with fewer than 50 or more than 150 terminals at the end of each simulation. We simulated 5,000 phylogenetic trees for each dataset.
As an alternative to the MEDUSA-based simulations described above, we also used a hierarchical Bayes approach to fit a non-phylogenetic “relaxed rate” model of diversification rate variation [14] to each of the 12 core subsets of the data (e.g., angiosperms, beetles, squamate reptiles) with substantial within-group sampling (see Figure 3). Here, we assumed that the net diversification rates for clades within each dataset were drawn from an uncorrelated lognormal distribution. We fit the model under both low (ε = 0) and high (ε = 0.99) relative extinction rates, where ε is the ratio of extinction to speciation rates. For each dataset (e.g., angiosperms), the model has two hyperparameters: the mean and standard deviation of the lognormal distribution of diversification rates. We used Markov Chain Monte Carlo (MCMC) to approximate the posterior distribution of all parameters and hyperparameters.
To assess whether this model could explain the lack of relationship between clade age and species richness, we conducted posterior predictive simulations by simulating species richness values for each clade under the fitted relaxed rate models. Unlike the MEDUSA analyses described above, these simulations treated the phylogenetic backbone tree as fixed; we thus performed phylogenetic GLS analyses on each simulated dataset. For each set of simulations, we computed the standardized effect size for the observed age-diversity relationship as SES = (βobs−βsim)/σsim, where βobs is the observed PGLS slope, and βsim and σsim are the expected mean and standard deviations of the slope from posterior predictive simulation. A negative SES value thus indicates negative displacement of the observed value relative to simulations.
|
10.1371/journal.pntd.0005502 | The novel nematicide wact-86 interacts with aldicarb to kill nematodes | Parasitic nematodes negatively impact human and animal health worldwide. The market withdrawal of nematicidal agents due to unfavourable toxicities has limited the available treatment options. In principle, co-administering nematicides at lower doses along with molecules that potentiate their activity could mitigate adverse toxicities without compromising efficacy. Here, we screened for new small molecules that interact with aldicarb, which is a highly effective treatment for plant-parasitic nematodes whose toxicity hampers its utility. From our collection of 638 worm-bioactive compounds, we identified 20 molecules that interact positively with aldicarb to either kill or arrest the growth of the model nematode Caenorhabditis elegans. We investigated the mechanism of interaction between aldicarb and one of these novel nematicides called wact-86. We found that the carboxylesterase enzyme GES-1 hydrolyzes wact-86, and that the interaction is manifested by aldicarb’s inhibition of wact-86’s metabolism by GES-1. This work demonstrates the utility of C. elegans as a platform to search for new molecules that can positively interact with industrial nematicides, and provides proof-of-concept for prospective discovery efforts.
| Many nematicides that have been used to kill plant and animal parasitic nematodes are being phased out over concerns of toxicity to humans. One potential solution to reduce toxicity is to use the nematicide at a lower concentration in combination with a second compound that together will produce a synergistic killing effect. That is, the use of either molecule alone at low concentrations is non-lethal, but when used together at these same concentrations, the cocktail is lethal. This strategy has two benefits. First, the killing effect is concentrated at the site of use and as the two molecules diffuse from the targeted site, toxicity is negated. Second, less of the toxic molecule is needed and therefore less is dispersed into the environment. Here, we describe our use of a model nematode called C. elegans to search for molecules that interact with aldicarb, which is one of the nematicides being phased out by environmental agencies. We identified 20 compounds that interact with aldicarb and describe how one of these, called wact-86, functions with aldicarb to kill worms. Our work provides proof-of-principle that C. elegans is a useful model for identifying compounds that positively interact with industrial nematicides and for understanding the nature of such interactions.
| Parasitic nematodes infect more than one billion people worldwide, negatively impacting human health and productivity [1,2]. Dramatic worldwide economic losses are incurred from nematode infections of commercially vital crops and livestock [3–5]. As a result of the growing resistance of nematodes to all of the major anthelmintic classes, the sustained utility of currently available treatments is in doubt, prompting the need for novel interventions [3,4,6,7]. Furthermore, unwanted toxicities associated with otherwise effective anti-nematode treatments has prompted usage restrictions and de-registrations for many nematicides [8,9], providing yet another avenue for attrition. Clearly, novel treatments targeted towards parasitic nematodes are desperately needed.
Aldicarb is one example of a particularly useful anti-nematode agent whose toxicity has limited its utility [10–12]. Aldicarb is a carbamate pesticide that has been used primarily to treat nematode, insect, and mite infections of various economically important crops including cotton and potato [13]. Aldicarb acts similarly to the organophosphate pesticides by inhibiting the enzyme acetylcholinesterase, which hydrolyzes and inactivates acetylcholine, resulting in the accumulation of acetylcholine at synapses [14]. The excess synaptic acetylcholine disrupts the neuromuscular activity of pest organisms, thereby restricting their mobility, arresting growth and impeding host infection. Aldicarb is also able to inhibit cholinesterase activity in non-parasitic animals, which is the mechanism by which it exerts its toxic effects [13,14]. Due to its improper use on watermelon crops in the early 1980s, over 2,000 people in California suffered cholinergic poisoning by aldicarb after eating the contaminated fruit [11,10]. In an effort to avoid additional poisonings, the environmental protection agency in the United States, and other similar agencies around the world, have enacted restrictions and bans on the use of aldicarb [9,15].
In principle, one approach to circumvent the toxicity of aldicarb, or any therapeutic with adverse toxicities, is to combine it with a distinct molecule that can potentiate its effects, such that lower concentrations can be used without compromising efficacy. Indeed, conjunctive therapies have been proposed to mitigate the toxicities of some cancer treatments [16]. In the case of aldicarb, potentiation would ideally not extend beyond the phyla of the parasites it is used to treat, so as to minimize unfavourable toxicity in the host and other non-target organisms.
Through its inhibition of acetylcholinesterase, aldicarb paralyzes and kills the free-living nematode Caenorhabditis elegans [17–19]. Thus, one way to find potentiators of aldicarb activity would be to screen chemical libraries, in combination with a sub-lethal dose of aldicarb, for compounds that interact with aldicarb to perturb C. elegans growth. Unlike many parasitic worms, C. elegans is readily amenable to high-throughput chemical screens, and it is cheap and easy to culture in the laboratory [20–22]. C. elegans is not a parasitic nematode, but the majority of commonly used anthelmintics are effective against C. elegans [23,24], and we and others have shown previously that C. elegans is a useful model for anthelmintic discovery [25,26]. All of these attributes provide a strong impetus for the use of C. elegans to screen for new chemical enhancers of aldicarb. Another salient feature of the C. elegans model is that the mode-of-action of newly discovered bioactive compounds can, in some cases, be determined using straightforward genetic and biochemical approaches [20,25]. An aldicarb interactor screen in C. elegans has the capacity to identify at least three different classes of compounds: those that potentiate aldicarb activity, those whose activity is potentiated by aldicarb, and those that show mutual potentiation or synergy with aldicarb. Molecules from all three classes hold promise as tools to combat parasitic nematode infection.
Here, we describe our screen of 638 worm-bioactive compounds for those that interact with aldicarb to perturb the growth of C. elegans. In total, we identified 20 compounds that interact with aldicarb. One of the hits from our screen is the novel worm-active and amide-containing compound wact-86. We use genetic and biochemical methods to demonstrate that wact-86 is hydrolyzed and detoxified in worms by the conserved carboxylesterase enzyme GES-1, and that aldicarb interacts with wact-86 by inhibiting its GES-1-dependent metabolism. Our work builds on ongoing efforts to discover and characterize new anthelmintic synergies [27], and provides proof-of-principle for future screening efforts aimed at identifying and characterizing chemical enhancers of other anti-nematode agents.
To find new compounds that interact with aldicarb we screened our in-house library of 638 worm-bioactive compounds [25], which we named the “wactive” library, in combination with a benign 10 μM dose of aldicarb, and assayed for combinations that disrupt the growth of C. elegans (S1 File; see Methods). As a single agent, aldicarb perturbs worm growth at concentrations above 1 mM (S1 Fig), so our aim was to uncover interactors that increase aldicarb potency by ~100-fold. The wactive library was screened in liquid media at a concentration of 1.5 μM–a condition where worm growth is indistinguishable from the solvent control for 95% of the compounds in the library (S1 File). Our screen identified 20 wactive compounds that perturb worm growth in combination with aldicarb, but are innocuous as single agents at the screening concentration (S2 Fig). The structures of the 20 compounds identified from our screen are shown in S3 Fig.
One of the strongest hits we obtained from our screen is wact-86 (N-{4-[(2-chlorobenzoyl)amino]-3-methoxyphenyl}-1-benzofuran-2-carboxamide; see S2 Fig), whose structure is shown in Fig 1A. We re-ordered wact-86 from a commercial source (see Methods), and verified its structure by mass spectrometry (S4 Fig). To validate the interaction between wact-86 and aldicarb we generated a combination dose-response matrix (Fig 1B). We found that maximal interaction is achieved when 20 μM aldicarb and 0.94 μM wact-86 are combined to kill C. elegans (Fig 1B). At these concentrations neither aldicarb nor wact-86 perturb the growth of worms as single agents (Fig 1B). This result is consistent with our primary screen data, and confirms the interaction between aldicarb and wact-86.
A search of SciFinder’s myriad chemical abstract databases revealed no published abstracts describing worm bioactivity for wact-86, or for any molecule sharing a pairwise structural similarity greater than or equal to 75% with wact-86, suggesting that wact-86 is a novel nematicide with an uncharacterized mechanism-of-action.
Towards better understanding the mode-of-action of wact-86, and by extension its mode of interaction with aldicarb, we carried out a genetic screen for wact-86 resistant mutants. This type of genetic approach has been used previously to identify the targets and the targeted pathways of bioactive compounds [20,25,28], as well as genes involved in drug detoxification and transport [29,30]. No resistant mutants were isolated in a screen of 100,000 mutagenized genomes in the second filial (F2) generation, suggesting that there are no recessive loss-of-function mutations that are sufficient to confer resistance to wact-86. We also screened 2.8 million mutagenized genomes in the F1 generation and were able to isolate three wact-86 resistant mutant strains (RP2809, RP2878, and RP2962). In contrast to wild-type worms, which are not viable at wact-86 concentrations greater than or equal to 1.88 μM, all three resistant strains display wact-86 resistance up to a concentration of at least 30 μM (Fig 2A).
To identify the wact-86 resistance-conferring mutations we sequenced the genomes of our three resistant strains and found that all three strains harbour missense mutations in the ges-1 gene (Fig 2 and S2 File). By contrast, no other gene in the genome has a protein-changing substitution in all three strains (S2 File). Furthermore, ges-1 is not mutated in 56 distinct mutagenized strains obtained from two unrelated genetic screens carried out by our group previously [25,31]. Thus, it is unlikely that ges-1 would be mutated in all three wact-86-resistant genomes by random chance alone. None of the ges-1 mutations are nonsense, frame-shifts, or deletions that are indicative of a loss-of-function. Instead, the missense mutations cause an A453V substitution in the RP2809 and RP2962 strains, and a M462V substitution in the RP2878 strain (Fig 2 and S2 File). These observations are consistent with the idea that these mutations confer a dominant gain-of-function phenotype that would be manifested in an F1 screen. Despite having the same missense mutation, RP2809 and RP2962 are clearly independently isolated mutants with distinct background mutations (S2 File). The RP2809 and RP2962 strains have greater wact-86 resistance compared with RP2878, perhaps indicating a correlation between ges-1 genotype and the wact-86 resistance phenotype (Fig 2A). Taken together, these data suggest that the ges-1 mutations may confer resistance to wact-86.
ges-1 encodes a carboxylesterase enzyme that catalyzes the hydrolysis of carboxylic ester bonds [32–34]. The GES-1 enzyme has relatively broad substrate specificity, and is thus classified as a non-specific esterase [32–34]. The expression of ges-1 is restricted to the intestine, pharynx, and rectum of C. elegans [32,35,36], but it is responsible for approximately half of the total esterase activity in worms [34]. In humans, orthologous liver carboxylesterases have been implicated in the hydrolysis of a number of drugs including cocaine and heroin [37], thereby facilitating their detoxification. In addition to carboxylic ester hydrolysis, carboxylesterases can also hydrolyse amide bonds, albeit less efficiently [38].
The GES-1 residues that are mutated in the wact-86 resistant mutants are in close proximity to residues that are conserved across phyla (Fig 2B). The high degree of conservation of these residues might implicate them as being important for enzymatic function. For instance, Ala453, which is mutated in two of the wact-86 resistant strains, is immediately C-terminal to His452, which is one of three conserved residues that make up the catalytic triad at the active site of the enzyme [39] (Fig 2B). Given their proximity to such highly conserved and functionally important residues, it is possible that the ges-1 mutations modify GES-1 activity.
Wact-86 contains two separate amide bonds that link three distinct aryl groups (Fig 1A). In light of the role carboxylesterases play in human drug metabolism, and given their ability to hydrolyze amide bonds, we hypothesized that the ges-1 mutations in our resistant mutants are gain-of-function, and that they confer wact-86 resistance by allowing for more efficient hydrolysis and detoxification of wact-86 by the GES-1 enzyme. This hypothesis is consistent with the expression of ges-1 in the pharynx and intestine, which is likely the point of entry for many xenobiotics into the tissues of the worm.
We have previously shown that drug metabolites in worm lysates can be separated, visualized, and quantified using a high performance liquid chromatography system coupled with a variable wavelength diode array detector (HPLC-DAD) [40]. We typically employ reversed-phase HPLC such that metabolites with greater aqueous solubility than the unmodified parent compound will elute earlier from the column than the parent structure (see Methods). To determine whether the wact-86 resistant mutants hydrolyze wact-86, we incubated RP2809, which contains a ges-1(A453V) mutation, and RP2878, which contains a ges-1(M462V) mutation, in 30 μM wact-86 for 2 hours, after which we lysed the worms and examined the contents of the lysates using our HPLC-DAD system. We identified two absorbance peaks in the wact-86-treated lysates that are absent from the DMSO control lysates (Fig 3A). The two peaks have retention times of 3.7 and 4.5 minutes, and absorbance maxima of 290 and 316 nm, respectively. The 4.5-minute peak likely corresponds to the wact-86 parent structure, since its retention time and absorbance spectrum are identical to the wact-86 standard (Fig 3A). The 3.7-minute peak is not present in the wact-86 standard, and it is absent from the lysates of heat-killed worms incubated in wact-86, suggesting that it may be a bona fide metabolite of wact-86 and not merely a wact-86 degradation product (Fig 3A).
To determine the structural identity of the presumptive wact-86 metabolite, we HPLC-purified it from the lysates of ges-1(A453V) and ges-1(M462V) mutants incubated in wact-86, and analyzed it by mass spectrometry (MS). For control purposes we collected the same HPLC fraction from lysates derived from worms incubated in DMSO alone and performed the same MS analysis. We identified a mass of 283.1 that was present in both of the mutant metabolite fractions, but was absent from both of the DMSO control fractions (Fig 3B and 3C). Hydrolysis of the amide bond that joins the 2-chlorophenyl and the anisole groups in wact-86 would produce two metabolites: N-(4-amino-3-methoxyphenyl)benzofuran-2-carboxamide (86-M1) and 2-chlorobenzoic acid (86-M2), which have exact masses of 282.1 and 156, respectively (see S5 Fig). The mass of 283.1 we identified in the mass spectra of our metabolite fractions is consistent with a protonated form of 86-M1. Fragmenting this mass by tandem MS/MS produced two abundant masses of 137.1 and 145.0, consistent with amide bond cleavage of 86-M1 (Fig 3B and 3C). Accurate mass determinations of the 283.1 mass confirm that the metabolite is indeed the wact-86 hydrolysate 86-M1 (S1 Table).
To test the hypothesis that our wact-86 resistant mutants hydrolyze wact-86 more efficiently than wild-type animals, we used our HPLC-DAD system to quantify the abundance of 86-M1 in the lysates of wild-type, ges-1(A453V), and ges-1(M462V) worms incubated in 30 μM wact-86 for 2 hours. Consistent with our hypothesis, the 86-M1 metabolite is 3 to 4-fold more abundant in the resistant mutants compared to wild-type worms (Fig 3A and 3D). If the resistant mutants metabolize and detoxify wact-86 more efficiently than wild-type animals then the amount of unmodified wact-86 should be greater in wild-type worms relative to the resistant mutants. Indeed, we found that the resistant mutants contain significantly less wact-86 in their tissues compared to wild-type worms (Fig 3E).
To test whether wact-86 hydrolysis depends on the activity of GES-1, we incubated two different ges-1 deletion mutants in 30 μM wact-86 and analyzed the worm lysates by HPLC-DAD. In contrast to the lysates obtained from wild-type and wact-86 resistant worms, we found that the deletion mutant lysates have no detectable wact-86 metabolite (Fig 3D), suggesting that wact-86 metabolism depends on GES-1 enzymatic activity. Altogether, these data support the idea that GES-1 hydrolyzes wact-86 in vivo, and that GES-1 activity is increased in the wact-86 resistant mutants, thus providing a mechanism for resistance.
In addition to inhibiting acetylcholinesterase activity, aldicarb is known to inhibit other carboxylesterase enzymes [41], including GES-1 [33]. Thus, one model to explain the interaction between aldicarb and wact-86 is that aldicarb inhibits the GES-1-dependent hydrolysis of wact-86, thereby preventing its detoxification and enhancing its nematicidal activity. This model espouses three predictions: 1) The wact-86 hypersensitivity of the ges-1 deletion mutants, if they are functionally null for wact-86 hydrolysis, should be similar to that of wild-type worms treated with aldicarb at a concentration affording maximal interaction with wact-86 (i.e. 20 μM aldicarb–see Fig 1B); 2) Aldicarb treatment should not further sensitize the ges-1 null mutant to wact-86; 3) Aldicarb should inhibit the GES-1-dependent hydrolysis of wact-86 in vivo.
In agreement with the first and second predictions, wild-type animals treated with 20 μM aldicarb phenocopy the wact-86 hypersensitivity exhibited by the ges-1(ok2716) deletion mutant (Fig 4A), and 20 μM aldicarb does not further sensitize this mutant to wact-86 (Fig 4A). The strain carrying the tm4694 deletion allele of ges-1 is also hypersensitive to wact-86, but less so than the strain carrying ok2716, suggesting that the tm4694 allele may retain some hydrolase activity. This result is perhaps not surprising, since the ok2716 deletion eliminates two residues of the catalytic triad, Ser198 and Glu319, the former being absolutely required for hydrolase activity [39,42], whereas the tm4694 deletion retains both of these residues (S6 Fig). Depending on its exact location, the tm4694 deletion may cause a premature stop codon upstream of His452, but the loss of this residue will not necessarily eliminate enzymatic function [43] (S6 Fig). The wact-86 hypersensitivity of the deletion mutants is consistent with these mutants containing 60 to 80% more wact-86 in their tissues relative to wild-type worms (Fig 3E). To test the third prediction, we incubated wild-type worms for two hours in 30 μM wact-86 together with 20 μM aldicarb, and analyzed the lysates using our HPLC-DAD system. Consistent with our prediction, aldicarb treatment inhibits the GES-1-dependent hydrolysis of wact-86, and results in the accumulation of relatively greater amounts of the wact-86 parent compound in worm tissue (Fig 4B). Taken together, our results suggest that aldicarb interacts with wact-86 to kill nematodes by inhibiting its GES-1-dependent hydrolysis and detoxification.
In addition to wact-86, our aldicarb interactor screen yielded 19 distinct compounds that interact positively with aldicarb, seven of which contain either an amide or ester group that could be metabolized by GES-1 (S2 and S3 Figs). To test whether GES-1 inhibition is the likely mechanism of aldicarb interaction for the additional hits, we performed dose-response assays for 13 out of the 19 compounds using wild-type worms, as well as the wact-86 resistant mutant RP2962 and the ges-1 deletion mutant RB2053 (S7 Fig). We found that the wact-86 resistant mutant is not consistently and robustly resistant to any of the compounds tested, suggesting that it is specifically resistant to wact-86. The ges-1 deletion mutant is weakly hypersensitive to 9 out of the 13 molecules tested, suggesting that inhibition of GES-1 activity may account for their interaction with aldicarb. Of these nine compounds, five do not contain an amide or ester group, suggesting that these compounds may have an amide or ester group introduced into their structure metabolically before GES-1 can hydrolyze them. For example, hydroxylation of the quinoline C2 carbon of wact-372, followed by enol-keto tautomerism, would reveal a secondary amide which could be hydrolyzed by GES-1, resulting in the opening of the quinoline ring and the potential inactivation of the compound. Regardless, four of the hits are interacting with aldicarb in a ges-1-independent manner, suggesting that their mode(s) of interaction are distinct from that of wact-86.
Here we used the free-living nematode C. elegans to screen for novel chemical interactors of a commercial nematicide. We identified 20 compounds that interact with aldicarb to perturb worm growth, and we characterized the mode of interaction for one of these, wact-86, in detail. Numerous lines of genetic and biochemical evidence show that the interaction between wact-86 and aldicarb derives from aldicarb’s inhibition of GES-1. We have shown that GES-1 hydrolyzes wact-86, which is lethal to C. elegans. Aldicarb’s inhibition of GES-1 therefore increases the potency by which wact-86 kills worms.
How wact-86 kills C. elegans remains unknown. Our previous work has shown that it is also able to kill C. briggase, and has some activity against at least two parasitic nematodes [25]. Exhaustive forward genetic screens for dominant and recessive C. elegans mutants that resist wact-86 failed to yield its target, suggesting that wact-86’s target is not genetically accessible. Furthermore, chemoinformatic searches using Scifinder Scholar and the Similarity Ensemble Approach online search tool [44] did not reveal any obvious candidate targets. Hence, other approaches will be needed to determine the mechanism by which wact-86 kills C. elegans.
GES-1 is the predominant esterase expressed in the intestine of C. elegans and likely has important roles in the metabolism of exogenous molecules and nutrients [34,45]. Because it is largely expressed in the intestinal lineage, it has been used as a marker for C. elegans gut development for nearly 30 years [32]. Despite its importance in metabolism, ges-1 mutants lack phenotypes that are obvious at the level of the dissection microscope [34]. However, animals deficient in GES-1 activity are hypersensitive to wact-86, while animals with increased GES-1 activity exhibit resistance to the lethal effects of wact-86. Hence, wact-86 is a new tool that can be exploited to genetically dissect ges-1 function.
Despite the mild interaction between aldicarb and wact-86, this work provides proof-of-principle that C. elegans can be a useful platform with which to: i) screen for new molecules that positively interact with known nematicides and, ii) understand the mechanism of their interaction. In addition to wact-86, our screen revealed 19 other compounds that interact with aldicarb, and a subset of these are likely interacting with aldicarb in a distinct, ges-1-independent manner. Future work may reveal the nature of these interactions with aldicarb.
The sources for the chemicals used in the aldicarb interactor screen are indicated in S1 File. For follow-up experiments, wact-86 was purchased from the ChemBridge Corporation and the Vitas-M Laboratory. Wact-86 from both vendors had comparable activity.
The N2 (wild-type) strain of C. elegans as well as the ges-1 deletion strain RB2053 were obtained from the Caenorhabditis Genetics Center (University of Minnesota). The strain harbouring the ges-1(tm4694) deletion allele was obtained from the Mitani Lab (Tokyo, Japan). All strains were cultured using standard methods [46]. The N2, RP2927, RB2053, and tm4694-containing strains were cultured at 20°C. To compensate for their relatively slower growth rates, RP2809 and RP2962 were cultured at room temperature (~22°C).
The aldicarb interactor screen was carried out in 96-well plates using our previously described C. elegans liquid-based chemical screening assay [25]. In brief, a saturated culture of HB101 E. coli was concentrated 2-fold with liquid nematode growth medium (see Ref. [25] for the NGM recipe). 80 μL of NGM+HB101 media was dispensed into the 96-well plate wells, and aldicarb (or DMSO alone for the control screens), was pinned into the wells using a pinning tool with a 300 nL slot volume (V&P Scientific). The wactive library chemicals were then pinned into the wells using the same pinning tool. Approximately 40 synchronized first larval-stage (L1) worms were added to each well in 20 μL of M9 buffer (see Ref. [47] for the M9 recipe). Synchronized L1s were obtained from an embryo preparation (see Ref. [47] for the protocol) performed the previous day. The final DMSO concentration in the wells was 0.6% v/v, the final aldicarb concentration was 10μM, and the final concentration of the wactive compounds was 1.5 μM. The plates were sealed with parafilm, placed upright into a Tupperware box containing many paper towels soaked with water, and then incubated at 20°C with shaking at 200 rpm for 6 days. After the 6-day incubation, a dissection microscope was used to count the number of viable worms in each well. The screen was repeated twice. Hits from the screen were identified as compounds that perturbed worm growth in combination with aldicarb in both replicates, but had no obvious effect on worm development as single agents.
The aldicarb dose-response experiments were carried out in 96-well plates using the liquid-based assay we have previously described (see above) [25]. 20 synchronized L1s were added to each well, and incubated for 6 days at 20°C. After 6 days, the number of viable worms in each well was counted, and the relative worm abundance was calculated by dividing the number of viable worms in a given aldicarb-containing well by the number of viable animals in the DMSO control well. Any well with 20 or more viable animals was counted as having twenty viable animals. Four technical replicates were performed, and the relative worm abundance was calculated as an average across the four replicates.
All of the dose-response experiments, with the exception of the aldicarb dose-response assay described above, were carried out in 24-well plates using a solid-based assay that we have described previously [47]. Briefly, in each well, the desired amount of wact-86 and/or aldicarb was dissolved in 1 mL of molten MYOB + 2% agar media (see Ref. [47] for the recipe), ensuring that the final concentration of DMSO (i.e. the vehicle) was 1% v/v. The plates were left overnight at room temperature to solidify. The following day, the plates were dried for 45 minutes in a sterile laminar flow hood, after which 25 μL of a saturated OP50 culture in LB media was deposited into each well. The plates were again allowed to dry overnight. The next day ~ 50 synchronized L1-stage larvae were added to each well in 10 μL of M9 buffer. The plates were wrapped in parafilm and stored upside down for 3 days at 20°C. On day 3, the number of viable worms in each well was counted. Relative worm abundance was calculated by dividing the number of viable worms in a given well by the number of viable animals in the DMSO control well. The dose-response experiments were performed at least three times, and the average relative worm abundance was calculated across the experimental replicates. Some of the wells had more worms deposited in them relative to the DMSO control, and so they have relative worm abundance values that exceed 1.
The forward genetic screen for wact-86 resistant mutants was carried out as previously described [20,25,47]. Briefly, wild-type parent (P0) worms were mutagenized with either 50 mM ethyl methanesulfonate (EMS) or 0.5 mM N-ethyl-N-nitrosourea (ENU) for 4 hours. For an individual screen, 100,000 synchronized L1s from the mutagenized F1 progeny were dispensed onto a 10 cm MYOB agar plate (see Ref. [47] for the protocol to make MYOB agar media) containing 50 μM wact-86. In total, 1.4 million mutagenized F1 animals were screened, which is equivalent to 2.8 million haploid genomes. Resistant worms were identified as those that can grow in the presence of the chemical. RP2878 was obtained from an EMS screen. RP2809 and RP2962 were obtained from ENU screens. Whole genome sequencing of the three wact-86 resistant mutants, and subsequent sequence analysis, was carried out as previously described (see Refs. [25] and [47] for a full description of our methods).
The multiple sequence alignment was carried out using Clustal Omega. The C. elegans sequence was obtained from WormBase (http://www.wormbase.org). All other sequences were obtained from the National Center for Biotechnology Information protein database.
Synchronized hatchlings were obtained from an embryo preparation of gravid adults (see Ref. [47] for the embryo preparation protocol). For the incubations, 60,000 hatchlings in 500 μL of M9 buffer (see Ref. [47] for an M9 buffer recipe) were treated with either 30 μM wact-86, 30 μM wact-86 in combination with 20 μM aldicarb, or DMSO alone for control purposes. The final concentration of DMSO in all samples was 1% v/v. Prior to the incubations, the hatchlings used for the dead worm controls were heat-killed at 37°C without aeration for 24 hours, and then at 95°C for 20 minutes. The incubations were carried out in standard 1.5-mL micro-centrifuge tubes on a nutating shaker, at 20°C for 2 hours. After the 2-hour incubation, the worms were transferred to the wells of a Pall AcroPrep 96-well filter plate (0.45-μm GHP membrane, 1-ml well volume), the buffer was drained from the wells by vacuum, and the worms were subsequently washed three times with 500 μL of M9 buffer. After washing, the worms were re-suspended in 35 μL of M9 buffer, transferred to a new standard 1.5-mL micro-centrifuge tube, and stored frozen at -80°C. The samples were later lysed by adding 35 μL of a 2X lysis solution (100 mM KCl, 20 mM Tris (pH 8.3), 0.4% SDS, 120 μg mL-1 proteinase K), and incubating the tubes at 56°C for 1 hour.
Prior to HPLC, 70μL of acetonitrile was added to the lysates. The samples were mixed by vortexing for approximately 10 seconds, and then centrifuged at 17,949g for 2 minutes. After centrifugation, 100 μL of the lysate was injected onto a 4.6 X 150 mm Zorbax SB-C8 column (5 micron particle size) and eluted with solvent and flow rate gradients over 5.2 minutes as indicated in Table 1.
UV-Vis absorbance was measured every 2 nm between 190 and 602 nm. Absorbance intensity data was converted to three-dimensional heat-mapped chromatograms using MATLAB (The MathWorks). Prior to processing the worm lysates, a 5 nmol amount of pure wact-86 was processed by HPLC to determine its elution time and absorbance spectrum. HPLC was performed using an HP 1050 system equipped with an autosampler, vacuum degasser, and variable wavelength diode-array detector. The column was maintained at room temperature (~22°C). HP Chemstation software was used for data acquisition and quantification. Area under the curve (AUC) was calculated using the Chemstation peak integration tool, using default settings. The AUC values plotted in Figs 3 and 4 are an average of at least three experimental replicates.
To purify the wact-86 metabolite, the HPLC fraction between 3.6 and 3.8 minutes was collected from two separate lysates, combined, and dried using a Genevac EZ-2 centrifugal evaporator. The identical fraction from DMSO control lysates was also collected and dried. The dried fractions were re-suspended in a minimal volume of 1:1 (v/v) methanol: 0.1% aqueous formic acid. Electrospray ionization mass spectrometry (ESI-MS) analyses were carried out using a 6538 UHD model quadrupole time-of-flight mass analyzer equipped with an atmospheric pressure ESI source and a 1260 Infinity model HPLC system (Agilent Technologies, Santa Clara, CA). Samples were analyzed via loop injection with mobile phase composed of 1:1 (v/v) methanol: 0.1% aqueous formic acid and flowing at a rate of 0.25 mL min-1. Mass spectra were recorded in the 2 GHz mode and the high-resolution MS analyses for molecular formula determinations were obtained using external calibration. Tandem MS/MS analyses were obtained via collision-induced dissociation using the targeted MSn function of the acquisition software. MS/MS spectra were recorded sequentially at three different fragmentation voltages (10, 20 and 30 V) and the resulting spectrum was composed of the average of those three collision energies.
|
10.1371/journal.pntd.0003160 | Achieving Population-Level Immunity to Rabies in Free-Roaming Dogs in Africa and Asia | Canine rabies can be effectively controlled by vaccination with readily available, high-quality vaccines. These vaccines should provide protection from challenge in healthy dogs, for the claimed period, for duration of immunity, which is often two or three years. It has been suggested that, in free-roaming dog populations where rabies is endemic, vaccine-induced protection may be compromised by immuno-suppression through malnutrition, infection and other stressors. This may reduce the proportion of dogs that seroconvert to the vaccine during vaccination campaigns and the duration of immunity of those dogs that seroconvert. Vaccination coverage may also be limited through insufficient vaccine delivery during vaccination campaigns and the loss of vaccinated individuals from populations through demographic processes. This is the first longitudinal study to evaluate temporal variations in rabies vaccine-induced serological responses, and factors associated with these variations, at the individual level in previously unvaccinated free-roaming dog populations. Individual-level serological and health-based data were collected from three cohorts of dogs in regions where rabies is endemic, one in South Africa and two in Indonesia. We found that the vast majority of dogs seroconverted to the vaccine; however, there was considerable variation in titres, partly attributable to illness and lactation at the time of vaccination. Furthermore, >70% of the dogs were vaccinated through community engagement and door-to-door vaccine delivery, even in Indonesia where the majority of the dogs needed to be caught by net on successive occasions for repeat blood sampling and vaccination. This demonstrates the feasibility of achieving population-level immunity in free-roaming dog populations in rabies-endemic regions. However, attrition of immune individuals through demographic processes and waning immunity necessitates repeat vaccination of populations within at least two years to ensure communities are protected from rabies. These findings support annual mass vaccination campaigns as the most effective means to control canine rabies.
| Canine-mediated rabies is a horrific disease that claims tens of thousands of human lives every year, particularly in Asia and Africa. The disease can be effectively controlled through mass vaccination of dogs with high-quality vaccines; however, questions remain over the effectiveness of vaccination where the health status of free-roaming dogs may be compromised and the life expectancy and access to these dogs may be limited. This study evaluated rabies-vaccine induced immune responses and vaccine delivery in previously unvaccinated, free-roaming dog populations in two rabies endemic regions in Asia and Africa, to better understand the effectiveness of vaccination campaigns. We found that the majority of dogs seroconverted to the vaccine regardless of health status. Excellent vaccination coverage was achieved through community engagement and door-to-door vaccine delivery, even where the majority of the dogs needed to be caught by net for vaccination. However, attrition of immune individuals through demographic processes and waning immunity reinforces the importance of frequent and regular vaccination campaigns to ensure effective vaccination coverage is maintained.
| Canine-mediated rabies is a viral zoonosis, causing at least 55,000 human deaths every year [1]. Mortality from rabies is highest in less developed communities in Asia and Africa, where domestic dogs are free-roaming [2]–[8]; with increasing evidence that the majority are owned [2], [3], [6], [9], [10] and, thus, generally accessible for vaccination [11], [12].
Canine rabies can be effectively controlled by vaccination [13]–[16] using readily available, high potency (antigenic value ≥1 IU/ml), inactivated cell-culture vaccines. These vaccines should provide protection from challenge in healthy dogs for the claimed period for duration of immunity [17], which is often two or three years. In free-roaming dog populations, vaccine-induced protection from rabies may be compromised for several reasons. These include: (a) insufficient vaccine delivery during vaccination campaigns [11], (b) lack of repeat vaccination campaigns, with loss of vaccinated individuals from populations through demographic processes [18], [19], and a substantial proportion of dogs probably vaccinated only once in their lifetime [20], despite them often living beyond three years of age [19]; and, (c) the possibility of immuno-suppression through malnutrition, infection or other stressors [21]–[23], which may reduce the proportion of dogs that seroconvert or the duration of immunity of those dogs that seroconvert. These constraints may result in a decline in the vaccination coverage between campaigns to below 20–45%, the threshold necessary to control rabies [24]. Consequently, investigating the effectiveness of vaccination campaigns under field conditions is critical.
The adaptive (B-cell humoral and T-cell cell-mediated) immune response to vaccination is complex. The humoral response generates virus neutralizing antibody (VNA), the primary correlate of protection induced by viral vaccines [18], [25]–[27]. Cell mediated immunity (CMI) is also important for the development of vaccine-induced immunity [28]–[30] and acts in synergy with the humoral response [27]. Ongoing protection from challenge depends on the persistence of long-lived plasma cells, continuing to generate antigen-specific antibody, and B- and T- memory cells. The primary antibody response following vaccination generally correlates with the strength of the memory response (B- and T-cell) and, thus, the ability to induce secondary responses to subsequent challenge [27], [31]–[35]. In healthy dogs the quality of the primary immune response to vaccination depends on several factors, including the type of vaccine, with modified-live vaccines generally inducing superior responses, the route of administration, and the dose of vaccine antigen [25], [27], [32], [33], [35]–[37].
Laboratory challenge studies in healthy dogs support these observations. Following seroconversion, protection from rabies virus challenge correlates with peak VNA titre and final titre prior to challenge for inactivated, DNA and modified-live vaccines, with increased susceptibility to challenge once titres drop to near negligible levels (VNA titres <0.1 IU/ml or mouse serum neutralizing antibody titres <1∶2 dilution) [31]–[35], [37]–[41]. These studies used comparable antibody assays [42], [43] and virus challenge doses. Titres measured repeatedly over 3–4 years initially peaked and then declined rapidly, followed by a more gradual decline [31], [33], [34], [44]. While a titre of 0.5 IU/ml demonstrates seroconversion following vaccination [45], the approximate threshold for protection following seroconversion may be 0.1 IU/ml [34], [37], [40], [46]. However, in the aforementioned experimental studies, only a proportion (<40%) of dogs with measureable titres following vaccination, but with negligible titres at the time of challenge succumbed to challenge, highlighting the importance of previously activated B- or T- cells allowing rapid response to challenge.
Although the same relationship between VNA titre and protection from challenge is expected in immuno-suppressed dogs as in healthy dogs [21], [23], no systematic comparison has been published to date. Reduced humoral immune responses have been shown in malnourished experimental dogs [22] and Gambian children vaccinated with human diploid-cell rabies vaccine [47], and pet dogs with anaemia or intestinal parasites vaccinated against rabies [37], [48]. Several studies have evaluated the immune response in previously unvaccinated, mostly healthy pet dogs to high potency, inactivated rabies vaccine under field conditions [48]–[54]. All of these studies report variable VNA titres up to 12 months following vaccination, including a proportion of dogs with titres ≤0.1 IU/ml (and generally a larger [17% to >42%] proportion with titres <0.5 IU/ml). These observations have serious implications for free-roaming dogs where their health status is more likely to be compromised. However, with the exception of one study in Peru [55], no study has evaluated variations in vaccine-induced VNA in previously unvaccinated free-roaming dogs where rabies is endemic. Furthermore, no study has properly evaluated the factors associated with these variations.
Cell mediated immunity is technically difficult to measure under field conditions [28], [56], however peripheral blood lymphocyte counts, which are predominately T-cells [57], may provide a straightforward, indirect assessment of CMI. Together with cytokine assays and measures of blastogenic responses of lymphocytes to mitogen, lymphocyte counts were used to assess immunomodulation in healthy dogs in response to vaccination [58]–[60] and protein-calorie malnutrition [22], and in humans in response to protein-calorie malnutrition [61]. In dogs, malnutrition induced declines in immunoglobulin and lymphocyte function and counts. Therefore, lymphocyte counts together with rabies vaccine-induced titres and nutritional status may correspond to the overall immune status of an individual and susceptibility to infection.
This study focused on evaluating temporal variations in vaccine-induced VNA, and factors associated with these variations, in three previously unvaccinated, owned free-roaming dog populations in South Africa and Indonesia, to better understand their effect on vaccination coverage. In addition, the efficiency of vaccine delivery and loss of vaccinated individuals from the cohorts were also assessed.
See Table S1 for a summary of the methodology. Data were collected from three cohorts of dogs, one in South Africa, and two in Indonesia. The cohorts were part of a larger ecological study that commenced in March 2008 [19]. The South African cohort was located in Zenzele, an informal settlement 10 km west of Johannesburg (26.15°S and 27.41°E). In Indonesia the cohorts were located in the study areas of Kelusa (8.26°S and 115.15°E) and Antiga (8.30°S and 115.29°E), two villages on the island of Bali. Kelusa, composed of six banjars (sub-villages), is inland. The study area encompassed the entire village except for Banjar Yehtengeh, separated from the rest of the village by rice fields and jungle, the southern half of Banjar Kelikikawan and the households scattered along the main road leading into the village. Antiga, a large village of six banjars, is located on the east coast. The bulk of the households are clustered into two banjars (Kaler and Kelod). The study area encompassed all of Kaler and Kelod. An additional area (Banjar Ketug) included households scattered along a 2.7 km stretch of road winding through the jungle north of Kaler and Kelod. Rabies is endemic in Indonesia and South Africa, with outbreaks occurring in Bali in 2008 and Gauteng Province in 2010.
The Zenzele research cohort included every available dog in the entire township (which was the study area) in February 2010 that had not been previously vaccinated by the Department of Agriculture (DoA) during a vaccination point (VP) on the outskirts of the township in October 2009 (Table S2). All the dogs vaccinated by the DoA were identified within one week of the one day VP through a rapid door-to-door search, with verification by owners and inspection of certificates. The DoA had also set-up a VP on the outskirts of Zenzele in May 2006, thus vaccination history and certification were checked with each owner at the start of the study. VNA titres were also evaluated for anamnestic responses to vaccination consistent with previous vaccination.
The Bali research cohorts included every available dog in the study areas of Kelusa and Antiga in January 2010 that had not been previously vaccinated by the Department of Livestock (DoL) as described below (Table S2). Prior to a rabies outbreak in 2008, vaccination against rabies was illegal in Bali and there had been no systematic vaccination programs in either village prior to commencement of the study. Vaccination points were set up by the DoL in two banjars in Kelusa in December 2009 and in one banjar outside of the study area in Antiga in February 2010. The VPs were poorly attended because of community awareness of the research vaccination program and because the owners could not readily handle their dogs. In Kelusa, 16 dogs from the study area attended the vaccination points. In Antiga only three dogs from the study area attended the vaccination point.
All of the dogs resident in the study area were owned and had been previously identified by household, name and appearance through intensive monitoring by direct observation and survey since March 2008. Intensive monitoring of all of the dogs in the study area continued until April 2011. Therefore, all of the dogs in the study population were readily identified at the individual level during the study period. There was no evidence for a resident population of unowned dogs [19], [62]. All dogs in their third month of life or older were photographed (standardised dorsal and lateral views). Pups in their first or second month of life were recorded but not photographed. The same enumerators had tracked the majority of the cohorts at the individual level since March 2008 and were familiar with the dogs.
Vaccine delivery was door-to-door for the research cohorts, and households were revisited repeatedly until the dog was caught for vaccination and blood sampling, or it was apparent that the dog could not be caught or the owner would not be available to give consent. A dog was also excluded from the study if the owner declined consent, the dog did not remain calm during restraint, there was a high index of suspicion that the dog may bite, or it was apparent the dog had a clinical condition that might have deteriorated as a result of restraint.
All the dogs were carefully restrained by experienced personnel using the correct equipment and under the direct supervision of a veterinarian. In Zenzele, dogs were gently restrained with a leash and soft muzzle. In Bali most dogs could not be safely restrained by leash and muzzle and required restraint by net. Vaccinations and blood sampling were undertaken by experienced veterinarians. High-quality, sterile consumables (i.e. needle, syringe and blood tubes) were used for each vaccination and blood sample. Dogs in the research cohorts were vaccinated with 1 ml of Rabisin [63], an inactivated rabies vaccine containing at least 1 IU/ml of rabies virus glycoprotein (GS57 Wistar strain) with an aluminium hydroxide adjuvant. Vaccine was administered subcutaneously into the neck or shoulder region. The vaccine cold chain was carefully preserved.
Rabisin and Galaxy DA2PPv, a polyvalent vaccine against common infectious pathogens, was administered by the DoA during the October 2009 VP in Zenzele. Some dogs vaccinated at the VP may have received ivermectin. The DoL administered Rabisin during the February 2010 VP in Antiga, and Rabivet Supra 92, a locally produced cell-culture vaccine, during the December 2009 VP in Kelusa. Vaccine administration and storage by the local authorities were not observed.
Different blood sampling schedules were required for Zenzele and Bali given the different methods of restraint and because the rabies outbreak in Bali escalated during 2009, forcing vaccination to be undertaken 6 months earlier than planned. Every dog in each research cohort, including neonates, was vaccinated at the start of the study (day 0) (Zenzele n = 259, Kelusa n = 284 and Antiga n = 259 vaccinated [Table S2]), and every available dog from about 6–8 weeks of age was blood sampled (see Table 1 and Table S3 for the number of dogs blood sampled at each time point).
Blood was collected from the Zenzele research cohort on day 0 (immediately prior to vaccination) and then approximately 30, 90, 180 and 360 days following vaccination. The dogs vaccinated by the DoA were also blood sampled 8–10 days after the VP. Samples were then collected approximately 30, 90, 180 and 360 days following the VP. In Zenzele, only those dogs that had been vaccinated were blood sampled. Rabies-vaccine induced VNA was measured at each time point. Complete blood counts (CBCs) were measured on days 0, 180 and 360 for the research cohort.
In Bali, samples were collected on day approximately 180 and 360 following vaccination. Every available dog, whether vaccinated or not, was blood sampled at both time points and analysed for rabies-vaccine induced VNA. Unvaccinated dogs constituted the control group, and included those dogs not caught for vaccination on day 0 and those that arrived into the study populations after day 0. The sixteen dogs in Kelusa and three dogs in Antiga vaccinated by the DoL, in December 2009 and February 2010 respectively, were blood sampled at the same time as the research cohort.
In all the sites, households were visited in approximately the same order at each time point, so the number of days between samples were similar for each dog.
For each sample, 5–7 ml of blood was collected from the jugular or cephalic vein and divided into plain and ethylene diamine-tetraacetic acid (EDTA) containing blood tubes. The blood tubes were immediately coded by date, house number and dog identification and placed in cool boxes with ice packs. Serum was separated by centrifugation within 8 hours of collection and refrigerated at 4–6°C for up to 48 hours prior to freezing. All the sera were transported frozen in dry shippers to the Weybridge Animal Health Veterinary Laboratory Agency in the United Kingdom for fluorescent antibody virus neutralization (FAVN) assays. EDTA whole blood samples were refrigerated and then tested within 48 hours of collection for CBCs. Approximately 10 grams of faeces was collected manually on day 0 from 107 dogs randomly selected from the Zenzele cohort for routine analysis. Upon collection, the faecal sample pots were similarly coded and kept in the cool boxes, then refrigerated until being tested. Complete blood counts and faecal analysis were undertaken by the Faculty of Veterinary Science, University of Pretoria. Suitable laboratory facilities were not accessible in Bali for these tests. Finally, 32 dogs from Kelusa and Antiga combined were selected on day 180 from those dogs diagnosed with generalised dermatitis during the preceding survey for deep skin scrapes (DSS) from affected areas of skin to determine the prevalence of Demodex spp. See text S1 and text S3 for an explanation of sample selection for the DSS and faecal analysis.
Factors that may influence the immune response to rabies vaccine were selected on their measurability under field conditions, particularly by vaccinators. These factors had been previously quantified at the individual level as part of the larger ecological study that commenced in March 2008, and the methods used to quantify the factors are described elsewhere [19]. In summary, the factors were categorical and measured by direct observation and questionnaire at the time of vaccination (gender, age class, pregnancy, lactation, sterilisation status [Bali only], intestinal parasites [Zenzele only]) or within 6 weeks of vaccination (body condition, clinical signs associated with serious illness, protein intake [Bali only], and generalised dermatitis [Bali only]) [21], [22], [37], [47], [48], [64]–[67]. See text S2 and Table S19 for a detailed description of the covariates. Time (points) was treated as a continuous variable.
The study was approved by the Ethics Committee, University of Cambridge [DVM/EC/1-2010], and the Animal Ethics Committee, University of Pretoria [v025-10 AUCC]. Permits to collect demographic data were granted by the Ministry for Research and Technology (RISTEK), Indonesia [03923/SIP/FRP/SM/IV/2010]. Blood samples were collected under the auspices of the Faculty of Veterinary Medicine, Udayana University, Bali [RG49780], and permits for vaccination and blood collection were granted by the Balinese provincial and regencies Departments of Livestock, the districts Centres of Animal Health (UPT) [RG49780], and Kesbang, Pol and Linmas (the combined Agencies for National Unity, Politics and Protection) [070/607.D.III and 070/015/D.II]. In all of the sites, informed consent was obtained prior to each survey and blood test from the community leaders and owners, who were kept fully informed of the purpose, approach and progress of the study. Vaccination and blood sampling were only carried out with the owner, or responsible adult delegated by the owner, present and their express consent.
Almost all of the dogs in the study populations were owned but free-roaming, with <10% confined continuously or frequently during the study period March 2008–April 2011. There was an approximately even ratio of male to female dogs in Zenzele, but the ratio was skewed towards males (approximately 75%) in Bali. Less than 2% of dogs were sterilised in Zenzele, but castration of juvenile male dogs by community members was common in Bali (approximately 14% in Kelusa and 27% in Antiga) [19]. Life expectancy was at least 3 years for the majority of dogs in the study populations (Table S4 and Figures S1a–S1c).
High vaccination coverage was achieved through door-to-door vaccine delivery: 82% (259/315) in Zenzele, 81% (284/351) in Kelusa and 79% (259/327) in Antiga. Similar coverage (75–86%) was achieved in Bali for blood sampling at day 180 and 360, despite many of the dogs having been caught on at least one previous occasion (Table S2). The characteristics of dogs that avoided capture are described in Table S5. The sex ratio and age distribution of these dogs were similar to the overall population (Figures S1a–S1c).
Attrition of the cohorts occurred during the study period through mortality, particularly of neonates, but also through the relocation and disappearance of dogs [19]. Of the 259 dogs vaccinated in Zenzele at the start of the study, 103 (40%) were sampled at the last time point. Similar proportions were recorded in Kelusa (44%, n = 124) and Antiga (49%, n = 126) (Tables S2 and S3).
In the Zenzele research cohort, upper outliers were defined as dogs with peak titres (on day 30) of 128 IU/ml or greater (n = 7). Some of these dogs were either in the study area in May 2006 or may have been previously independently vaccinated by their owner. Baseline titres of the upper outliers were ≤0.25 IU/ml, most with a titre of ≤0.09 IU/ml. The history of those individuals with the next highest titre (91 IU/ml) varied, and included seven dogs that were born in Zenzele after October 2009.
It is unlikely that any of the dogs vaccinated by the DoA four months prior to initiation of vaccination of the research cohort were inadvertently included in the research cohort. The day 0 titres of the research cohort (including upper outliers ranged from 0.06–1 IU/ml with a GMT of 0.1 IU/ml) were substantially lower than the day 90 titres of the DoA cohort (including upper outliers ranged from 0.06–128 IU/ml with a GMT of 2.8 IU/ml). Thirteen (20%) of the dogs vaccinated by the DoA had titres ≤1 IU/ml 90 days after vaccination, of which 6 had titres <0.5 IU/ml and four of these were non-responders (i.e. day 30 titre of <0.5 IU/ml). Only five dogs in the research cohort had day 0 titres ≥0.5 IU/ml, and of these none appeared to have an anamnestic response to the vaccine (day 30 titres ranged from 1.4–45 IU/ml) (Tables S6 and S11). There were no differences in the distributions of titres for dogs in Zenzele probably present in May 2006, when the DoA vaccinated, and those that arrived into the population after May 2006 (Table S12).
In the Bali research cohorts, upper outliers were defined as dogs with day 180 titres of 11.3 IU/ml or greater (n = 4 in Kelusa; n = 15 in Antiga). For some of these dogs, information provided by their owner, breed, source and geographical location was suggestive of vaccination undertaken independently by their owner or as part of vaccination campaigns outside of Kelusa and Antiga. Several (n = 15) unvaccinated controls had titres ≥0.5 IU/ml (Tables S7, S8, S9). The titres of the unvaccinated controls are summarised in Table S10.
The longitudinal, individual-level data from this study provides the most detailed serological data currently available for domestic dogs in rabies endemic areas, and provides valuable support for planning rabies vaccination programmes.
This study reinforces the importance of frequent and regular vaccination campaigns to ensure effective vaccination coverage is maintained. Dogs with lower peak titres had correspondingly lower titres at the end of the study, with titres <0.5 IU/ml at the last time point (day 360) for 20–40% of the dogs and <0.1 IU/ml for 3–8% of the dogs (Table S14); the implication being an increased susceptibility to natural exposure with time in the dogs with low titres [27], [34], [37], [40], [46]. Robust demographic data from these study populations indicates, two years after a pulse campaign which achieved 80% vaccination coverage, at least 20–45% vaccination coverage would remain [19], which is the critical threshold necessary to prevent rabies [24]. However, from our model predictions (Table S17), we speculate that a substantial proportion of the dogs remaining in Zenzele two years after vaccination may have titres <0.1 IU/ml, potentially dropping effective vaccination coverage to below the critical threshold. Models were constrained to two time points for the Bali cohorts, but predicted similar declines in the GMT for Kelusa.
The vast majority of the dogs seroconverted following vaccination (with a peak titre of ≥0.5 IU/ml), regardless of health status. However, there was considerable variation in titres at each time point for all the cohorts. Peak titres were not measured for the Bali cohorts, however day 180 titres were comparable to Zenzele, therefore it is likely that a similar proportion of dogs to Zenzele seroconverted following vaccination. Identification of risk factors associated with lower titres may promote targeted boostering to maintain vaccination coverage. Clinical conditions around the time of vaccination reduced the immune response to the vaccine in all the cohorts; in particular, generalised dermatitis provided a ‘visible marker’ for a reduced immune response, with practical implications for rabies control. While demodicosis was assumed to be an important cause of generalised dermatitis associated with immuno-suppression in Bali, the mostly negative skin scrapes suggests that dermatophytosis may be more likely, consistent with both the tropical climate and immuno-suppression [65], [66]. This warrants further investigation given that a substantial proportion of the dogs (37%–46% Table S24) were affected, potentially reducing the effectiveness of vaccination. Lactation at the time of vaccination in Zenzele and the Bali cohorts combined was significant statistically, however its biological significance is unclear. Lactation is associated with loss of body condition in all the research sites [19], consistent with immuno-suppression observed in Zenzele. The reason for the opposite effect in Bali cannot be readily explained [81], [82]. While this incongruity may warrant further investigation in larger study populations on balance lactating bitches should be vaccinated, with re-vaccination following weaning.
Our study demonstrated an advantage of community engagement and door-to-door vaccine programmes over the use of simple vaccination points. We consistently achieved vaccination coverage above 70% through door-to-door vaccine delivery, even in Bali where the majority of the dogs needed to be caught by net on successive occasions. Similar coverage was achieved across the rest of the island through door-to-door vaccine delivery in 2010 and 2011 [83]. This compares to a vaccination coverage of only 27% through the vaccination point in Zenzele and a very low vaccine uptake (5%) in Kelusa. The utility of vaccination points is likely to differ between locations according to local circumstances. Similar to other communities in Africa, Europe and central Asia where free-roaming dogs are handleable [11], [13], [84], [85], it is likely that the majority of the dogs in Zenzele could have been delivered to the vaccination point by their owners, and the low vaccination coverage was probably the result of inadequate advertising [86] and limited operating hours during a work/school day. Vaccine uptake in Kelusa was, in part, affected by community awareness of the research vaccination program, however the majority of the dogs could not be handled by their owners or the vaccinators, thus necessitating restraint by net [83]. The reasons for the difference in handleability between locations are unclear. Restraint by net is more stressful to the dog, time consuming and costly than by leash and muzzle. In order to improve welfare, facilitate more cost-effective and efficient delivery of vaccines (and other prophylactics), and improve evaluation of the dogs in Bali and similar communities, extending our studies to evaluate the differences in husbandry, environment and other factors influencing the temperament of the dogs in the sites is warranted.
This research has generated valuable data that may contribute to rabies control, including through improving epidemiological models. However, understanding variation between dogs in titres measured from field studies is challenging. Some covariates that may impact on titres, such as lactation and health status, are measurable, whereas others such as genetics and stress are harder to assess in real time. Further evaluation of factors associated with variation in immunity over time since vaccination, including both serological responses and direct assessment of CMI, and recording vaccine failures is warranted and may require larger populations studied and over longer time periods.
This study demonstrates that the vast majority of free-roaming dogs, in two regions of Africa and Asia where rabies is endemic, seroconverted to rabies vaccine regardless of health status, producing titres that exceeded 0.5 IU/ml, the level considered necessary to protect against rabies. Declines in vaccination coverage following a vaccination campaign occur through mortality/emigration of vaccinated dogs and birth/immigration of unvaccinated, susceptible dogs. Robust demographic data from the study populations show that two years after vaccinating at least 70% of dogs during a pulse vaccination campaign, vaccination coverage remained within 20–45% [19], the range necessary to control rabies (Hampson 2009). However, our serological data indicates that dogs with lower peak (day 30) titres had correspondingly lower end point (day 360) titres. We speculate that a proportion of vaccinated dogs remaining in the study populations after two years will probably have titres below the approximate threshold for protection (<0.1 IU/ml) thus dropping effective vaccination coverage to below the critical threshold (of 20–45%). This emphasizes the importance of re-vaccinating within two years. Vaccination of all dogs during annual campaigns is therefore recommended as the most effective means of ensuring that individual immunity and population coverage are both maintained at sufficient levels to control rabies.
|
10.1371/journal.pcbi.1007061 | The ability of transcription factors to differentially regulate gene expression is a crucial component of the mechanism underlying inversion, a frequently observed genetic interaction pattern | Genetic interactions, a phenomenon whereby combinations of mutations lead to unexpected effects, reflect how cellular processes are wired and play an important role in complex genetic diseases. Understanding the molecular basis of genetic interactions is crucial for deciphering pathway organization as well as understanding the relationship between genetic variation and disease. Several hypothetical molecular mechanisms have been linked to different genetic interaction types. However, differences in genetic interaction patterns and their underlying mechanisms have not yet been compared systematically between different functional gene classes. Here, differences in the occurrence and types of genetic interactions are compared for two classes, gene-specific transcription factors (GSTFs) and signaling genes (kinases and phosphatases). Genome-wide gene expression data for 63 single and double deletion mutants in baker’s yeast reveals that the two most common genetic interaction patterns are buffering and inversion. Buffering is typically associated with redundancy and is well understood. In inversion, genes show opposite behavior in the double mutant compared to the corresponding single mutants. The underlying mechanism is poorly understood. Although both classes show buffering and inversion patterns, the prevalence of inversion is much stronger in GSTFs. To decipher potential mechanisms, a Petri Net modeling approach was employed, where genes are represented as nodes and relationships between genes as edges. This allowed over 9 million possible three and four node models to be exhaustively enumerated. The models show that a quantitative difference in interaction strength is a strict requirement for obtaining inversion. In addition, this difference is frequently accompanied with a second gene that shows buffering. Taken together, these results provide a mechanistic explanation for inversion. Furthermore, the ability of transcription factors to differentially regulate expression of their targets provides a likely explanation why inversion is more prevalent for GSTFs compared to kinases and phosphatases.
| The relationship between genotype and phenotype is one of the major challenges in biology. While many previous studies have identified genes involved in complex genetic diseases, there is still a gap between genotype and phenotype. One of the difficulties in filling this gap has been attributed to genetic interactions. Large-scale studies have revealed that genetic interactions are widespread in model organisms such as baker’s yeast. Several molecular mechanisms have been proposed for different genetic interaction types. However, differences in occurrence and underlying molecular mechanism of genetic interactions have not yet been compared between gene classes of different function. Here, we compared genetic interaction patterns identified using gene expression profiling for two classes of genes: gene specific transcription factors and signaling related genes. We modelled all possible molecular networks to unravel putative molecular differences underlying different genetic interaction patterns. Our study proposes a new mechanistic explanation for a certain genetic interaction pattern that is more strongly associated with transcription factors compared to signaling related genes. Overall, our findings and the computational methodologies implemented here can be valuable for understanding the molecular mechanisms underlying genetic interactions.
| Understanding the relationship between genotype and phenotype of an organism is a major challenge [1,2]. One of the difficulties for unravelling genotype-phenotype relationship has been genetic interactions, when combinations of mutations lead to phenotypic effects that are unexpected based on the phenotypes of the individual mutations [3–5]. Large-scale analyses of single and double deletion mutants have revealed that genetic interactions are pervasive in many model organisms [6–11]. Recently, efforts have been initiated to investigate genetic interactions in human cell lines too, using large-scale RNA interference and Crispr-Cas9 knock downs [12–15]. Our understanding of the molecular mechanisms that underlie genetic interactions lags behind our ability to detect genetic interactions. Understanding the molecular basis of genetic interactions and their interplay with cellular processes is important for unraveling how different processes are connected [16–18], to what degree genetic interactions shape pathway architecture [6], as well as for understanding the role genetic interactions play in human disease [5,19].
One of the phenotypes that is frequently used to investigate genetic interactions is cell growth [6,20–28]. Based on this phenotype, genetic interactions can be broadly subdivided in two types, negative genetic interactions where the double mutant is growing slower than expected given the growth rate of the single deletion mutants, and positive genetic interactions where the double mutant is growing faster than expected [3]. Negative genetic interactions have frequently been associated with a redundancy relationship between two functionally related genes [29]. The redundancy mechanisms by which two genes can compensate for each other’s loss has been linked with close paralog genes or redundant pathways [30,31]. Positive genetic interactions have been associated with genes participating in the same protein complex or pathway [32]. There are however many exceptions to these rules and it also has become clear that there are many other hypothetical mechanisms underlying these genetic interactions that require further investigation [3,18].
Another phenotype that has been much less frequently used for investigating genetic interactions is gene expression [16,17,33–36]. Expression-based genetic interaction profiling provides detailed information at the molecular level which is beneficial for unraveling mechanisms of genetic interactions [16,17,33–36]. Unlike growth-based profiling, which gives a subdivision into either positive or negative interactions, expression-based genetic interaction profiling provides further subdivision into more specific genetic interaction patterns. These patterns have recently been systematically classified and include buffering, quantitative buffering, suppression, quantitative suppression, masking and inversion [17]. A more detailed sub classification that includes information on expression of downstream genes, can also contribute to understanding the underlying mechanisms by which two genes interact [16,17,37].
To provide mechanistic insights into biological networks, Boolean modeling has been used successfully [38,39]. It has also been applied to unravel regulatory networks underlying genetic interaction patterns between kinases and phosphatases [16]. Due to their intrinsically simple nature, such Boolean network models allow exhaustive enumeration of network topologies. The outcomes of these models can then be easily compared to the patterns observed in experimental data. Boolean operators however, are limited to on and off values and cannot easily accommodate quantitative measurements, which limits the types of genetic interaction patterns that can be investigated using this approach. Unravelling the regulatory network underlying genetic interaction patterns would benefit from modeling approaches that allow some degree of quantitativeness to be introduced while still being computationally feasible to exhaustively explore all potential models. In this way, Petri nets may be considered an extension of Boolean modeling that provides more flexibility, without the need to incorporate detailed prior quantitative knowledge [40–44]. Petri nets are able to capture both qualitative and quantitative traits and have successfully been applied to investigate genetic interactions before [45,46]. Petri net modeling would therefore allow investigation of all possible genetic interaction patterns in an exhaustive and semi-quantitative manner.
It is evident that genetic interactions are widespread in Saccharomyces cerevisiae [6] as well as other organisms [7,8]. Nevertheless, extensive characterization of the molecular mechanisms underlying genetic interactions, as well as a comparison of the molecular mechanisms underlying genetic interactions between different functional classes have, as yet, not been performed. Here, based on two existing data sets and corresponding functional classes, gene specific transcription factors (GSTFs) and signaling related genes (kinases and phosphatases) have been compared with regard to negative genetic interaction patterns and the possible underlying molecular mechanisms. This revealed that the two most common genetic interaction patterns are buffering and inversion. The prevalence of inversion however, is much stronger in GSTFs. Inversion, whereby genes show opposite behavior in the double mutant compared to the corresponding single mutants, as well as the underlying mechanism of inversion, are poorly understood. Exhaustive enumeration of network topologies using Petri net modeling reveals that the minimum requirement for observing inversion is having a quantitative difference in interaction strength (edge weight) from the two upstream transcription factors to a shared downstream gene. In addition, this quantitative edge difference is frequently accompanied by an intermediate node, that displays a buffering pattern. The proposed model provides a mechanistic explanation for inversion, thereby further aiding a better understanding of genetic interactions. GSTFs, more so than kinases/phosphatases, can modulate or fine-tune the activation levels of their target genes, which suggests quantitative differences in regulating downstream target genes are important for the functioning of GSTFs. This is consistent with the fact that inversion occurs more often between GSTFs than between signaling genes, as well as our observation that quantitative edge differences are required for inversion to occur and provides a likely explanation why inversion is more prevalent for transcription factors.
To investigate potential differences in mechanisms of genetic interactions between groups of genes with a different function, data from two previously published datasets using the same technical setup and platform were combined [16,17]. Both datasets include DNA microarray gene expression measurements as a result of deleting genes in the yeast Saccharomyces cerevisiae and have been subjected to rigorous quality control and statistical analyses [47]. The first dataset consists of genome-wide gene expression measurements of 154 single and double gene-specific transcription factor (GSTF) deletion mutants [17]. The second dataset contains genome-wide gene expression measurements of 54 single and double kinase/phosphatase (K/P) deletion mutants [16]. These studies applied different criteria to select for interacting pairs. Whereas the GSTF dataset includes both positive and negative genetic interactions, the kinase/phosphatase dataset was restricted to negative genetic interactions only. To avoid potential systematic biases, the selection criteria of the kinase/phosphatase dataset [16] were adopted and applied to both datasets. To summarize, selection was based on pairs having a significant growth-based negative genetic interaction score (adjusted p-value < 0.05, Methods) to include redundancy relationships that influence fitness. In addition, for a given double mutant, at least one of the corresponding single mutants has an expression profile similar to wildtype (WT) to ensure that genetic interactions such as redundancy are considered. An expression profile is considered similar to wildtype if no more than eight genes are changing significantly (adjusted p-value < 0.05, fold-change > 1.7). These selection criteria yield a uniform dataset consisting of 11 GSTF double mutants and 15 kinase/phosphatase double mutants as well as their respective single mutants (63 single and double mutants in total; S1 Table).
Genetic interactions can be investigated in different ways. Here, both growth as well as genome-wide gene expression is used to compare genetic interactions between GSTFs and kinases/phosphatases, as described before [17]. To summarize, a growth-based genetic interaction score εgrowth,XY between two genes X and Y is obtained by comparing the observed fitness for double mutant WxΔyΔ to the fitness that is expected based on both single mutants WxΔ · WyΔ (εgrowth,XY = WxΔyΔ - WxΔ · WyΔ) [48]. A gene expression-based genetic interaction score between two genes X and Y is calculated in two consecutive steps [17]. First, the effect of a genetic interaction between two genes X and Y on any downstream gene i is calculated as the deviation between the expression change observed in the double mutant Mi,xΔyΔ and the expected expression change based on the corresponding single mutants Mi,xΔ + Mi,yΔ (εtxpn_i,XY = |Mi,xΔyΔ –(Mi,xΔ + Mi,yΔ)|). The overall genetic interaction score between gene X and Y is then obtained by counting the total number of genes for which εtxpn_i,XY is greater than 1.5 [17]. Gene expression changes from single and double mutants were subsequently grouped into the six genetic interaction patterns, buffering, suppression, quantitative buffering, quantitative suppression, masking and inversion, as previously described (Fig 1A) [17]. Gene expression profiles of GSTFs and kinases/phosphatases do not show any obvious differences in the number of genes significantly changing (30 vs. 27 on average), show similar gene expression ranges (Fig 1B and 1C) and correlations between pairs involving either kinases/phosphatases or GSTFs are highly similar (S1 Fig). When investigating the genetic interaction profiles of GSTFs (Fig 1B) as well as kinases/phosphatases (Fig 1C), it is clear that buffering is prevalent in many of the larger genetic interaction profiles, but the degree of buffering differs for the smaller genetic interaction profiles.
Hierarchical clustering was applied to group pairs with similar genetic interaction patterns (S2 Fig), thereby disregarding the identity of individual downstream genes. From this clustering, it is clear that there is no distinct separation between pairs consisting of GSTFs and kinases/phosphatases. Instead, most pairs are characterized by large buffering effects, grouped together in a single large cluster (S2A Fig, red branch labeled as 1). This is not surprising, since all pairs are selected for having a significant growth-based negative genetic interaction score. This in turn is based on double mutants growing slower than expected based on the single mutants. Slow growing strains are known to display a common gene expression signature [49,50]. This slow growth gene expression signature is caused by a change in the distribution of cells over different cell cycle phases [51]. To facilitate investigating mechanisms of genetic interactions, such effects are better disregarded. Removing the slow growth gene expression signature is therefore expected to improve identification of direct effects and thereby aid systematic unravelling of the underlying mechanisms. As described before [51], the dataset was transformed by removing the slow growth signature (Methods). Removing the slow growth signature and thereby reducing indirect effects should also improve the identification of direct downstream target genes of the GSTFs included in the dataset. Four GSTF double deletion mutants have binding data available [52]. Investigating the enrichment before and after data transformation for direct downstream targets of Hac1/Rpn4 (S3A Fig), Met31/Met32 (S3B Fig), Gat1/Gln3 (S3C Fig) and Cbf1/Hac1 (S3D Fig) show a clear improvement in enrichment after data transformation for three out of four GSTF pairs as also shown before for individual GSTFs [51]. These results confirm that removing the slow growth signature improves the identification of direct effects and is therefore probably more suited when investigating mechanisms of genetic interactions.
Hierarchical clustering of the slow growth corrected genetic interaction profiles was then applied to unravel potential differences in observed genetic interactions patterns between GSTFs and K/P (Fig 2A–2C). Three striking differences emerge when comparing this clustering with the clustering of the original, untransformed data (S2 Fig). First, pairs are grouped into distinct clusters, whereas previously, most were grouped into a single large cluster. Second, a cluster of predominantly kinase/phosphatase pairs emerges (Fig 2A, green branch, labeled as 1). These contain mixtures of different genetic interaction patterns, corresponding to ‘mixed epistasis’ [16]. Third, a smaller cluster dominated by buffering appears (Fig 2A, red branch, labeled as 2). This cluster also has strong growth-based negative genetic interaction scores (Fig 2C), which are known to be associated with redundancy.
The ‘buffering’ cluster, with its strong growth-based negative interactions, mostly consists of pairs with a high sequence identity (average 43.7%) compared to the others (average 21%). These include Nhp6a-Nhp6b, Met31-Met32, Ecm22-Upc2 and Ark1-Prk1, for all of which redundancy relationships have been described previously [53–56]. The high sequence identity here indicates a homology-based redundancy, in which both genes can perform the same function [30,31,57,58]. The only exception here, is the kinase/phosphatase pair Elm1-Mih1. This pair may be explained through pathway-based redundancy where two parallel pathways can compensate for each other’s function [59]. Elm1 is a serine/threonine kinase, and Mih1 a tyrosine phosphatase, which are both involved in cell cycle control (S4 Fig, left panel) [60,61]. Mih1 directly regulates the cyclin-dependent kinase Cdc28, a master regulator of the G2/M transition [61]. Elm1, on the other hand, indirectly regulates Cdc28 activity by promoting Swe1 degradation through the recruitment of Hsl1 [62,63]. The timing of entry into mitosis is controlled by balancing the opposing activities of Swe1 and Mih1 on Cdc28, and both Swe1 and Mih1 are key in the checkpoint mediated G2 arrest [64,65]. Deletion of Elm1 does not result in many gene expression changes (Fig 1C) which can be explained through compensatory activity of Mih1 (S4 Fig, middle panel). Downregulation of Mih1 activity has also been suggested before as an effective mechanism to counter stabilization of Swe1, as neither stabilization of Swe1 or elimination of Mih1 in itself is sufficient to promote G2 delay, but simultaneous stabilization of Swe1 and elimination of Mih1 does cause G2 arrest [63]. Simultaneous deletion of Elm1 and Mih1 leads to higher levels of inactive Cdc28 causing a G2 delay and stress (S4 Fig, right panel) [63]. All pairs within this cluster can therefore be associated with a redundancy mechanism.
Taken together, these results suggest that the clustering of the slow growth corrected genetic interaction profiles is able to discern potential differences in mechanisms. Even though most pairs in the four clusters (Fig 2A) show negative genetic interactions (Fig 2C), different mechanisms are likely underlying each individual cluster.
Within the slow growth corrected genetic interaction profiles another interesting cluster stands out: the orange branch where five out of six pairs involve GSTFs which predominantly show the inversion pattern (Fig 2A, branch 3). This suggests that inversion may be strongly associated with a particular group of GSTFs, whereas this does not seem to be the case for kinases and phosphatases. The overall percentage of genes showing inversion is already much higher for GSTFs (28.6%) than for kinases/phosphatases (18.7%) (S2 Table). When investigating the GSTF pairs within the cluster, it is clear that these display an even higher percentage of inversion compared to kinases and phosphatases (Fig 2D; adjusted p-value = 0.00026) as well as compared to other GSTF pairs (Fig 2D; adjusted p-value = 0.0043). In order to determine whether inversion was specific to the set of GSTFs analyzed here, or part of a more general phenomenon common to GSTFs, we included both positive and negative genetic interactions between GSTF pairs, expanding the number of GSTF pairs to 44. Clustering of all 44 GSTF pairs (S5 Fig) also shows that a large fraction of the GSTF pairs contain many genes showing inversion, with most of the inversion dominated GSTF pairs still clustering together (S5 Fig, indicated with an asterisk). Note though, that because the 44 GSTF pairs include both positive and negative genetic interactions, the results are not directly comparable to the kinase/phosphatase pairs as these only include negative genetic interactions. Taken together, this indicates that not only is inversion more frequently associated with GSTFs compared to kinases and phosphatases, but one particular subset of GSTFs is also predominantly defined by inversion.
Unlike buffering, where redundancy is a likely mechanistic explanation, the underlying mechanism of inversion is still unknown [17]. The GSTF pairs within the inversion dominated cluster also do not share a common biological process, function, pathway or protein domain other than general transcription related processes and functions. To investigate potential mechanisms of inversion, an exhaustive exploration was initiated. Previously, Boolean modeling has been applied to exhaustively explore all mechanisms underlying two genetic interaction patterns for the Fus3-Kss1 kinase phosphatase pair [16]. However, to explore all potential mechanisms underlying inversion, a Boolean approach may not suffice as more subtle, quantitative effects, may be needed to obtain inversion. At the same time, any modeling approach must remain computationally feasible. For this purpose, a modeling approach based on Petri nets was devised to exhaustively evaluate all possible three and four node models but taking into account the possibility of quantitatively different effects (Fig 3, Methods). Interactions between nodes (edges) can be activating (positive) or inhibiting (negative). In order to incorporate quantitative differences, both strong and weak edges were used (Methods). Counting all possible combinations of different edges results in 152,587,890,625 possible edge weight matrices. To reduce the number of models, three conditions were imposed, as used previously [16]. In short, nodes contain no self-edges, the number of incoming edges on any node is limited to two and the model includes at least two edges from one of the regulators (R1, R2) to the downstream genes (G1, G2). Applying these requirements and filtering for mirror edge weight matrices results in 2,323,936 matrices. By including AND/OR logics the final number of models to be evaluated was 9,172,034 (Methods). Petri net simulations were then run and genetic interaction patterns determined for G1 and G2, analogous to what was done for the original data (Methods) (Fig 1A). Depending on the topology, Petri net models can be stochastic, in other words, they do not show the same behavior when simulated multiple times and therefore result in unstable models. Only 2.3% of the models were found to be unstable, i.e. showed inconsistent genetic interaction patterns for G1 and G2 across five simulation runs. Thus, stochasticity hardly influences the observation of genetic interaction patterns in our simulations (Fig 3). Nevertheless, unstable models were excluded from further analysis. In total, 168,987 models (1.8%) show inversion in either G1, G2, or both downstream nodes.
To investigate which potential regulatory patterns underlie the 168,987 models showing inversion, low complexity models with few edges were analyzed first. Two interesting observations can be made. First, although there are many high complexity models involving four nodes and many edges (up to eight), three nodes and three edges are sufficient to explain inversion (Fig 4A). Second, only two three-node models exist that show inversion (Fig 4A). These two models only differ in the strength of the inhibiting edge from R1 to R2. Both models involve inhibition of R2 through R1 and weak activation of G1 by R1 in combination with a strong activation of G1 by R2, i.e. a quantitative edge difference between the incoming edges of G1. Deletion of R1 in these two models results in activation of R2, and therefore upregulation of G1 due to a strong activating edge. Deletion of R2 however, will not result in any changes compared to WT as it is normally inhibited by R1. Deletion of both R1 and R2 will lead to downregulation of G1 as the weak activating edge from R1 to G1 is lost. Taken together, the analysis of the low complexity models indicates that a quantitative difference in interaction strength is required to explain inversion.
To investigate whether this requirement also holds for higher complexity models, all models containing two to eight edges were further analyzed. Inversion models were grouped by the number of edges (complexity) and then analyzed for their relative frequency of having a quantitative edge difference (Fig 4B, top left panel, note that the number of possible models grows exponentially with the number of edges). Almost all of these models show a quantitative edge difference, with only a very small fraction (1.3% overall) of models not having a quantitative edge difference. To exclude these results being based on a particular choice of edge weights (1 for weak and 5 for strong, or ‘1/5’ for short), we repeated the simulations with strong interactions represented by an edge weight of 9 (named ‘1/9’). Of the 168,987 models that show inversion in the ‘1/5’ simulations, 144 299 (85.4%) also show inversion in the ‘1/9’ simulations. Moreover, both of the three edge models (Fig 4A) also show inversion in the ‘1/9’ simulations. Finally, also in the 144,299 ‘1/9’ inversion models, only 1,696 (1.18%) have no quantitative edge difference. Except for masking, the other genetic interaction patterns show different behavior, indicating that the relative ratio of quantitative versus non-quantitative edges is not an inherent network property. Based on both the low complexity models as well as the high complexity models showing inversion, it is evident that a quantitative difference in interaction strength of two genes or pathways acting on a downstream gene is required to explain inversion.
With the exception of the two models discussed above, all other inversion models consist of four nodes with two regulator nodes and two downstream effector nodes. To better understand the interplay between all four nodes, besides the node displaying inversion (G1), the second downstream gene (G2) was also analyzed for the occurrence of different genetic interaction patterns (Fig 5A). Most G2 nodes tend to have no genetic interaction pattern (27%). The most common genetic interaction patterns are buffering (23%) and quantitative buffering (18%). These both are very alike in their genetic interaction pattern (Fig 1A) and only show slight differences in their quantitative behavior. They may therefore be considered as part of the same superclass of “buffering”. As can be expected, the buffering node is frequently positioned upstream of the inversion node (Fig 5B). The combination of inversion and buffering is also significantly overrepresented within inversion models when compared to all models (Table 1, p < 0.005). Taken together this shows that a quantitative difference in interaction strength of two genes or pathways acting on a downstream gene is frequently accompanied by an intermediate gene or pathway that displays buffering.
One gene pair within the inversion dominated GSTF cluster (Fig 2A, branch 3; Fig 6A) that largely consists of inversion is Gat1-Gln3. By combining the three node model derived from the Petri Net modelling (Fig 4A, left panel) with existing literature, a potential mechanistic explanation for the interaction between this pair can be obtained (Fig 6B). Both Gln3 and Gat1 are activators involved in regulating nitrogen catabolite repression (NCR) sensitive genes [66–68]. When cells are grown under nitrogen rich conditions, as was done here, Gat1 is repressed by Dal80 [67]. Dal80 in turn can be activated by Gln3 [67,69], which provides a plausible mechanism for the predicted inhibition edge between Gln3 and Gat1 (Fig 6B). The degree to which Gln3 and Gat1 influence downstream genes has also been reported to differentiate between individual genes [70], which is fully consistent with the quantitative edge difference as predicted in the model (Fig 6B). The set of inversion related genes (Fig 6A, gene set 1) is enriched for nuclear encoded mitochondrial respiratory genes compared to non-inversion related genes (Fig 6A, denoted with a dot, adjusted p-value 3.2x10-17). Previously, NCR has been linked with mitochondrial-to-nuclear signaling through the retrograde signaling pathway [71,72], although an alternative mitochondrial-to-nuclear signaling pathway, such as the intergenomic signaling pathway, may instead be involved [73]. Taken together, this suggests that Gat1 and Gln3 might differentially influence mitochondrial-to-nuclear signaling, although additional experiments would be needed to confirm this initial hypothesis.
Another interesting pair of genes within the GSTF cluster dominated by the inversion pattern (Fig 2A, branch 3) is Hac1-Rpn4. This pair displays a substantial amount of both inversion as well as buffering (Fig 7A) and lends itself well for testing some of the model predictions. Hac1 and Rpn4 are both involved in the processing of inappropriately folding proteins, either by activating genes of the unfolded protein response [74] (UPR, Hac1) or via the endoplasmic reticulum-associated degradation [75] (ERAD, Rpn4). Two genes that display inversion, Pdr5 and Pdr15, show stronger expression changes compared to the other genes in the same gene set (Fig 7A, gene set 1). Both Pdr5 and Pdr15 are multidrug transporters involved in the pleiotropic drug response [76]. Expression of these two genes is tightly regulated by Pdr1 and Pdr3 [77,78]. Pdr5 is also positively regulated by expression of Yap1, a basic leucine zipper transcription factor that is required for oxidative stress tolerance [79]. Of the three transcription factors Pdr1, Pdr3 and Yap1, only PDR3 shows a clear upregulation in the hac1Δ rpn4Δ double mutant and hardly any change in the respective single mutants (Fig 7B). This is consistent with the role of the intermediate buffering gene as derived from our Petri net modeling results. If Pdr3 acts as the intermediate buffering gene mediating the quantitative effect as predicted based on our model, it is also expected that deletion of PDR3 leads to a more severe downregulation of PDR5 and PDR15 expression levels when compared to expression levels of PDR5 and PDR15 in the rpn4Δ mutant. To test this prediction, mRNA expression changes of PDR5 and PDR15 where investigated in the pdr3Δ and rpn4Δ mutants. As expected, deletion of PDR3 results in a much stronger downregulation of PDR5 (adjusted p-value = 7.26x10-4) and PDR15 (adjusted p-value = 5.95x10-5) compared to deletion of RPN4 (Fig 7C), thereby confirming the model prediction. Taken together, these results provide a likely mechanistic explanation where Pdr3 acts as the intermediate buffering gene in regulating Pdr5 and Pdr15 (Fig 7D).
To investigate genetic interactions in a high-throughput manner, growth-based assays have frequently been deployed, resulting in the identification of an overwhelming number of both negative and positive genetic interactions [6,20–28]. Based on these surveys, several theoretical mechanisms have been proposed to explain genetic interactions [3,18,80,81]. More efforts, also using different types of assays, are however still needed to systematically and thoroughly investigate the underlying mechanisms. Alongside growth-based genetic interactions, genome-wide gene expression measurements have been applied to elucidate potential molecular mechanisms underlying genetic interactions [16,17,33–36]. Although more laborious, expression-based genetic interactions potentially allow for more in-depth characterization of the genetic interaction landscape. Here, we show that buffering is the most frequently occurring pattern underlying most negative genetic interactions. These are however to a large degree related to slow growing strains, hindering the investigation of the underlying mechanisms. By applying a slow growth transformation that removes a cell cycle associated gene expression signature, many such effects can be filtered out [51]. The transformation results in distinct clusters that can be more easily aligned with potential underlying mechanisms. Recent advances using Crispr-Cas9 single and double knock-down screens, followed by single cell RNA sequencing have also shown that results are greatly influenced by the cell-cycle phase in which different cells are found [35,82]. It is therefore essential for future studies on genetic interactions to incorporate methods that decompose such large confounding effects, as they greatly influence the ability to deduce mechanism.
To infer underlying mechanisms from the genetic interaction landscape as obtained from genome-wide gene expression measurements, systematic modeling approaches are warranted [3,18]. Various modeling techniques have been instrumental in understanding various aspects of experimental data (reviewed in [83]). Different modeling methods have different applications, depending on the question asked and available data types. To understand the underlying mechanisms for many genetic interactions, an approach is needed that is able to exhaustively explore the complete genetic interaction landscape while at the same time incorporating (semi-)quantitative values. Thus, the simulated gene expression levels are coarse-grained semi-quantitative representations of the actual expression levels and cannot be linearly translated to experimental output. Therefore, we here used Petri net models to exhaustively explore more than nine million models. Inversion, a pattern strongly associated with a group of GSTF pairs was investigated in more detail, resulting in the striking conclusion that a quantitative difference in interaction strength is needed to explain inversion, independent of the particular value of the edge strength parameter chosen in the model. The approach taken here, by combining slow growth corrected genome-wide gene expression measurements with the exhaustive semi-quantitative Petri-net modeling thus highlights the benefits of using such an approach to understand mechanisms of genetic interactions. Applying this approach to other types of genetic interactions or across many more genetic interaction pairs can help us in further characterizing mechanisms of genetic interactions and relating these to pathway organization and cellular states.
Previously, a mechanism termed “buffering by induced dependency” was proposed to explain parts of the genetic interaction patterns observed between Rpn4 and Hac1 (Fig 8, dotted inset) [17]. This mechanism links the endoplasmic reticulum-associated degradation (ERAD) by the proteasome (Rpn4) with the unfolded protein response (UPR, Hac1), two distinct processes dealing with misfolded and unfolded proteins. By combining the “buffering by induced dependency” mechanism with the model proposed for inversion here, most genetic interaction patterns observed for Rpn4 and Hac1 can be explained (Figs 7A and 8). The combined model introduces a third, compensatory process, the pleiotropic drug response (PDR; Fig 8, bottom light gray inset). Even though the exact relationship between ERAD, UPR and pleiotropic drug response is unclear, the interplay between UPR and drug export has been shown in mammalian cells [84]. In yeast, Pdr5 and Pdr15 have been implicated in cellular detoxification [78,85] and may also be required for cellular detoxification under normal growth conditions [85]. Both Pdr5 and Pdr15 have been reported to be regulated through Pdr1 and Yap1 [79,86], as well as through Rpn4 [87,88]. This is also confirmed here by downregulation of both Pdr1 and Yap1 as well as downregulation of their target genes Pdr5 and Pdr15 in rpn4Δ (Fig 7B and 7C). It is therefore likely that in the wildtype situation when Rpn4 is active, both ERAD and the PDR are functioning (Fig 8). Deletion of RPN4 leads to deactivation of the ERAD and PDR pathways and activation of the UPR through Hac1 (Fig 8, rpn4Δ dotted red line). Deletion of both RPN4 and HAC1 results in a major growth defect and accumulation of misfolded and unfolded proteins, most likely leading to a stronger activation of the PDR through Pdr3 compared to the wildtype situation (Fig 7B and 7C; Fig 8, hac1Δ rpn4Δ dotted red line) [77,78]. Taken together, this model thus provides a potential regulatory mechanism in which two redundant processes, each with slightly different efficacies, can be differentially regulated, or fine-tuned, through a third, compensatory process. The requirement to fine-tune slightly different efficacies of different cellular processes then also provides a potential explanation why inversion is observed more frequently for gene-specific transcription factors since these allow for more fine-grained control than protein kinases and phosphatases.
In conclusion, we have shown how exhaustive exploration of regulatory networks can be used to generate plausible hypothetical regulatory mechanisms underlying inversion. Almost all models showing inversion contain a quantitative difference in edge strengths, which suggests quantitative differences in regulating downstream target genes are important for the functioning of GSTFs. These hypothetical mechanisms have subsequently been tested against known and new experimental data. For GSTFs we show a validated example of Hac1-Rpn4 where differential regulation of gene expression is key to understanding the genetic interaction pattern inversion.
Two selection criteria were applied to select genetically interacting GSTF and kinase/phosphatase pairs. First, one of the mutants of each individual pair should show genome-wide gene expression measurements similar to wildtype (WT). DNA microarray data from Kemmeren et al [47] was used to determine whether a single deletion mutant is similar to WT. A deletion mutant is considered similar to WT when fewer than eight genes are changing significantly (adjusted p-value < 0.05, FC > 1.7) in the deletion mutant gene expression profile, as previously described [16]. Second, selected pairs should show a significant growth-based negative genetic interaction score. Growth-based genetic interaction scores for GSTF [28] and kinase/phosphate [26] pairs were converted to Z-scores. A negative Z-score significance of p < 0.05 after multiple testing correction was used as the significance threshold. Applying these selection criteria resulted in 11 GSTF pairs and 15 kinase/phosphatase pairs (S1 Table).
Genome-wide gene expression measurements of single and double mutant GSTF pairs were obtained from Sameith et al [17]. Genome-wide gene expression measurements of single and double mutant kinase/ phosphatase pairs were obtained from van Wageningen et al [16]. Genome-wide gene expression measurements of pdr3Δ and rpn4Δ were obtained from Kemmeren et al [47]. Statistical analysis of these gene expression profiles was performed as previously described [47]. In summary, mutants were grown in Synthetic Complete (SC) medium with 2% glucose and harvested during exponential growth. WT cultures were grown alongside mutants in parallel to monitor for day to day effects. For each mutant statistical analysis using limma was performed versus a collection of WTs [16,47]. Reported FC for each transcript is the average of four replicate expression profiles over a WT pools consisting of 200 WT strains.
Growth measurements for single and double mutant GSTF and kinase/phosphatase pairs were obtained from Sameith et al [17] and van Wageningen et al [16] respectively. Growth-based genetic interaction scores were calculated for both GSTF and kinase/phosphatase pairs as performed before [17]. In summary, the fitness W of single and double mutants was determined as the ratio between the WT growth rate and the mutant growth rate. The growth-based genetic interaction score ɛgrowth,XY was calculated as the deviation of the observed fitness in a double mutant from the expected fitness based on the respective single mutants (ɛgrowth,XY = WxΔyΔ - WxΔ. WyΔ). P-values were assigned to genetic interaction scores based on the mean and standard deviation of a generated background distribution [17]. P-values were corrected for multiple testing using Benjamini-Hochberg. Adjusted p-values lower than 0.05 were considered significant. Fitness values of all single and double mutants, as well as calculated genetic interaction scores can be found in S1 Table.
Expression-based genetic interaction scores were calculated for both GSTF and kinase/phosphatase pairs as described before [17]. In summary, the effect of a genetic interaction between two genes X and Y on gene i is calculated as the deviation between the observed expression change in the double mutant and the expected expression change based on the corresponding single mutants (εtxpn_i,XY = |Mi,xΔyΔ − (Mi,xΔ + Mi,yΔ)|). The overall genetic interaction score between X and Y is calculated as the sum all genes i for which εtxpn_i,XY > log2(1.5). All genetic interaction scores consisting of at least 10 genes were kept for further downstream analyses. Genes with similar gene expression changes were divided into the 6 different patterns (buffering, quantitative buffering, suppression, quantitative suppression, masking, inversion), as previously described [17] (Fig 1A).
Genetic interaction profiles for both classes of proteins were grouped together based on the number of occurrences of the six different patterns using hierarchical clustering. Average linkage was applied for the clustering. Identity of genes in each genetic interaction profile was disregarded.
Slow growth signature transformation of the gene expression profiles was performed as previously described [51]. In short, for each mutant, the correlation of its expression profile with the first principal component of 1,484 deletion strains [47] was removed, thus minimizing correlation with the relative growth rate. The transformation reduces correlation with the relative growth rate from 0.29 to 0.10 on average [51].
Exhaustive modeling of possible network topologies underlying the genetic interaction patterns was carried out by creating Petri net models consisting of four nodes, representing two regulator genes (R1 and R2) and two downstream genes (G1 and G2). With four nodes and directed edges, there are 42 = 16 possible edges, and 216 = 65536 possible edge weight matrices, which is a tractable number. However, each interaction can in addition be positive or negative, and weak or strong (and absent), leading to 516 = 1.5⋅1011 possible interaction graphs (edge weight matrices), which becomes intractable. Many of these models, however, will be irrelevant for the understanding the biological behavior of genetic interaction patterns of two genes. To exclude these types of models, the following conditions were applied: 1) No self-edges are allowed. 2) The number of incoming edges on any node must be limited to two. 3) At least two incoming edges from at least one of the regulators (upstream nodes) to the genes (downstream nodes). Applying these conditions reduces the number of relevant edge weight matrices to 9,287,616. Furthermore, most generated matrices have mirror counterparts, therefore only one of the matrices was included in downstream analyses. Applying this filtering step results in 2,323,936 matrices. Fig 3 gives an overview of the various filtering steps, and shows which representation of the models was relevant in different stages of the filtering. Edge weight matrices were generated in R, version 3.2.2 (the function expand.grid was used to generate all combinations of edges per row in a given matrix).
Regulatory effects of two potentially interacting genes (R1 and R2) on two downstream genes (G1 and G2) were simulated using a Petri net approach [42,89–91] to recapitulate genetic interaction patterns observed in the gene expression data.
In the Petri net notation, nodes in a given model are represented by places (denoted as circles). Tokens are denoted in the places and indicate the availability of the resource represented by the place. Interactions between nodes always go via a transition (denoted as squares), connected via directed arcs (drawn as arrows). An incoming arc to a transition can be either activating or inhibiting. The weight on arcs going to a transition is always fixed to 1. The weight on arcs going from a transition to a place depends on the edge weight between two nodes, 1 for weak and 5 for strong (Fig 3). To establish sensitivity of our results with respect to the particular edge weights chosen, we also performed simulations with an edge weight of 9 for the strong interaction.
For nodes with two incoming edges, one has to decide how these two inputs should be combined: does the transition require both inputs to be activated (AND logic), or can one or the other activate it (OR logic). To incorporate this, for each pair of incoming edges with the same weight, two Petri net models were generated: one using the AND logic, and one using the OR logic (Fig 3, bottom right panel). For two incoming edges with different weights only the Petri net model using the OR logic was generated. For cases with two incoming edges to a node with two different directions, activation and inhibition, inhibition dominates.
To simulate the regulatory effects of two upstream genes (R1 and R2), 200 tokens were provided to represent the mRNA resources for each regulator, except when one of the regulators has an incoming edge from the other regulator as shown in (S6A Fig). Each step in the simulation process comprises of firing all enabled transitions (maximal parallel execution) [92,93]. A transition is enabled to fire when resources (tokens) in the input place(s) match or exceed the weight(s) on the respective incoming arc(s) to the transition (S6B Fig). In total 50 consecutive transition firing steps were performed.
To incorporate deletion mutants in the simulation process, tokens were removed from corresponding regulators. To prevent accumulation of tokens in deleted regulators, each outgoing arc from a transition to the corresponding deleted places were also removed in simulated deletion strains. The number of tokens in G1 and G2 after 50 steps of firing transitions represent their expression levels. The final token levels are coarse-grained semi-quantitative representations of expression levels. Since they cannot be linearly translated to the experimental output, we compare the relative difference between single and double mutants and the WT situation where both R1 and R2 are active. To avoid division by zero one token was added to the total number of tokens in G1 and G1. These fold changes were then log2 transformed (M values).
Simulation-based genetic interaction scores for G1 and G2 were calculated based on the deviation between observed M values in the double mutant and the expected M value based on the single mutants, as follows: εsim,R1R2i = |MR1ΔR2Δi − (MR1Δi + MR2Δi)|, where i can be either G1 or G2. Each node with εsim,R1R2i > log2(1.7) was further divided into genetic interaction patterns, as defined before based on gene expression data [17]. Simulated expression levels for single and double mutants are considered to be increased relative to WT when M > log2(1.7) and decreased when M < -log2(1.7).
Functional enrichment analyses were performed using a hypergeometric testing procedure on Gene Ontology (GO) biological process (BP) annotations [67] obtained from the Saccharomyces Cerevisiae Database [68]. The background population of genes was set to 6,359 and p values were corrected for multiple testing using Bonferroni.
Models were visualized in R, version 3.2.2, using diagram package (version 1.6.3). Weak and strong activation/inhibition edges are represented as thin and thick lines, respectively.
|
10.1371/journal.ppat.1007107 | Methyl-CpG-binding (SmMBD2/3) and chromobox (SmCBX) proteins are required for neoblast proliferation and oviposition in the parasitic blood fluke Schistosoma mansoni | While schistosomiasis remains a significant health problem in low to middle income countries, it also represents a recently recognised threat to more economically-developed regions. Until a vaccine is developed, this neglected infectious disease is primarily controlled by praziquantel, a drug with a currently unknown mechanism of action. By further elucidating how Schistosoma molecular components cooperate to regulate parasite developmental processes, next generation targets will be identified. Here, we continue our studies on schistosome epigenetic participants and characterise the function of a DNA methylation reader, the Schistosoma mansoni methyl-CpG-binding domain protein (SmMBD2/3). Firstly, we demonstrate that SmMBD2/3 contains amino acid features essential for 5-methyl cytosine (5mC) binding and illustrate that adult schistosome nuclear extracts (females > males) contain this activity. We subsequently show that SmMBD2/3 translocates into nuclear compartments of transfected murine NIH-3T3 fibroblasts and recombinant SmMBD2/3 exhibits 5mC binding activity. Secondly, using a yeast-two hybrid (Y2H) screen, we show that SmMBD2/3 interacts with the chromo shadow domain (CSD) of an epigenetic adaptor, S. mansoni chromobox protein (SmCBX). Moreover, fluorescent in situ hybridisation (FISH) mediated co-localisation of Smmbd2/3 and Smcbx to mesenchymal cells as well as somatic- and reproductive- stem cells confirms the Y2H results and demonstrates that these interacting partners are ubiquitously expressed and found within both differentiated as well as proliferating cells. Finally, using RNA interference, we reveal that depletion of Smmbd2/3 or Smcbx in adult females leads to significant reductions (46–58%) in the number of proliferating somatic stem cells (PSCs or neoblasts) as well as in the quantity of in vitro laid eggs. Collectively, these results further expand upon the schistosome components involved in epigenetic processes and suggest that pharmacological inhibition of SmMBD2/3 and/or SmCBX biology could prove useful in the development of future schistosomiasis control strategies.
| Schistosomiasis, caused by infection with blood fluke worms, is responsible for chronic disability and debilitating pathology in millions of infected individuals living in deprived regions of the developing world. Currently, schistosomiasis is primarily controlled by administration of a single drug (praziquantel) with a currently unknown mechanism of action and an inability to prevent reinfection or kill juvenile blood flukes. Therefore, to limit the spread and lower the global prevalence of this neglected infectious disease, praziquantel replacement or alternative strategies are urgently needed. One such strategy is to identify molecular targets essential for schistosome biology and to characterise how their loss of function affects parasite developmental processes. By doing so, new drug targets or vaccine candidates can be progressed. Here, we extend our work on the characterisation of Schistosoma mansoni epigenetic processes and reveal that two interacting components (S. mansoni methyl-CpG-binding domain protein, SmMBD2/3 and S. mansoni chromobox protein, SmCBX) are instrumental for maintaining the proliferative capacity of the single most important parasite cell population—proliferating schistosome stem cells (or neoblasts). We additionally demonstrate that these two proteins are necessary for maintaining schistosome egg production, a lifecycle feature responsible for human pathology and disease transmission. Developing drugs that disrupt the interaction of these epigenetic participants or inhibit their activity could highlight a novel approach for controlling schistosomiasis.
| Characterised by a complex lifecycle alternating between two different hosts (snail and mammal) and a fresh water ecosystem, schistosomes are highly evolved human pathogens responsible for the neglected infectious disease schistosomiasis. Predominantly found in sub-tropical and tropical regions of resource-poor communities, schistosomiasis kills thousands of individuals per year and causes chronic disability in millions more [1]. Until a immunoprophylactic vaccine can be developed, existing treatment relies on chemotherapeutic administration of praziquantel (PZQ) to individuals living in endemic communities [2]. Use of a single anti-parasitic drug with a currently unknown mechanism of action (perhaps acting as a G-protein coupled receptor agonist [3]) and limited efficacy against juvenile schistosomes [4], however, raises serious concerns in meeting the ambitious targets set by the World Health Organisation for achieving schistosomiasis elimination in selected regions and countries by 2020 [5]. Therefore, furthering our understanding into how schistosomes respond to diverse environmental stimuli (water, snail or human) may simultaneously reveal the molecular processes essential for lifecycle transmission as well as the specific components suitable for next-generation anti-schistosomal drug or vaccine development. Epigenetic processes that shape chromatin modifications as well as regulate both heritable and environmentally influenced phenotypes present a rich molecular area in which to identify these key schistosome components [6].
While others have explored the role of histone modifying enzymes (HMEs) [7] in mammalian host infection, cercariae to schistosomula transformation, parasite viability, egg production and sexual differentiation of adults [8–13], we have investigated the epigenetic biology of the core schistosome DNA methylation machinery components DNA methyltransferase 2 (SmDNMT2) and methyl-CpG binding domain protein (SmMBD2/3) [14]. After first demonstrating that the co-regulated expression of Smdnmt2 and Smmbd2/3 throughout the schistosome lifecycle mirrored the abundance of DNA methylation, we provided important evidence that SmDNMT2 is, indeed, a functional DNA methyltransferase. In this previous study, RNA interference (RNAi) suppression of Smdnmt2 led to a significant decrease in global DNA methylation in adult schistosomes. Together with 5-azacytidine (5-AzaC) mediated inhibition of adult worm DNA methylation, egg production, embryo maturation and ovarian development, this functional genomics datum strongly suggested an important, but enigmatic [15], role for DNA methylation (and SmDNMT2 activity) in schistosome biology and oviposition. However, in depth functional analysis of the DNA methylation reader, SmMBD2/3, was not fully explored in our previous study [14]. As methyl-CpG binding domain (MBD) proteins importantly link DNA methylation to higher order chromatin structures [16], SmMBD2/3 characterisation could provide further insight into the downstream action of an intact DNA methylation machinery in this parasite.
Despite the diversity of MBD family members found within vertebrates (MBDs 1–6 and MeCP2), invertebrate genomes generally only contain one ancestral form, MBD2/3 [16, 17]. Within the invertebrates, the most comprehensive functional data for MBD2/3 proteins have been obtained from Drosophila melanogaster (dMBD2/3), largely due to its unusual mCpT/A-binding capability [18]. Non-CpG methylation has been associated particularly with DNMT2 activity [14, 15, 19], and so it appears that dMBD2/3 has evolved to adapt to these conditions in this DNMT2-only organism. Amongst the Platyhelminthes, only MBD2/3 from the planarian Schmidtea mediterranea (SmedMBD2/3) has been characterised to date [20]. An arginine to lysine mutation at position 17, which is known to interact directly with mCpGs [21], is likely to render SmedMBD2/3 incapable of binding to methylated DNA. Interestingly, Smedmbd2/3 expression was exclusively found in adult somatic stem cells (ASCs), called neoblasts, as well as germline cells, and was deemed essential for their differentiation during tissue regeneration and homeostasis. RNAi-mediated knockdown of Smedmbd2/3 resulted in a failure to correctly regenerate several organs, including the eyes, gut and pharynx. This observation, combined with a lack of detectable DNA methylation, suggests an entirely DNA methylation-independent role for SmedMBD2/3 in this non-parasitic platyhelminth species.
Due to its currently unknown role in schistosome molecular or epigenetic processes, we herein have conducted a thorough investigation of the first parasitic platyhelminth MBD2/3, SmMBD2/3. Using a combination of experimental approaches, we demonstrated that SmMBD2/3 is a nuclear localised, functional 5-methyl cytosine (5mC) binding protein capable of interacting with an epigenetic adaptor protein—S. mansoni chromobox protein (SmCBX). Functional genomics-led analyses of SmMBD2/3 and SmCBX have further indicated that both gene products are required for schistosome neoblast proliferation and oviposition. Together, our data provide additional roles for ancestral MBD2/3 function in DNMT2-only organisms and highlight the SmMBD2/3-SmCBX protein complex as a novel target for combating schistosomiasis.
All mouse procedures performed at Aberystwyth University (AU) adhered to the United Kingdom Home Office Animals (Scientific Procedures) Act of 1986 (project license PPL 40/3700) as well as the European Union Animals Directive 2010/63/EU and were approved by AU’s Animal Welfare and Ethical Review Body (AWERB). In adherence to the Animal Welfare Act and the Public Health Service Policy on Humane Care and Use of Laboratory Animals, all mouse procedures performed at UT Southwestern Medical Center were approved by the Institutional Animal Care and Use Committee (IACUC) of the UT Southwestern Medical Center (protocol approval number APN 2014–0072).
A Puerto Rican strain (NMRI) of Schistosoma mansoni was used throughout the study and passaged between Mus musculus (Tuck Ordinary; TO) and Biomphalaria glabrata (NMRI albino and pigmented hybrid [22]) hosts. Cercariae were shed from both B. glabrata strains by exposure to light in an artificially heated room (26 oC) for 1 hr and used to percutaneously infect M. musculus (200 cercariae/mouse) [23]. Adult schistosomes were obtained from M. musculus at 7 wks post-infection and used for RNA interference (RNAi), fluorescence in situ hybridization (FISH) and for the generation of nuclear protein extracts.
Comparison of methyl-CpG binding domains (MBDs, PF01429) from multiple MBD family members was performed by multiple sequence alignment, generated using MUSCLE v3.8 [24]. The NCBI accession numbers of these MBD family members comprise: Apis mellifera MBD1, XP_003250634.1; Homo sapiens MBD1, NP_002375.1; M. musculus MBD1, NP_038622.2; H. sapiens MeCP2, NP_001104262.1; M. musculus MeCP2, NP_001075448.1; Xenopus laevis MeCP2, NP_001081854.1; H. sapiens MBD4, NP_001263201.1; M. musculus MBD4, NP_034904.2; H. sapiens MBD2, NP_003918.1; M. musculus MBD2, NP034903.2; Gallus gallus MBD2, NP_001012403.1; H. sapiens MBD3, NP_001268382.1; M. musculus MBD3, NP_038623.1; X. laevis MBD3, AAD55389.1; Bombyx mori MBD2/3, XP_004929675.1; Ancylolomia pectinifera MBD2/3, ACF05483.1; Hemicentrotus pulcherrimus MBD2/3, ACF05485.1; S. mansoni MBD2/3, CCD59176.1; Drosophila melanogaster MBD2/3, NP_731370.1. For Acheta domesticus MBD2/3, no NCBI accession number could be located; therefore, the amino acid sequence for this family member was retrieved from Uno et al. [25]. The sequences were inspected for homology and the presence of conserved and semi-conserved residues. Residues that directly bind to the DNA backbone or 5mC decorated dinucleotides in the published solution structures [21, 26] were manually annotated, along with residues previously shown to result in reduced 5mC-binding after mutagenesis.
The three-dimensional structure of the MBD within SmMBD2/3 (GenBank ID: AEK05283.1) was derived by homology modelling using M4T [27, 28]. The MBD template selected for SmMBD2/3 modelling was the three-dimensional structure of G. gallus MBD2 (Protein Databank [29] identification code: 2KY8) [26]. The sequence identity between SmMBD2/3 and the chicken MBD2 template was 42% with a sequence coverage over 90%, hence well within the acceptable range for comparative modelling techniques [30]. The quality and stereochemistry of the SmMBD2/3-5mC model was assessed using Prosa-II [31] and PROCHECK [32] respectively.
Full-length cDNAs encoding SmMBD2/3 (HM991455.1) and red fluorescent protein (RFP) were cloned into the pkFLAG vector and expressed in NIH-3T3 cells (ECACC number 86041101) with a C-terminal FLAG tag (SmMBD2/3-FLAG and RFP-FLAG). Briefly, NIH-3T3 fibroblasts were cultured at 37°C in a 5% CO2 environment in DMEM supplemented with penicillin (100 U/ml), streptomycin (100 μg/ml), 10% v/v new born calf serum (Sigma Aldrich) and 2 mM L-glutamine. At 24 hr prior to transfection, 50,000 cells were seeded per chamber slide well (1.7cm2) to give 70–90% confluency the next day. Cells were transfected with either 1 μg of pkFLAG-RFP or pkFLAG-SmMBD2/3 per well of a chamber slide using 2 μl of Turbofect transfection reagent (Thermo Scientific) according to the manufacturer’s instructions. For negative transfection controls, an equal volume of water was used in replacement of plasmid. Transfections proceeded for 24 hr before preparation of slides for microscopy.
For SmMBD2/3-FLAG immunolocalisation, cells were fixed in 5% v/v formaldehyde solution (5 min), washed 3 X 5 min with PBS, permeabilised with 0.15% v/v Triton X-100 (2 min) and washed again with PBS (3 X 5 min). Slides were then incubated for 1 hr with a 1:400 dilution (PBS/1% BSA) of anti-FLAG primary antibody (M2 clone, F3165, raised in mice, Sigma Aldrich) and washed once in PBS/1% BSA for 5 min before secondary antibody incubation. Here, Alexa Fluor 488 (F(ab’)2 fragment, anti-mouse IgG (H + L), raised in goat) conjugated Abs (1:200 dilution in PBS/10% v/v new born calf serum) were added to the slides and incubated for 1 hr. Slides were finally washed 3 X 5 min in PBS. For ‘water’ transfected cells (negative control), staining was also performed in this manner. For RFP-FLAG transfected cells, only the initial formaldehyde fixation step and PBS wash was required.
Prior to visualisation, cells were immersed in mounting solution (Vectashield, Vector Labs) supplemented with DAPI (4',6-diamidino-2-phenylindole, 1.5 μg/ml). Fluorescent microscopic images were captured on a Leica TCS SP5II laser scanning confocal microscope (LSCM) fitted with a 63 X (oil immersion) objective (2.3 X zoom) using the Leica LAS-AF software. Green (SmMBD2/3-FLAG) and red (RFP-FLAG) fluorescence were visualised with an argon or diode-pumped, solid state (DPSS) laser at 488 nm and 561 nm, respectively. DAPI was visualised using a 405 nm blue diode laser. Obtained images were deconvolved using AutoQuant X2 (Media Cybernetics) and analysed in Imaris 7.3 (Bitplane).
ImageJ was used for quantitative analysis of immunofluorescence localisation of SmMBD2/3-FLAG and RFP-FLAG in NIH-3T3 cells. The microscopic fields of view selected for image analysis contained a total of approximately 50 cells (~25% of which were transfected). Every transfected cell within each image was used in the analysis, counting top to bottom, to ensure no subjective bias could influence the selection. A total of 40 transfected cells were analysed across three images for both pkFLAG-SmMBD2/3 and pkFLAG-RFP constructs. The total pixel intensity (derived from green or red fluorescence produced by SmMBD2/3-FLAG or RFP-FLAG, respectively) was measured for manually annotated nuclear and whole cell regions in ImageJ. Nuclear regions were annotated according to DAPI staining. No background subtraction of values was required, because pixel intensity readings for the surrounding area (verified individually for each image) were zero. Nuclear fluorescence readings were normalised as a percentage of the total cell fluorescence to account for variation in whole image fluorescence intensity, varying cell size and variable levels of protein expression within each cell. Nuclear and cytosolic fluorescence values were analysed and compared by one-way ANOVA and Tukey’s Honest Significant Difference (HSD) test to confirm statistical significance.
Full-length SmMBD2/3 (HM991455.1) was cloned into the pET30a vector (Novagen, UK) and expressed in One Shot BL21 (DE3) E. coli competent cells (Invitrogen, UK) to contain a C-terminal poly-histidine tag (His6) as previously described for S. mansoni venom allergen like 9 (SmVAL9) [33]. Briefly, isopropyl β-D-1-thiogalactopyranoside (IPTG) induced bacterial cell pellets were resuspended in 15 ml of lysis buffer (50 mM NaH2PO4, 300 mM NaCl, 10 mM imidazole (IMDZ) + protease inhibitors (cOmplete, mini, EDTA-free tablets, Roche)) and lysed in a Cell Disruption System (Constant Systems) at 30,000 Psi. The lysates were centrifuged at 21,000 x g, 4°C for 20 min to yield the soluble protein fraction. The soluble protein fraction was passaged 3 x over a column containing 500 μl Ni-NTA agarose beads (Qiagen). Purification of recombinant (r) SmMBD2/3 was achieved using wash buffers (50 mM NaH2PO4, 300 mM NaCl, + protease inhibitors (cOmplete, mini, EDTA-free tablets, Roche)) containing increasing concentrations of IMDZ. An initial 40 mM IMDZ wash buffer was used (30 ml per 800 ml culture) followed by a 100 mM IMDZ wash (10 ml). rSmMBD2/3 elution was achieved using 10 ml wash buffer containing 250 mM IMDZ. For the “un-induced” negative control protein sample, bacteria were processed identically, except the IPTG induction step was omitted. The resulting cell pellets were subjected to the purification scheme used for IPTG-induced rSmMBD2/3. This process produced a soluble protein fraction that was enriched for the Ni-NTA co-purifying E. coli products present in the purified rSmMBD2/3 sample. Western blot analysis of purified rSmMBD2/3 and “un-induced’ Ni-NTA co-purifying E. coli products was performed essentially as described [33]. The HisProbe-Horseradish peroxidase (HRP) conjugate (ThermoScientific) used to detect rSmMBD2/3- His6 was used at a 1:4000 dilution.
Amino acid sequence evaluation of recombinant SmMBD2/3 (rSmMBD2/3) was confirmed by matrix assisted laser desorption ionisation time of flight (MALDI-TOF) mass spectrometry. The protein band corresponding to rSmMBD2/3-His6 (36.5 kDa) was excised from Coomassie blue stained polyacrylamide gels and subjected to in-gel trypsin digests followed by MALDI-TOF mass spectrometry at the Leiden University Medical Center (LUMC) as previously described [34].
Nuclear protein extracts from adult male and female schistosomes (~80 parasites/gender) were extracted with the EpiQuik Nuclear Extraction Kit I (Epigentek) according to the manufacturer’s instructions (for tissues). The same technique was used for extraction of nuclear proteins from NIH-3T3 cells. NIH-3T3 cells were cultured to confluency in T75 flasks under the conditions described above, and nuclear protein extraction was performed according to EpiQuik’s instruction for monolayer and adherent cells.
Induced and “un-induced” rSmMBD2/3-His6 soluble protein fractions were dialysed to exchange IMDZ wash buffer for TBS (Tris-buffered saline; 50 mM Tris-HCl, 150 mM NaCl, pH 7.6) prior to assessment of 5mC binding. The 5mC binding activities of induced and “un-induced” rSmMBD2/3-His6, nuclear extracts of 7 wk adult male and female parasites and nuclear extracts of NIH-3T3 cells were quantified using the EpiQuik MBD2 Binding Activity Assay Kit (Epigentek) according to the manufacturer’s instruction. Colorimetric readouts of 5mC binding (in the CpG context) were measured by a BMG Labtech Polarstar Omega plate reader. For induced and “un-induced” SmMBD2/3, 5mC binding of 1 μg of soluble protein sample was used (1 μg bovine serum albumin, BSA, was additionally used as a negative control). For nuclear protein extracts, 10 μg of soluble protein was used per sample. Input sample volumes were made up to a total of 3 μl using TBS. As a measure of background, blank samples containing TBS only were used, and this colorimetric reading was subtracted from all other samples. All conditions were set up in triplicate, and results are representative of three replicates derived from single protein samples. For statistical analysis, one-way ANOVA and Tukey’s HSD test were used.
A yeast two-hybrid (Y2H) library was synthesised using the BD Matchmaker Library Construction and Screening kit (Clontech, UK) according to manufacturer instructions. Double-stranded cDNA for the library was made from 2 μg of mixed sex, adult S. mansoni RNA (isolated from parasites using TRIzol, Invitrogen). The library was constructed by transforming competent Saccharomyces cerevisiae (strain AH109) with double-stranded cDNA and the pGADT7-Rec plasmid and transformants were selected on SD/-Leu plates, harvested and stored in 1 ml aliquots at -80°C. The full length SmMBD2/3 coding sequence (GenBank: HM991455.1) was cloned into the Gal4-BD fusion vector, pGBKT7, and expressed within S. cerevisiae strain Y187. Production of the cDNA library, transformations, toxicity tests, auto-activation tests, mating procedures and the Y2H screen were performed according to the Matchmaker Library Construction & Screening Kits User Manual (Clontech, UK). Toxicity and auto-activation tests of expressed SmMBD2/3 Gal4-AD fusion proteins were performed in both AH109 and Y187 strains and found to be negative. The Y2H screen was performed on SD plates lacking Tryptophan, Leucine, Histidine and Adenosine plus X-α-gal (4 μg/ml) (SD SD/-Trp/-Leu/-His/-Ade + X-α-gal). Positive colonies were subsequently screened for LacZ reporter gene activity using the X-β-gal filter assay (Matchmaker Library Construction & Screening Kits User Manual, Clontech, UK). X-β-gal filter assay colonies were visually categorised according to the intensity of blue, and those with the highest intensity had cDNA library Gal4-AD fusion interacting partners identified by PCR, as described in the Matchmaker Library Construction & Screening Kits User Manual (Clontech, UK). Each in-frame, identified interacting partner was retested by co-transformation into the Y2HGold strain (Clontech, UK) and plating on SD/-Trp/-Leu/-His/-Ade + X-α-gal with additional Aureobasidin A (60 ng/ml). Positive (p53 + LgT) and negative controls (LamC + LgT) were also produced and used as references in the screen.
The pellet X-β-gal (PXG) assay, as described by Möckli et al. [35], was used to quantify SmMBD2/3 + SmCBX interactions. Appropriate negative (pGBKT7 + SmCBX-Gal4-AD, SmMBD2/3-Gal4-BD + pGADT7, pGBKT7 + pGADT7, LgT + LamC) and positive (LgT + p53) controls were produced in the Y2HGold strain (Clontech, UK) and assayed alongside SmMBD2/3 + SmCBX samples. Three colonies of each Y2HGold strain produced were assayed per interacting partner. The plate was scanned using a GS-800 calibrated densitometer and Quantity One (v4.6) software (Biorad, UK). The pixel intensity of each well was quantified using ImageJ. The average pixel intensity value of the LgT + LamC negative controls was used to blank all other samples. Pixel intensities were expressed as a percentage of the average positive control LgT + p53 value. One-way ANOVA and Tukey’s HSD test were used for statistical analysis.
Data from the 37,632 element S. mansoni long-oligonucleotide DNA microarray studies of Fitzpatrick et al. [36] was interrogated to find the expression profile of Smcbx across 15 different lifecycle stages. Raw and normalised fluorescent intensity values are available via Array Express under the experimental accession number E-MEXP-2094.
Parasite fixation, permeabilisation, and whole mount fluorescence in situ hybridisation (FISH) were performed as previously described [37]. To detect hybridisation signals, Tyramide Signal Amplification was employed using methods previously described [38].
Following the perfusion of 7-week infected mice, adult worms were recovered and RNAi performed as previously described [14]. Smmbd2/3, Smcbx and non-specific Luciferase (Luc) siRNA duplexes were purchased from Sigma (siRNA sequences defined in S1 Table). Briefly, 10 adult females or 5 worm pairs were transferred to 4mm electroporation cuvettes containing DMEM (5.4 g/L D-Glucose, Sigma) supplemented with 2 mM L-glutamine, 10,000 Units/ml penicillin and 10,000 μg/ml streptomycin. siRNA duplexes (5 μg) were subsequently added and worms were electroporated with a single pulse at 125V for 20ms using a ECM-830 Square Wave Porator (BTX). For double knockdowns, 5 μg of each Smmbd2/3 and Smcbx siRNA duplex was used and compared to 10μg of siLuc duplexes. Mixed sex adult worms (for knockdown assessment by quantitative reverse transcription PCR, qRT-PCR) and adult females (for stem cell quantification) were cultured at 37°C in DMEM (5.4 g/L D-Glucose, Sigma) supplemented with 10% fetal calf serum, 2 mM L-glutamine, 10,000 Units/ml penicillin and 10,000 μg/ml streptomycin in an atmosphere of 5% CO2 with a 70% media exchange performed every 24 hr.
Following RNAi with siSmcbx, siSmmbd2/3 and siLuc, mixed-sex adult worms were incubated for a total of 48 hr before processing them for RNA isolation. Briefly, worms were homogenised using a TissueLyser LT (Qiagen, UK) in TRIzol Reagent (Invitrogen, UK) before isolation of total RNA using the Direct-zol RNA Kit (Epigentek, UK). cDNA was then generated, qRT-PCR performed and data analysed as previously described [36]. qRT-PCR primers are defined in S1 Table.
In vitro 5’-ethynyl-2’-deoxyuridine (EdU) labelling was performed as previously described [37]. Briefly, RNAi-manipulated adult females were cultured for six days and pulsed with 10 μM EdU for 24 hr at day six. On day seven, female schistosomes were fixed, stained and prepared for laser scanning confocal microscopy (LSCM) imaging. Anterior regions and ovaries were imaged and used to determine the relative number of EdU-labelled nuclei for treatment as well as control groups. For quantification, LSCM images were acquired using a Leica TCS SP5II confocal microscope and a 40X lens (NA 1.25), accruing a total of 15 sections for each Z-stack. For each Z-stack, the fluorescent intensity of the DAPI and EdU channels were used to calculate the total volume (μm3) occupied by each fluorophore using the Surface tool in Imaris v8.2 (Bitplane). The percentage of EdU positive nuclei was calculated by dividing the volume of the EdU channel by the volume of the DAPI channel. To investigate significant differences between the siRNA treatments, a one-way ANOVA followed by Tukey HSD test was performed.
Our previous studies suggested that SmMBD2/3 and some, but not all, related platyhelminth MBD2/3 homologs contained structural features critical for 5mC binding and diagnostic for this family [6, 14, 39]. This is also true for other representative eukaryote MBDs (Fig 1).
While these previously described features included a methyl-CpG binding domain (PF01429) and a C-terminal domain of methyl-CpG binding protein 2 and 3 (PF140489), our current examination (using cNLS mapper [40] and WormBase-Parasite [41]) of SmMBD2/3 has additionally revealed the presence of two putative atypical bipartite nuclear localisation signals (13QTKRSSYANYGKQPQNSMSGQQPHHHQQ40, 271PMIKTFIVTDDDIRRQEARVKELRKKLEIA300) and a coiled-coil domain (283 IRRQEARVKELRKKLEIARKK303) (Fig 1A). Further sequence analyses of the methyl-CpG-binding domain (PF01429) within SmMBD2/3 and other MBD proteins highlighted the molecular basis for a proposed difference in functional activity. While variation was found in the conservation of SmMBD2/3 residues likely to be important in DNA binding (Fig 1B, ‘:’), the amino acid residues necessary for 5mC interactions (Table 1) were generally well conserved (9 out of 12 residues being identical, ‘*’ in Fig 1B).
Amongst the differences, semi-conservative (S45N) and non-conservative (S27G) substitutions were observed at 2 out of 3 residues in SmMBD2/3. Of particular importance, however, is the finding that SmMBD2/3 retains K30 and Y34. Together with R22 and R44, K30 and Y34 form a tetra-amino acid archetypal binding pocket well-conserved in all MBD binding proteins shown to interact with 5mC (Fig 1C and Table 1) and suggested a functional role for SmMBD2/3 in the nuclei of schistosome cells.
To determine if SmMBD2/3 is capable of translocating to nuclear compartments, transient transfection of a SmMBD2/3-FLAG tagged construct into a surrogate NIH-3T3 M. musculus fibroblast system was performed and SmMBD2/3 nuclear versus cytoplasmic localisation was quantified (Fig 2).
Here, SmMBD2/3 was found predominantly localised to nuclear compartments of transfected cells in contrast to the more evenly distributed nuclear and cytoplasmic localisation of cells transfected with RFP (representative images, Fig 2A). Quantification of these experiments demonstrated that ~80% of SmMBD2/3 transfected cells contained nuclear-dominated, as opposed to, cytoplasmic-dominated localisation (Fig 2B).
Nuclear-dominated localisation of SmMBD2/3 in transient transfected NIH-3T3 cells, as well as our previous description of Smmbd2/3 biased expression in females (vs males) [14], prompted us to investigate 5mC binding activities in nuclear extracts derived from adult male and female schistosomes (Fig 3).
In both nuclear samples derived from schistosome adults, 5mC binding activity was above background levels (BSA; negative control) and was comparable to that measured in nuclear extracts derived from NIH-3T3 cells (positive control) (Fig 3A). Consistent with Smmbd2/3’s female biased expression in adult schistosomes [14], significantly greater 5mC binding activity was detected in female compared to male nuclear extracts (p < 0.01). Recombinant expression of SmMBD2/3 (rSmMBD2/3) in E. coli cells followed by Ni2+-NTA purification (Fig 3B) allowed us to directly validate the 5mC binding activity (CpG context) of this nuclear protein. When compared to un-induced rSmMBD2/3 or BSA control samples, purified rSmMBD2/3 demonstrated significantly greater 5mC binding (p < 0.05) confirming its role as a functional methyl-CpG-binding protein. rSmMBD2/3 binding to non-methylated DNA targets or to methylated cytosines in diverse nucleotide contexts (i.e. CpA, CpT or CpC) was not assessed.
SmMBD2/3’s nuclear localisation in a heterologous transfection system and rSmMBD2/3’s binding to 5mC provided further evidence for a functional DNA methylation machinery operating within schistosome parasites [6, 14, 39]. As MBD proteins are recognised ‘readers’ of DNA methyltransferase enzymatic ‘writers’, they importantly serve as an epigenetic bridge between DNA and proteins involved in the formation and regulation of diverse chromatin states [16]. Thus, using Y2H screening of adult schistosome cDNA libraries, we subsequently investigated whether SmMBD2/3 interacted with other known epigenetic regulators or adaptors (Fig 4).
The single most abundant SmMBD2/3-interacting protein identified in these Y2H assays was a C-terminal, in-frame, truncation of Smp_179650 (Fig 4A); this truncation (SmCBX: Δ1–160) was identified five independent times (or 36% of all hits; Y2H results summarised in S2 Table). Full-length Smp_179650 encodes a protein with sequence similarity to epigenetic adaptor heterochromatin- (HP) or chromobox- (CBX) proteins [51, 52] and contains an N-terminal chromodomain (CD; PFAM PF00385, AAs 18–67), a monopartite nuclear localisation signal (NLS; AAs 109–119) and a C-terminal chromo shadow domain (CSD; PFAM PF01393, AAs 167–218). Chromobox CDs are responsible for binding to tri-methylated lysine (K) 9 of histone 3 (H3K9me3) [53, 54] and full-length SmCBX contains all critical AAs within this domain necessary for ‘reading’ H3K9me3 modifications (S1 Fig). Chromobox CSDs are necessary for initiating and maintaining protein-protein interactions (PPIs) [55]; the presence of the CSD within SmCBX (Δ1–160) likely explains all five, Y2H-detected, SmMBD2/3 interactions.
To quantify the interactive strength of CSD-containing SmCBX (Δ1–160) to SmMBD2/3, a modified X-β-gal based assay [35] was performed (Fig 4B). Here, binding of SmMBD2/3 and SmCBX (Δ1–160) was significantly higher than the empty vector negative control (pGBKT7+pGADT7; p<0.01). Similar to the negative controls, undetectable reporter expression was observed in yeast cells transfected with either SmMBD2/3 or SmCBX (Δ1–160) constructs alone and confirmed that auto-activation was not responsible for the specific SmMBD2/3-SmCBX (Δ1–160) interaction detected in the Y2H assays. Lifecycle expression profiling demonstrated that Smcbx was abundantly transcribed in all schistosome developmental forms analysed (Fig 4C). Together, these data suggest that SmMBD2/3 interacts with the highly abundant epigenetic adaptor protein SmCBX and this interaction is stably maintained by SmCBX’s CSD.
To provide supportive evidence for the Y2H-identified, SmMBD2/3-SmCBX interactions, localisation of both Smmbd2/3 and Smcbx transcripts in adult schistosomes was explored by fluorescence in situ hybridisation (FISH) (Fig 5).
Here, both Smmbd2/3 and Smcbx transcripts were found widely distributed throughout schistosome mesenchymal tissues (Fig 5A). In many (if not all) mesenchymal cells, Smmbd2/3 and Smcbx were spatially co-expressed (Fig 5A, upper row; white boxes). Interestingly, both Smmbd2/3 and Smcbx were also found in a sub-population of mesenchymal cells co-expressing Smhistone H2B (Fig 5A, lower two rows; white boxes), a known marker for proliferating neoblasts in adult parasites [37]. Amongst the reproductive tissues, Smmbd2/3—Smhistone H2B and Smcbx—Smhistone H2B co-localisation was also broadly expressed in many cells of the male (Fig 5B) and female (Fig 5C) gonads. Supporting the Y2H PPI results (Fig 4), these FISH data provided complementary evidence for SmMBD2/3 and SmCBX interactions in adult schistosomes and demonstrated that both genes were expressed within proliferating (H2B+) and differentiated (H2B-) cells.
RNAi was subsequently used to investigate the function of Smmbd2/3 and Smcbx in adult worms (Fig 6).
Here, siRNAs targeting either Smmbd2/3 or Smcbx in adult worm pairs led to a greater than 50% reduction in transcript abundance when compared to control worms (57% for siSmcbx treated worm pairs, 54% for siSmmbd2/3 treated worm pairs) (Fig 6A). Together, these data confirmed that RNAi could reduce the pools of Smmbd2/3 and Smcbx in adult schistosomes.
As our FISH results showed co-localisation of Smmbd2/3 and Smcbx to schistosome neoblasts and reproductive tissues (Fig 5), we next investigated whether either Smmbd2/3 or Smcbx knockdown could affect aspects of schistosome stem cell biology. Adult females were chosen for these experiments due to greater Smmbd2/3 expression [14] and 5mC binding (Fig 3) found in this gender compared to males. In either Smmbd2/3 or Smcbx knockdown conditions, adult females contained noticeably fewer EdU+ somatic cells (58% and 46% less, respectively, to control siLuc treated worms) throughout their bodies compared to controls (representative anterior regions; Fig 6B). In contrast, ovarian stem cell proliferation was not significantly affected by Smmbd2/3 or Smcbx knockdown. Females treated with siRNAs targeting both Smmbd2/3 and Smcbx (double knock-down) showed a similar neoblast deficiency phenotype (60% less EdU+ somatic cells compared to control siLuc treated worms, S2 Fig).
As a defect in neoblast proliferation was observed in adult females treated with either Smmbd2/3 or Smcbx siRNAs, other gross phenotypic abnormalities were additionally sought in these in vitro manipulated schistosomes (Fig 7).
Here, despite an incomplete reduction in Smcbx or Smmbd2/3 transcript levels (Fig 6A), a significant decrease in the number of normal (oval, containing a lateral spine with regular surface autofluorescence) schistosome eggs was consistently observed in siRNA treated parasites compared to siLuc controls (Fig 7A). This decrease in oviposition of normal eggs was also associated with the increased production of abnormal eggs (Fig 7B). Noticeable phenotypes observed in both treatment (siSmcbx and siSmmbd2/3) conditions included exemplars without lateral spines, individuals demonstrating reduced egg volumes and entities containing irregular autofluorescence (Fig 7C). Regardless of siRNA treatment, and despite these gross morphological differences, vitellocytes (DAPI+ cells) were present in all eggs examined (Fig 7C).
Schistosome development is influenced by interactions with three distinct niches (freshwater ecosystem, snail intermediate host and mammal definitive host) and is molecularly controlled by genetic as well as epigenetic processes [6, 36]. While schistosomes do not harbour the extreme developmental plasticity potential exhibited by nematodes [56], their ability to remain responsive to diverse environmental signals assists in the establishment of heritable variations critical for infection success [57]. Therefore, elucidating how schistosome epigenetic components cooperatively regulate key parasitological processes and shape heritable traits will likely uncover new targets for schistosomiasis control. Here, we provide evidence for the role of SmMBD2/3 and SmCBX in the biology of schistosome somatic stem cells (neoblasts) and additionally suggest that pharmacological disruption of these interacting partners will lead to defects in the most important aspect of schistosome mediated pathology and lifecycle maintenance, egg production.
While not every eukaryotic MBD binds to methylated DNA (reviewed in [58]), our data indicate that SmMBD2/3 contains the necessary features responsible for 5mC recognition (Fig 1), nuclear localisation (Fig 2) and functional 5mC binding (Fig 3). These particular findings are in contrast to those obtained from a detailed study of the only other platyhelminth MBD protein characterised to date, SmedMBD2/3 [20]. In this previous study, Jaber-Hijazi et al. demonstrated that SmedMBD2/3’s function in planarian tissue homeostasis was independent of 5mC binding [20]. The most likely explanation for these differential results between related platyhelminth MBD homologs is amino acid substitutions of critical 5mC-binding residues in the SmedMBD2/3 MBD domain (also observed in both HsMBD3 and MmMBD3, Fig 1 and Table 1), which are well-conserved in SmMBD2/3 [39]. MBD sequence divergence, along with variable levels of detectable DNA methylation (detectable levels found in S. mansoni [14, 15], undetectable levels found in S. mediterranea [20]), strongly suggests that the core platyhelminth DNA methylation machinery (MBDs and DNMTs) is diversely utilised within this group of animals and may be involved in other nuclear functions in addition to or in replacement of ‘reading’ DNA methylation marks [14] (Fig 3). For example, human MBD1 can bind to unmethylated cytosines via a CxxC domain as well as 5mC via its MBD [59]. Similarly, human MBD4 has an additional 5mC binding function, DNA repair, and this particular activity is facilitated by the presence of a DNA glycosylase domain [60]. Finally, while HsMBD5 and HsMBD6 both contain a MBD, they do not bind to 5mC; the presence of PWWP domains (present in MBD5) and P-rich domains (located in both MBD5 and MBD6) likely defines their role in other biological activities [61]. While none of these motifs are present in SmMBD2/3, a C-terminal coiled-coil region is clearly identifiable (Fig 1C). As Tatematsu et al., demonstrated that coiled-coil regions are essential for homo-dimerisation of HsMBD2 [62], the presence of this C-terminal domain within SmMBD2/3 is likely responsible for PPIs important for regulating higher-order chromatin structure in schistosome nuclei. Evidence to support this contention was derived from our Y2H studies confirming that SmMBD2/3 specifically interacts with an epigenetic adaptor protein SmCBX (Fig 4).
Chromobox (CBX) proteins (also known as heterochromatin protein 1; HP1) are non-histone, chromatin-interacting proteins involved in the regulation of heterochromatin, transcription and development [52]. Previous studies have demonstrated that both HsMBD1 and HsMeCP2 interact with HsHP1; the consequence of these interactions results in heterochromatin formation and transcriptional repression [51, 63]. Our results, therefore, are in line with these previous reports and illustrate that schistosomes maintain this conserved molecular interaction (Fig 4). As SmCBX’s CD contains all of the features necessary for H3K9me3 (a transcriptionally repressive histone mark; [64]) interactions (S1 Fig) and SmCBX’s CSD is associated with SmMBD2/3 binding (Fig 4A and 4B), this protein complex (along with other, yet to be identified proteins) is well-positioned to link the epigenetic processes of schistosome DNA methylation and post-translational histone modifications. Therefore, this data provides the first evidence within the Platyhelminthes that epigenetic cross-talk can occur and may have particular relevance in the context of schistosome chromatin structure, genome function and phenotypic manifestations. Further investigations exploring how SmMBD2/3-SmCBX interactions shape or modulate parasite developmental processes or transcriptional regulation could reveal novel (epigenetically-directed) strategies for anti-schistosomal control.
As a first step towards this goal, we investigated the spatial distribution of both Smmbd2/3 and Smcbx within adult schistosomes and found co-localisation to mesenchymal cells (histone H2B-) as well as histone H2B+ germ line cells and somatic neoblasts (Fig 5). While the function of the SmMBD2/3-SmCBX protein complex in schistosome mesenchymal cells was not explored, their importance in stem cell biology was investigated due to this cell population’s role in adult schistosome development and host interactions [37, 65]. Here, RNAi-mediated knockdown of either Smmbd2/3 or Smcbx (or both Smmbd2/3 and Smcbx, S2 Fig) in adult schistosomes led to significant reductions in the numbers of proliferative neoblasts, but not ovarian stem cells (Fig 6). This discrepancy (affecting neoblast but not ovarian stem cell proliferation) may be due to: 1) additional function (s) associated with non-5mC-mediated DNA binding [58, 59], 2) capacity to form other multi-protein complexes [58], 3) differential role (s) in neoblast/ovarian stem cells or 4) the incomplete knockdown of both Smmbd2/3 (54%) and Smcbx (57%) transcript levels in female schistosomes (Fig 6A). As ovarian stem cells appear to contain greater quantities of both Smmbd2/3 and Smcbx compared to neoblasts (Fig 5C vs 5A and [66]), residual levels of these two epigenetic regulators after RNAi may be sufficient to maintain proliferation in ovarian stem cells, but not neoblasts. Nevertheless, partial depletion of Smmbd2/3 or Smcbx transcript pools both significantly affected schistosome egg production and phenotype (Fig 7). While vitellocyte production in siSmmbd2/3 or siSmcbx treated females was comparable to siLuc controls (all normal/abnormal eggs contained DAPI+ vitellocytes, Fig 7C), this cell population’s role in egg-shell tanning appeared altered (noticeable difference in autofluorescence were observed). In addition to tanning, the size and shape of in vitro laid eggs would suggest that deficiencies in Smmbd2/3 and Smcbx also affect ootype and Mehlis’ gland contributions to oviposition. Therefore, further studies are necessary to understand how Smmbd2/3 and Smcbx contribute to vitellaria as well as Mehlis’ gland function, egg production rates, ootype biology and normal egg-shell tanning. However, similar egg-laying defects were observed in parasites treated with the DNA methylation inhibitor 5-AzaC [14], which, together with our current findings, provides further evidence for a functionally relevant DNA methylation machinery in schistosomes.
Our RNAi results are also consistent with studies conducted in mammalian systems where knockdown of Cbx2, Cbx3, Cbx4 or Cbx8 all resulted in decreased stem cell proliferation [67–70]. Where studies have been conducted in non-parasitic platyhelminth species (i.e. planarians), critical roles have also been established for both MBD2/3 and CBX in neoblast function [20, 71, 72]. However, due to differences (sometimes undetectable [20]) in underlying levels of genome methylation amongst platyhelminth species [39], the function of MBD2/3 and CBX proteins within the phylum is also likely to differ. For example, while SmMBD2/3 displays ubiquitous spatial expression throughout adult schistosomes (Fig 5), SmedMBD2/3 is exclusively expressed in planarian proliferating ASCs and germ line cells only [20]. Additionally, and also in contrast to our results where SmMBD2/3 appears vital for schistosome neoblast (but not ovarian stem cell) proliferation (Fig 6B), planarian neoblast proliferation does not seem to involve SmedMBD2/3 [20]. These data, together with those indicating a role for SmedCBX1 in regulating planarian neoblast function [71], strongly support differing functions of MBD2/3-CBX protein complexes within the platyhelminths. Indeed, characterising the differing functions of related platyhelminth epigenetic components (amongst the backdrop of divergent DNA methylomes) represents an exciting area of future research into the evolution of this phylum and control of its parasitic species [6].
Together, our data provides growing evidence supporting the view that schistosomes have an intact DNA methylation machinery (SmDNMT2 [14] and SmMBD2/3, this study). A suggestive link between DNA methylation and post-translational histone modifications (mediated by SmMBD2/3-SmCBX interactions) also indicates that the wider schistosome epigenetic pathway operates similarly to other characterised eukaryotes containing measurable DNA methylomes. Where differences in the roles ascribed to platyhelminth epigenetic components do occur, it is likely that these are related to loss of gene function (mutations in core components), lack of DNA methylation, divergent developmental biology pathways (free-living or parasitic species) or a combination of all three. The molecular details as to how SmMBD2/3-SmCBX interactions modify chromatin, influence other PPIs, regulate neoblast proliferation or shape other aspects of schistosome genome/transcriptome biology leading to egg production defects awaits further investigations. The results of such studies will lead to a greater understanding into how schistosome epigenetic components shape the developmental biology of this pathogen responsible for a devastating neglected infectious disease and perhaps shed light on novel ways for controlling its public health significance.
|
10.1371/journal.pcbi.1005095 | Inhomogeneity Based Characterization of Distribution Patterns on the Plasma Membrane | Cell surface protein and lipid molecules are organized in various patterns: randomly, along gradients, or clustered when segregated into discrete micro- and nano-domains. Their distribution is tightly coupled to events such as polarization, endocytosis, and intracellular signaling, but challenging to quantify using traditional techniques. Here we present a novel approach to quantify the distribution of plasma membrane proteins and lipids. This approach describes spatial patterns in degrees of inhomogeneity and incorporates an intensity-based correction to analyze images with a wide range of resolutions; we have termed it Quantitative Analysis of the Spatial distributions in Images using Mosaic segmentation and Dual parameter Optimization in Histograms (QuASIMoDOH). We tested its applicability using simulated microscopy images and images acquired by widefield microscopy, total internal reflection microscopy, structured illumination microscopy, and photoactivated localization microscopy. We validated QuASIMoDOH, successfully quantifying the distribution of protein and lipid molecules detected with several labeling techniques, in different cell model systems. We also used this method to characterize the reorganization of cell surface lipids in response to disrupted endosomal trafficking and to detect dynamic changes in the global and local organization of epidermal growth factor receptors across the cell surface. Our findings demonstrate that QuASIMoDOH can be used to assess protein and lipid patterns, quantifying distribution changes and spatial reorganization at the cell surface. An ImageJ/Fiji plugin of this analysis tool is provided.
| Plasma membrane organization is fundamental to cellular signaling, transport of molecules, and cell adhesion. To achieve this, plasma membrane proteins and lipids are spatially organized: they form clusters, aggregate in signaling platforms, distribute into gradients on polarized cells, or randomly distribute across the membrane. It is also clear that these organizations can be affected in various contexts. For example, in aging or neurodegenerative diseases, the composition of the plasma membrane is altered and, consequently, the protein and lipid distributions in the membrane fluctuate. In addition, cancer progression is characterized by changes in cellular polarity, lipid content, and the redistribution of cell surface receptors and adhesion molecules. Here we have developed a method to quantify such alterations that, unlike current tools, is compatible with diverse types of cellular organization, including polarity. Our tool can be employed to screen for changes in a straightforward manner and to elucidate distributions of cell surface components in different disciplines, ranging from neurobiology to cancer research.
| The function of cell surface proteins and lipids is tightly coupled to their spatial organization [1–3]. Membrane constituents cluster in nano- and micro-domains originating from lipid affinity (e.g., lipid rafts) [4], protein-protein interactions (e.g., tetraspanin domains) [5], and constraints imposed by the cytoskeleton [6]. Plasma membrane organization is also inherently asymmetric in polarized cells, such as migrating cells [7,8], epithelial cells [9], and neurons [10,11]. The ability to detect this plasma membrane organization is crucial for unraveling the dependency of signaling events, and understanding membrane regulation in a disease-related context [12].
One approach to assess the distribution of plasma membrane molecules is to consider these as bi-dimensional point processes that can be analyzed by spatial statistics. Point processes can be classified as: (I) Homogeneous or random, characterized by a constant density of points; (II) Inhomogeneous, characterized by a non-constant density of points; (III) Regular, with points equally dispersed; and (IV) Clustered, where points are grouped [13].
For investigating the spatial organization and especially for studying clusters, Ripley’s K-function [14] and pair correlation (PC) function approaches have been established. These measure the number from neighbors within a certain distance of a protein to determine the amount of clustering [15,16]. Several modifications and extensions of Ripley’s K-function have been made, including a model-based Bayesian approach [17], the extension to co-clustering [18], and an adaptation to account for limited localization precision in single molecule localization microscopy [19]. The PC function has been applied to images acquired by Photoactivatable Localization Microscopy (PALM) to quantify the heterogeneity of protein distributions on the plasma membrane [20]. In addition, the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm, a density based tool, was used to identify clusters of varying shape against a background in super-resolution microscopy [21]. Recently, the Ordering Points To Identify the Clustering Structure (OPTICS) algorithm was made available to the single molecule community to measure local density changes (overcoming some of the limitations of DBSCAN) [22,23]. Moreover, the Getis-Ord G statistic [24] has been used to quantify the degree of local protein clustering in super-resolution images [25] and compared to other methods including DBSCAN and PC analyses.
A key limitation of the current toolset is the lack of assessment on both clustering and polarity, equally important from a biological point of view. While the ability to detect specific clusters is of great importance, Sengupta et al. [20] among others, have shown that molecules like glycosylphosphatidylinositol (GPI)-anchored proteins are organized into different populations of clusters as well as residing as single molecules. All these organizations should be accounted for when analyzing the overall distribution. Furthermore, current tools require knowledge of the precise localization of the proteins and are thus limited to super-resolution or electron microscopy images. Therefore, we aimed to create an analysis tool that investigates biomolecule distributions without upfront information about their organization, and that can be applied to a variety of microscopy methods, including super-resolution and widefield microscopy. The main principle of our approach is to investigate spatial patterns of proteins and lipids and to quantify any deviation from random towards clustered or polarized organization as a measure of increased inhomogeneity.
We exploit a geometrical approach, called tessellation, to divide the image into tiles, and use the distribution of tile areas to characterize different patterns. The image analysis algorithm uses both the information of the neighbor relations of segmented objects and the intensity of the tiles in which they are confined. By using fluorescence intensity information for tile area correction, we have made the tool applicable to images acquired by a range of microscopy techniques. We have termed this analysis tool Quantitative Analysis of the Spatial-distributions in Images using Mosaic segmentation and Dual parameter Optimization in Histograms (QuASIMoDOH). An ImageJ/Fiji [26] plugin for QuASIMoDOH analysis is available along with instructional documentation (S1 File).
For QuASIMoDOH analysis we consider single fluorescent emitters distributed on the cell surface as a point process P in a bi-dimensional space. Considering a number N of individual points p of the process placed on the support S, the distribution pattern P is defined as:
P=∑i=1Npi with pi ∈ S,∀i∈{1,2,3…N},
(1)
where P describes a homogenous, clustered, or inhomogeneous pattern. Considering the brightness f of the fluorescent emitters (represented by p) and assuming the emitters being equally bright, the distribution pattern can be defined as:
P=∑i=1Npif with pi ∈ S, ∀ i ∈{1,2,3…N}.
(2)
When the point process P (S1A Fig) is imaged by an optical system, each point of position r, p(r), will be diffracted by the Point Spread Function (PSF). The resulting microscope image g(r) is typically approximated by the convolution (⊗) of the point p(r) with the PSF [27]:
g(r)=p(r) ⊗ PSF(r).
(3)
The result is a blurred image (S1B Fig). Due to the band limitation of the PSF, points of the pattern P placed at a distance shorter than the band limit will not be resolved as single points. In the blurred image, the intensity of the pixels depicting the diffraction pattern is directly related to the number of points contributing to this diffracted pattern.
Aiding the development of QuASIMoDOH, we generated in silico images of points (shown as single pixels) dispersed on a surface (Fig 1A, S1 Text). Blur and noise were then added into the image (Fig 1B) to mimic the acquisition process of a fluorescence microscope (see S1 Fig and S1 Text for detailed description). The individual steps of QuASIMoDOH analysis are depicted in Fig 1 and described in detail below:
To verify the ability of QuASIMoDOH to detect different distributions, we generated in silico images of different point patterns (see ‘Creation of discrete point pattern images’ in S1 Text and S3 Fig for a detailed description of the patterns). As described before, blur and noise were added to simulate the acquisition process of a fluorescence microscope (S1 Text, S1 Fig). We simulated images from widefield (WF) and Total Internal Reflection Fluorescence (TIRF) microscopy (resolution ~200 nm), Structured Illumination Microscopy (SIM) (resolution ~100 nm), and PALM (resolution ~20 nm). We decided to generate biologically relevant patterns, namely: random distributions (Fig 2A); random distributions of clusters with diameter d = 80 nm (Fig 2B); random distributions of clusters with diameter d = 240 nm (Fig 2C); polar distributions (Fig 2D); and polar distributions of clusters with diameter d = 240 nm (Fig 2E) for WF (Fig 2F–2J) and SIM images. For PALM, clusters with increasing sizes were simulated (see S1 and S3 Figs and S1 Text).
The correction factor used in the analysis of simulated images was determined by maximizing the accuracy of the analysis to give the correct number of known points. The accuracy is defined as the ratio between the number of areas, obtained by tile size correction, and the total number of points in the image (see S1 Text and S4 Fig). When the tessellation (Fig 2K–2O) and intensity correction are applied to images of different point patterns, the result is a set of tile areas typical for each pattern, captured by the histograms of corrected tile areas (Fig 2P–2T). The histograms display an increasingly steep peak for increasingly inhomogeneous patterns, moving from random to polar clusters. Modeling the distribution of tile areas by the Inverse Gamma PDF allows for the discrimination between different point patterns. The two parameters of the function are specifically associated with the histogram fit shape and its broadness. These parameters ultimately vary based on the underlying point pattern. In particular, they decrease with an increase in the inhomogeneity of the distribution, as shown in Fig 2U where the shape and scale parameters are plotted together.
The average (and standard error of the mean, SEM) from the analysis of 50 images for each point pattern is presented in Fig 2U. The analyzed images have the same density, ρ, defined as:
ρ=NA/AreaImage,
(10)
where NA is the number of areas obtained upon tile size correction and AreaImage is the total area of the image in μm2. Fig 2U shows the results from analyzing images with density ρ ~5 (tiles/μm2); the images in Fig 2F–2J are examples of this density. Fig 2V shows a plot of the shape and scale parameters obtained by analyzing images with different patterns and densities (1 ≤ ρ ≤ 7 tiles/μm2). The value of the shape parameter increases with increasing density, while the scale parameter decreases. Examples of images of density ρ ~1 and ρ ~7 (tiles/μm2) are given in Fig 2W top and bottom, respectively. S5 Fig shows the densities associated with each point on the reference graphs for microscopy techniques with different resolution.
The deviation from a random distribution (calculated as the Euclidean distance from the random reference point) is used as a measure of the inhomogeneity of a point pattern. This is quantified in Fig 2X. Here the inhomogeneity is calculated for images with ρ ~5 (tiles/μm2) (Fig 2U). To provide comparable results for each density, the interval between the two extremes (random and polar clusters) was normalized to one. Analysis of images similar to the one shown in S6A Fig can result in shape and scale parameters larger than shape and scale parameters for the corresponding random distribution reference (S6B Fig). The inhomogeneity measure, in this case, will have a negative value (S6C Fig) since the coordinates of the random distribution are used as the origin from which the inhomogeneity is calculated. Shape and scale parameters 6–8 times larger than the random distribution reference can be further explained by a regular pattern where points are equally dispersed (S6D–S6F Fig).
Taken together, these results show that QuASIMoDOH can be used to quantify the inhomogeneity of spatial distributions and demonstrate the validity of the approach at the level of simulated images. We subsequently validated our method using images from fluorescence microscopy.
Real and simulated images can exhibit analogous features, but they may differ as a consequence of microscope/detector sensitivity and dynamic range. Therefore, the correction factor C for actual microscopy images was selected to establish a comparable degree of correction as that determined by simulated results (see S1 Text and S7 Fig). We took advantage of plasma membrane sheets (S8 Fig) [36,37] to acquire images of the cell surface without fluorescent background arising from the cytoplasm [38,39]. To control for the integrity of the plasma membrane, cells were initially incubated with the lipophilic fluorescent stains DiO or DiI, depending on the combination of subsequent protein/lipid staining.
We tested whether QuASIMoDOH analysis was able to replicate findings made using microscopy techniques with different resolution and measured the inhomogeneity value from similar samples. We computed the cluster size of the lipid raft marker glycosylphosphatidylinositol (GPI)-anchored protein by comparing the results against the reference graph for simulated WF, SIM, and PALM images (S5 Fig). Images of GPI-anchored protein tagged with green fluorescent protein (GFP) or photoactivatable GFP (paGFP) [40], were acquired by WF, SIM, and PALM (S9A–S9C Fig). QuASIMoDOH analysis of SIM and WF data gave comparable results (S9D and S9E Fig). The analysis of 16 PALM images resulted in a GPI cluster size between 50 and 100 nm in diameter (two regions were identified as random and one region with a low r2 was excluded) (S9F Fig). PC-PALM analysis of the same 16 images resulted in an average GPI cluster size of approximately 80 nm in diameter (one region was identified as random). Obtained values for GPI cluster size were also consistent with previously published work [41,42]. Thus, QuASIMoDOH is capable of providing fast and quantitative analysis of super-resolution data with respect to cluster size. Finally, GPI inhomogeneity measurements from WF, SIM, and PALM data provided similar results when using appropriate tile area intensity corrections (see S9G Fig), indicating the applicability of the approach across microscopy techniques with various resolutions.
To further confirm that QuASIMoDOH detects different distributions we labeled three proteins that have well described and distinct spatial arrangements in Mouse Embryonic Fibroblasts (MEFs) using indirect immunostaining, and imaged these by WF microscopy. We used the sodium-potassium pump (Na+/K+ ATPase) [43] (Fig 3A and 3B), the transferrin receptor (TfR) [44] (Fig 3C and 3D), and caveolin1 (Cav1) [45] (Fig 3E and 3F). The Na+/K+ ATPase is an integral membrane protein that is randomly distributed across the cell surface. The TfR is an integral membrane protein responsible for ferric ion uptake, which clusters in clathrin-coated pits (200 nm in size) after ligand binding and prior to endocytosis. Cav-1 is a cytosolic peripheral-membrane protein that functions, together with cavin, in the formation of caveolae. Cav-1 is predominantly localized to the trailing edge of migrating cells, and thus shows a polarized distribution. Analysis by QuASIMoDOH demonstrates that the different distribution patterns of immunolocalized Na+/K+ ATPase, TfR, and Cav-1 can be reliably distinguished based on their inhomogeneity (Fig 3G and 3H). Comparing the results to simulated images indicates that the distribution pattern of Na+/K+ ATPase most closely resembles simulated images with a random distribution, whereas TfR matches those with a clustered distribution, and Cav1 aligns with simulated images highlighting a polar distribution of clusters. Thus, QuASIMoDOH effectively captures different protein distributions. Furthermore, these results demonstrate that our analysis can be successfully applied to images of proteins detected by immunolabeling.
The cell surface is also patterned by the wide variety of lipid molecules present in the plasma membrane. We, therefore, explored whether QuASIMoDOH detects lipid distributions, including changes that result from perturbation of lipid trafficking (Fig 4A–4D). For this purpose, we treated MEF cells with U18666A, an amphipathic steroid that causes cholesterol and sphingomyelin accumulation in late endosomes and lysosomes [46,47] (3 μg/mL for 18 h). The treatment was followed by fixation and incubation with Lysenin (a toxin that specifically binds sphingomyelin in the membrane [48,49]) and immunostaining. We reasoned that blocking endosomal sphingomyelin trafficking would deplete this lipid from the cell surface where it is normally organized into clusters, possibly causing reorganization to a more homogeneous distribution. Indeed, QuASIMoDOH analysis detects a significant difference in sphingomyelin organization between drug-treated and control cells (t-test: p = 0.045; two-tailed; unpaired, Fig 4E and 4F). Thus, QuASIMoDOH analysis can also detect plasma membrane lipid distributions and how these are altered in response to cellular perturbations.
To further explore QuASIMoDOH analysis in combination with different techniques we turned to TIRF imaging (with widefield resolution). We monitored the internalization of the epidermal growth factor receptor (EGFR) [50] (S10 Fig). We performed this in HeLa cells transiently transfected with EGFR-GFP, and treated with two concentrations of EGF reported to direct EGFR internalization distinctly [51,52]: a low EGF dose (2 ng/ml) induces EGFR internalization via the clathrin-mediated route, while a high EGF dose (20 ng/ml) directs EGFR internalization through both clathrin-coated pits and caveolae. TIRF images of cells fixed at specific time points (t0, before stimulation, t1 = 2, t2 = 5, t3 = 7, t4 = 10, and t5 = 15 minutes following EGF addition) were acquired, and the organization of EGFR at the plasma membrane was analyzed by QuASIMoDOH (Fig 5). This revealed that EGF indeed altered the EGFR distribution (Fig 5G–5J), corresponding to an increase in the measure of inhomogeneity. Furthermore, as predicted, this occurred more rapidly for cells exposed to the high concentration of EGF. These results demonstrate that QuASIMoDOH can be used to follow dynamic processes occurring at the cell surface and distinguish those processes with different rates.
The EGFR internalization assay raises the question: Can different spatial distributions observed in the plasma membrane (Fig 5E and 5F) be analyzed on a local scale as opposed to the global scale described above? To address this, we expanded QuASIMoDOH analysis to assess local distribution differences. For a local analysis, the tile area distribution analysis (see above ‘5. Tile area distribution analysis’) is applied to a subset of tiles in the images. This subset of tiles is selected by a circle that moves from the center of one tile to the next (Fig 6A and 6B). A color is assigned to the tile in the center based on the distribution detected using the selected tiles. Magenta represents a random distribution, green represents a distribution in clusters (for simplicity, the distinction between clusters of different sizes is omitted), and red represents polar distributions (for simplicity, the distinction between polar and polar clusters is omitted). When switching to a local analysis, the fitting quality (see S11 Fig) can be affected. To limit incorrect distribution assignment, we set 0.45 as a minimum r2 and a cut off for the outliers that, despite a high enough r², are too far away from any reference point (distance > 1, for widefield images), resulting in tiles without color.
To test the local analysis, we first created in silico images where the upper part contained a random distribution and the lower part consisted of a clustered distribution (see example in Fig 6A). These test images were simulations of WF images with a size of 512 x 512 pixels and 1600 points, on average. To maximize the number of tiles with detected distributions, the diameter of the circle must be in the range of 4 to 10 μm (Fig 6C, S1 Text and S12 Fig). As expected, from the local analysis we obtained a random distribution in the upper third of the image, a clustered distribution in the lowest third, and a polarized distribution in the center due to the neighborhood abrupt change from one part of the image to the other in this transition region (Fig 6D).
We next applied the local analysis to a fixed cell image from our EGFR internalization assay (treated for 10 min with high dose of EGF, Fig 6E). In Fig 6F, we applied the local analysis by setting the circle diameter in the range of 4 to 15 μm (similar to the simulation, the maximum diameter is about half of the image). A clear clustered organization of the receptor becomes apparent at the periphery of the cell following the analysis. A random distribution, however, covers the center of the cell and a gradient of random and clustered receptors mark the transition between these regions. Finally, an image of EGFR on the surface of a live cell, stimulated by 20 ng/mL of EGF, was used to further investigate the successful application of this local analysis approach. We were able to observe the local variation of receptor distributions following the time dependence of stimulation (see S1 Movie). These results indicate that QuASIMoDOH can be used to assess both the global and local changes in the distribution of fluorescent patterns at the plasma membrane.
Here we present QuASIMoDOH as a new approach to measure the inhomogeneity of a spatial distribution as a deviation from random towards clustered and polarized patterns (S13 Fig). Different from methods such as PC-PALM [20], Ripley’s K-function [14], nearest neighbor approaches [13], and DBSCAN [22,23], QuASIMoDOH is compatible with polarized distributions (see Caveolin-1 in Fig 3). It can detect and measure polarized distributions independent from their orientation, while other tools must acquire information on cell morphology or other features to assess the spatial phenotype of polarized molecules [53]. Compared to grid-based algorithms, like the Hoshen-Kopelman algorithm [54] that divides space into a grid and identifies clusters as continuously occupied areas, QuASIMoDOH can correct for unresolved points by complementing the tile area dataset with the information on tile intensity and is thus applicable beyond single molecule imaging techniques.
Where SpIDA [55] measures protein interactions and aggregation by multiple fitting of pixel intensity histograms, QuASIMoDOH analyzes the fit from tile intensity histograms to extract a measure of plasma membrane inhomogeneity. Comparable to Number and Brightness (N&B) approaches [31], which analyze temporal fluctuations, in QuASIMoDOH the tile intensity IT depends on the number of molecules (p) in the tile and their brightness (f): IT = pf. After calibration, incorporating background intensity, N&B can be mapped to absolute values with good spatial resolution. For ease of use in QuASIMoDOH, we decided to estimate the correction factor from the image by an analysis of the distribution of tile intensities. Determining the correction factor, however, leaves the possibility of misinterpretation due to background, non-uniform illumination, or the presence of artifacts. Pre-processing steps to address these conditions are provided in the QuASIMoDOH documentation (S1 File). S14 Fig offers information on how to determine if an image is suitable for QuASIMoDOH analysis.
We applied QuASIMoDOH to PALM, SIM, WF, and TIRF images (Figs 3–5, S9 Fig). However, in principle, this tool can also be applied to other microscopy modalities, including confocal microscopy and (d)STORM imaging. We have demonstrated that results obtained for the cluster analysis of GPI-anchored proteins are in reasonable agreement across microscopy techniques with different resolution and with previously published results obtained by PC-PALM analysis [42]. Additionally, QuASIMoDOH correctly detected the differential distribution of Na+/K+ ATPase, TfR, and Cav-1 proteins. Notably, QuASIMoDOH does not require the presence of a reference molecule in the sample with a known distribution. As a result, comparisons between protein/lipid organization in different cell types or within the same cell and under different experimental conditions are possible, e.g., evaluating changes in a lipid distribution following drug treatment. We specifically demonstrated this by analyzing the distribution of sphingomyelin (Fig 4). The observed increase in homogeneity for the lipid probe Lysenin in cells treated with U18666A compared to untreated cells emphasizes the potential of our analysis tool to readily test a range of cell perturbations. Additionally, the fact that affinity probes, fusion proteins, and lipid probes can be detected highlights the versatility of our analysis method. We further demonstrated the ability of this approach to reveal the dynamic nature of EGFR organization at the plasma membrane upon stimulation, using both fixed and live cells.
A list of advantages and limitations of QuASIMoDOH is provided in Table 1. An important feature of QuASIMoDOH analysis is its direct application to microscopy data with no detailed prior knowledge of the sample. Overall QuASIMoDOH serves as a quick, straightforward, and automatable method to measure distribution patterns of proteins and lipids on the cell surface. It can be used to study events at the cell surface related to cell signaling or remodeling as these appear, for example, in the reprogramming of cancer cells or neuronal differentiation.
For QuASIMoDOH development and testing, a Dell Optiplex7010 computer was used (Intel CITM) i7-3770 CPU @ 3.40GHz processor and 4.00 GB RAM, running Windows 7 Professional). Images were run in batch mode, each 256 x 256 pixels in size. In total, approximately 25,000 points were identified, and using a standard four year old personal computer, we were able to analyze the entire dataset of 50 images in 22 seconds. Pixel-by-pixel image processing of both simulated and fluorescent images was carried out using custom-made routines in MATLAB and ImageJ/Fiji [26] according to the schemes described in the main text. For plasma sheet samples, background correction was applied (where necessary) by subtracting a background value either using the Dip Image function “backgroundoffset” or manually. Background subtraction on TIRF images of intact cells was carried out using the ImageJ/Fiji function “Rolling Ball” [26].
The WF, SIM, and TIRF images are initially filtered by the ImageJ/Fiji filter ‘Sigma Filter Plus’ and then smoothed. For the separation of signal and background in our simulated images, we used the default threshold in ImageJ/Fiji [57], which is based on Isodata. For the widefield images of Na+/K+ ATPase and Caveolin-1, the threshold Li [58] was used. GPI, TfR, and Lysenin widefield images, as well as GPI SIM and EGFR TIRF images, were thresholded using Mean [59]. Apparently, the best-suited threshold for an image can depend on the imaging modality and was chosen upfront based on the obtained images (see S1 Text). For skeletonization, the function implemented in ImageJ/Fiji version 1.47 was used. PALM super-resolution images were generated by analyzing datasets obtained from Peak Selector software (Research Systems, Inc.). Analysis of paGFP-GPI images was performed on 16 square regions of 7–18 μm2 obtained from 8 cells, with an average of 75 ± 5 localized peaks/μm2 and average localization precision of 15 nm. PC-PALM analysis was performed similarly to the previously described method [20]. For QuASIMoDOH analysis, localized peaks were grouped using a group radius of 3 x maximum localization precision and a maximum dark time of 5 s using Peak Selector. The maximum dark time was obtained experimentally using sparse paGFP. Similar to the method previously reported by Annibale et al. [60], a best fit of the observed fluorophore counts as a function of the dark time was used to determine the effective number of molecules present in the sample. Grouped peak coordinates (in pixels) were subsequently fed into ImageJ/Fiji for QuASIMoDOH analysis. Peaks were plotted following a conversion to nanometers by a user-defined pixel size. The image width and height was calculated from the coordinates as the difference between the maximum and minimum values of X and Y coordinates, respectively. The generated image was then scaled to a 2.5 nm pixel size, and subsequently analyzed as described before.
All graphing and statistics were prepared using MATLAB, GraphPad Prism (GraphPad, La Jolla, USA), and Excel (Microsoft, Redmond, USA). For the analysis of distribution patterns of proteins and lipids, images of cells from three different preparations/coverslips were analyzed. Images were pooled for further QuASIMoDOH analysis.
Mouse Embryonic Fibroblast (MEF) cells were maintained at 37°C and 5% CO2 in DMEM/F12 (+L-glutamine + 15 mM HEPES) supplemented with 10% of fetal bovine serum (FBS). MDA-MB-468 cells were cultured in DMEM supplemented with 10% FBS. Transient transfection of MDA-MB-468 cells was performed using Jetprime (PolyPlus, following manufacturer instructions) with 2 μg of paGFP-GPI similarly as described in detail elsewhere [42]. Transient transfection of MDA-MB-468 cells was performed using FuGENE (Promega, following manufacturer instructions) with 2 μg of GFP-GPI. HeLa cells were grown in DMEM/F12 (+L-glutamine + 15 mM HEPES) supplemented with 10% FBS. One day after plating cells, transfection was performed using FuGENE6 with 0.5 μM of plasmid pGFP encoding for EGFR-GFP.
Plasma membrane sheets were prepared as previously described [36]. In short, MEF cells were cultured in DMEM/F12 (+L-glutamine + 15 mM HEPES) supplemented with 10% of fetal bovine serum. Cells were grown on coverslips to approximately 60% confluence. At 4°C, cells were washed with PBS+/+ and subsequently with coating buffer (20 mM MES, 135 mM NaCl, 0.5 mM CaCl2, 1 mM MgCl2, pH5.5). Next, they were incubated in coating buffer with 1% of silica beads for 30 min. They were rinsed with deionized water (10 min) followed by three washing steps with PBS+/+. To prepare plasma membrane sheets, shear force was applied to the coverslip using a syringe held at a 30° angle (on the coverslip). As the upper surface of adherent cells is made rigid by the silica coating, shear forces break off membranes at the edges releasing all soluble contents and retaining only the basal plasma membrane adherent to the coverslip. These remaining sheets were then fixed (4% paraformaldehyde in PBS, 15–20 min at RT) (see schematic in S8 Fig).
Fixed plasma membrane sheets were rinsed with PBS+/+ and blocked (2% FCS, 2% BSA, 0.2% gelatin, 5% goat serum in PBS-/-) for 1 h. For immunofluorescence, the following antibodies were used: mab to murine Cav-1 (BD Biosciences; San Jose, USA) (1:200); mab against Transferrin receptor (Zymed/Life Technologies) (1:100); mab against Na+/K+ ATPase alpha (Novus Biologicals, Littleton USA) (1:100). Alexa488/555/568 conjugated secondary antibodies were used (Life Technologies) (1:1000). Lysenin (1:40) was purchased from Peptide Institute (Osaka, Japan) and immunolocalized using anti-Lysenin pab (1:100) followed by goat anti-rabbit Alexa-555 (Life Technologies) secondary antibody. U18666A was purchased from Sigma. DiI and DiO (Life Technologies) (1:100) were used to stain lipid bilayers. Staining was performed according to manufacturer recommendations.
Twenty-four hours after transfection, HeLa cells were serum starved overnight. For the preparation of the fixed samples, after starvation, the cells were incubated at 37°C with either 2 ng/mL or 20 ng/mL EGF for different time intervals (2, 5, 7, 10, and 15 minutes), then fixed with 4% paraformaldehyde in PBS, for 15–20 minutes at room temperature. The fixed samples were then imaged by TIRF. Additionally, living cells were imaged by TIRF upon stimulation with 20 ng/mL of EGF (S1 Movie).
Widefield and structured illumination images were acquired with a structured illumination microscope (Elyra S1 (Carl Zeiss, Jena, Germany) equipped with a 63x oil objective lens with a numerical aperture (NA) of 1.4 and an Andor iXon 885 EM-CCD camera). PALM imaging was performed using a Nikon Instruments Ti Eclipse inverted microscope with a 100x/1.49 NA TIRF objective (Apo) and a 488 nm laser (Agilent, MLC-MBP-ND laser launch) with an EM-CCD camera (Andor Technology, iXon DU897-Ultra). The microscope was equipped with a Perfect Focus Motor to minimize axial drift over the duration of imaging. paGFP was simultaneously activated and excited with the 488 nm laser at an intensity set to 1.45–1.9 mW (as measured at the optical fiber). Exposure time was set at 100 ms. Imaging was performed until paGFP was completely exhausted, typically after 20,000 frames. TetraSpeck beads (Life Technologies) were used as fiducial markers for drift-correction during image acquisition. TIRF images were acquired by a Nikon Ti Eclipse inverted microscope, equipped with a 100x/1.49 NA oil objective lens and a Hamamatsu Orca D2 camera, or Hamamatsu Orca Flash 4 Lite, for living cell imaging. Confocal images were acquired using the same microscope equipped with Nikon A1R using a 60x 1.4 NA oil objective. Co-localization analysis was carried out using the Fiji plugin JACoP [61].
|
10.1371/journal.pbio.2000737 | Four simple rules that are sufficient to generate the mammalian blastocyst | Early mammalian development is both highly regulative and self-organizing. It involves the interplay of cell position, predetermined gene regulatory networks, and environmental interactions to generate the physical arrangement of the blastocyst with precise timing. However, this process occurs in the absence of maternal information and in the presence of transcriptional stochasticity. How does the preimplantation embryo ensure robust, reproducible development in this context? It utilizes a versatile toolbox that includes complex intracellular networks coupled to cell—cell communication, segregation by differential adhesion, and apoptosis. Here, we ask whether a minimal set of developmental rules based on this toolbox is sufficient for successful blastocyst development, and to what extent these rules can explain mutant and experimental phenotypes. We implemented experimentally reported mechanisms for polarity, cell—cell signaling, adhesion, and apoptosis as a set of developmental rules in an agent-based in silico model of physically interacting cells. We find that this model quantitatively reproduces specific mutant phenotypes and provides an explanation for the emergence of heterogeneity without requiring any initial transcriptional variation. It also suggests that a fixed time point for the cells’ competence of fibroblast growth factor (FGF)/extracellular signal—regulated kinase (ERK) sets an embryonic clock that enables certain scaling phenomena, a concept that we evaluate quantitatively by manipulating embryos in vitro. Based on these observations, we conclude that the minimal set of rules enables the embryo to experiment with stochastic gene expression and could provide the robustness necessary for the evolutionary diversification of the preimplantation gene regulatory network.
| The first 4.5 days of mammalian embryo development proceeds without maternal information and is remarkably robust to perturbations. For example, if an early embryo is cut in half, it produces 2 perfectly patterned, smaller embryos. Where does the information guiding this development come from? Here, we explore this issue and ask whether a model composed of a simple set of rules governing cell behavior and cell—cell interactions produces in silico embryos. This agent-based computational model demonstrates that 4 rules, in which a cell makes decisions based on its neighbors to adopt polarity, make lineage choices, alter its adhesion, or die, can recapitulate blastocyst development in silico. By manipulating these rules, we could also recapitulate specific phenotypes at similar frequencies to those observed in vivo. One interesting prediction of our model is that the duration of cell—cell communication through fibroblast growth factor (FGF) signaling controls scaling of a region of the blastocyst, and we confirmed this experimentally. Taken together, our model specifies a set of rules that provide a framework for self-organization, and it is this self-organizing embryogenesis that may be an enabler of stochastic variation in evolution.
| Early mammalian development is a fascinating example of how deterministic spatiotemporal patterns emerge at the level of cell populations from highly stochastic regulatory components. During mouse preimplantation development, 2 sequential lineage decisions take place [1] (Fig 1), and these decisions are marked by the expression of lineage-determining transcription factors. The first decision happens between embryonic day (E) 2.5 and 3.0, as the morula is formed. The outer cells of the embryo express the transcription factor caudal-related homeobox 2 (Cdx2) and form the trophectoderm (TE), while the inside cells express sex-determining region Y-box 2 (Sox2) [2] and form the inner cell mass (ICM). The morula then cavitates, forming the blastocyst, and the ICM differentiates into 2 lineages: Gata6-expressing cells form the primitive endoderm (PrE), an epithelial layer adjacent to the blastocoel cavity, and Nanog-expressing cells form the epiblast (EPI) enclosed by the TE and the PrE. The specification of EPI and PrE is a gradual process that involves the initial specification of cell types in a salt-and-pepper distribution throughout the ICM and then their progressive segregation by E4.5, the time of implantation [3,4]. All future lineages of the embryo, including the germ line, are generated from the EPI. The TE and PrE lineages will produce the support structures required for placental and yolk sac development.
These early decisions are remarkable in that they proceed in the apparent absence of maternal information, and that the cells undergoing these differentiation decisions remain competent for respecification up to around E3.5. Either the removal of blastomeres or the aggregation of multiple morulae as late as E3.5 can produce developmentally competent embryos, albeit at a lower success rate [5,6]. In fact, single blastomeres from 32-cell stage mouse blastocysts can generate entire mice [7,8].
While the analysis of mutant phenotypes has suggested the broad outlines of several regulatory mechanisms involved in embryo development, the robust and regulative nature of early patterning cannot currently be explained. Based on the literature, we have identified 4 major regulatory themes:
At compaction, E2.5–E3.0, the outer cells of the embryo become polarized, express the transcription factor Cdx2, and differentiate into TE. At E3.0, the apical membrane of outer cells expresses several proteins that have a known role in cell polarity [9–13]. The acquisition of polarity starts with compaction at the 8-cell stage, in which apical domain is developed at the contact-free surface. The apical domain is inherited asymmetrically at the next cell division and was shown to play an important role in defining inner and outer cells through cells sorting due to differential contractility [14]. How the TE fate becomes limited to only outer cells is not fully understood, but it is suggested to be a combined effect of contractility, Hippo, and Notch pathways. The Hippo signaling pathway is normally activated at high cell densities and, in this context, is specifically activated in the inner cells [15–18] in which it induces phosphorylation and degradation of the transcriptional coactivator Yes-associated protein 1 (Yap). In outer cells with higher contractility, the levels of phosphorylated Yap are higher [14]. In the absence of Hippo activation, the TEA domain family member 4 (Tead4) binds Yap and cooperates with Notch signaling to induce the transcription factor Cdx2 and specify the TE [19]. In addition, tight junctions are formed between the TE cells in a plane perpendicular to the polarity axis [20,21], and this may further reinforce TE polarity.
At E3.0, the inside cells down-regulate Cdx2 but express octamer-binding transcription factor 4 (Oct4) and become ICM. The cells of the ICM initially express both Gata6 and Nanog, but early variations in expression are thought to be propagated by the production of fibroblast growth factor (FGF) 4 downstream of Nanog and higher levels of the FGF receptor (FGFR2) in cells expressing higher levels of Gata6. EPI precursors expressing Nanog secrete FGF4, promoting a PrE fate in neighboring cells [22–26]. Consistently, ex vivo manipulation of the FGF pathway from E2.5 to E4.0 can change the fate of ICM cells [27–30]. It has been shown in vitro that Nanog and Gata6 repress each other intracellularly [31–36]. Moreover, FGF/extracellular signal—regulated kinase (ERK) signaling enhances Gata6 expression while repressing Nanog [36–39]. Finally, modulating the FGF4 level is sufficient to convert all ICM cells to either PrE (high FGF4) or EPI (low FGF4) [40,41].
During the period that cells are making a lineage decision between EPI and PrE, cell movement occurs within the ICM [42]. Chazaud et al. [3] showed that initially EPI and PrE progenitors arise in a heterogeneous mosaic pattern and later physically segregate into the appropriate cell layers, which are finally separated by a basal lamina. It was proposed that the cell movements contribute to cell sorting and may be due to differential adherence of progenitor cell types, which has been observed in vitro [43,44]. There have also been several reports on differences in the expression level of the adhesion molecule integrin β1 receptor during PrE differentiation in vitro between the 2 ICM lineages [45–47]. Several other mechanisms contribute to the formation of the “layered” pattern [48], including down-regulation of transcriptional programs in inappropriately positioned cells or apoptosis [4].
As the blastocyst expands, the ratios of the PrE and EPI are self-regulating, as paracrine interactions control proliferation and apoptosis. In particular, the cytokine Leukemia inhibitory factor (LIF) appears to regulate the relative size of the PrE and EPI. LIF is secreted by TE cells, and the corresponding receptor complex is found in the ICM [49]. LIF has been shown to act both on EPI and PrE fate. It blocks maturation in the EPI, and it supports proliferation and cell survival in the PrE [50,51]. In addition, atypical protein kinase C (aPKC) and platelet-derived growth factor (PDGF) signaling promote survival of PrE precursors that reach the surface of the ICM [52,53]. Furthermore, a considerable number of ICM cells undergo apoptosis around the time of PrE formation [4,54,55]. Plusa et al. [4] showed that there is a steady increase in the rate of apoptosis from E4.0–E4.5. They reported that PrE precursors are more likely to undergo apoptosis when they are deep within the ICM than when they are positioned along the cavity lining.
Here, we hypothesized that a combination of these 4 themes could together explain the robust nature of blastocyst formation. We have conceptualized and unified these themes as rules in a rule-based model to investigate their relative contribution to the robustness of early embryo development.
Both the initial specification of the TE and the ICM and differentiation and segregation of PrE and EPI have been modeled in silico at different levels. Chickarmane et al. [56] focused on intracellular transcription networks generating the 3 stable states (EPI, PrE, and TE). Bessonnard et al. [57] modeled 25 static ICM cells on a grid and addressed how cell—cell communication via the FGF/ERK pathway establishes the right proportion of EPI and PrE cells. Krupinski et al. [58,59] modeled the mechanical interaction of cells, focusing on the role of polarity in Cdx2 partitioning, as well as differential adhesion and directed movements for the segregation of PrE, EPI, and TE into 3 distinct layers. In these models, the growth of the blastocyst is driven by the growing cavity, and all cells have similar apolar interactions, albeit of different strength. None of these models address the role of polarity in TE cell—cell interaction, apoptosis, or aspects of the emergence of blastocyst scaling [60–62].
The existing in silico models provide important insight into the individual mechanisms driving cell specification during preimplantation development but do not provide a unified framework of early embryo development as a self-organizing system [63]. Recently, such a framework, using rule-based modeling, has been proposed for the specification of synaptic partner cells [64]. Here, we use a similar approach to propose a minimum set of rules to quantitatively model early blastocyst development. Our aim is not to recapitulate the precise timing of mouse development, but to show that with a simple set of rules we could capture blastocyst patterning. As evolution can produce changes in the timing and wiring of the gene regulatory network, the patterning of the mammalian embryo should be able to tolerate stochasticity; our aim is to show that the 4 simple rules can enable this robustness.
We focus on 4 rules that include polarity, cell—cell communication via FGF4 signaling, differential adhesion, and apoptosis. Using a series of in silico 2D simulations, we quantify the relative contribution of these 4 elements to early embryonic development. To facilitate comparison to published genetic studies, we have validated this approach in 3D simulations. By introducing polar interactions between TE cells, we show that the development (including cavity formation) is self-organized and does not require an a priori assumption of the growing cavity. Moreover, based on these 4 rules, we found that we could effectively simulate experimental embryo manipulation: our model successfully reproduces a range of experimentally observed mutant phenotypes and predicts that the time point of FGF activation could be a clock that dictates the relative size of EPI and PrE in scaling experiments. Consistent with the notion that the timing and duration of FGF/ERK activation is the essential variable in proportioning these 2 lineages, we found that delaying ERK activation by 24 hours resulted in a quantitative reduction in PrE specification.
In a growing blastocyst, cells are tightly packed and adhere to each other. Similarly to earlier in silico models [58,65], we simulate this by introducing an interaction potential in which cells repel each other at a distance smaller than their typical size and attract at longer distances (see Fig 2 and Materials and methods for details of the potential). The interactions between all cell types are the same, except in 2 cases. First, to simulate differential adhesion, the attraction is set to be weaker with and among PrE cells. Second, in contrast to previous models, the physical forces between TE cells are assumed to depend on cell polarity such that 2 TE cells adhere to each other when their polarity is pointing in the same direction, and cells are positioned next to each other in the plane perpendicular to the polarity axis. Biologically, this would correspond to a well-known phenomenon of tight junctions forming in a plane perpendicular to the polarity axis [66].
The modeled blastocyst grows as cells divide. Cell division is simulated by selecting a cell at random and introducing a daughter cell between the mother and its nearest neighbor. In case of TE, 1 daughter cell inherits polarity including the orientation of the polarity from the mother cell.
To conceptually capture the 4 major themes outlined in the Introduction, we have formulated the following 4 rules:
At E3.0, we define the outer cells by counting the number of nearest neighbors (shown in Fig 2a). Cells with fewer than 5 nearest neighbors are assigned TE fate and polarity, pointing radially outwards from the center of the cell mass. We do not aim at recapitulating how the polarity is established and how inner and outer cells are defined in a real embryo. Instead, we focus on the role of polarity in outer cells after it has been established. To take into account that TE cells are about twice as big as the rest of the cells, each TE cell is simulated by 2 unit circles. Polarity of the TE cells is assumed to lead to polar interactions that can be thought of as tight junctions forming in the plane perpendicular to the TE polarity. We simulate this by a polarity-dependent attraction factor, S (see Eq 1), that is maximal for 2 neighboring cells if their polarity is oriented in the same direction and perpendicular to the position vector (illustrated in Fig 2b). At every time point, the strength of attraction, S, between 2 neighbor TE cells is given by:
S=−1.4 (e^1 × r^12)⋅(e^2 × r^21)
(1)
Where ê is the polarity unit vector of a TE cell, and r^ is the unit distance vector between 2 neighbor TE cells. The prefactor of 1.4 assures that TE cells are tightly packed but nonoverlapping. Notice that this polar interaction favors the formation of a single-layer sheet with cells positioned perpendicular to the polarity axis and disfavors compact structures (Fig 2b).
The dynamics of the polarity vectors is governed by simple damped dynamic equations as in Eq 3 (see Materials and methods). Proliferation of TE was reported to be about twice as fast compared to ICM [67], so we set the rate of TE cell division to be 2-fold that of ICM cells. In all lineages, daughter cells inherit the mother’s fate and polarity.
From E3.0 to E4.0, the FGF signaling pathway becomes important for lineage segregation in the ICM. Bessonnard et al. [57] suggests that the FGF/ERK pathway coupled with the intracellular mutual inhibition between Nanog and Gata6 act together to ensure the fidelity of initial EPI and PrE specification. The simplified logic behind this process can be reduced to the intracellular inhibition and extracellular activation between Nanog and Gata6 (shown in the network in Fig 1 at E3.5):
The mutual inhibition between Nanog and Gata6 is, in effect, an intracellular positive feedback loop. When reduced to 1 variable (e.g., Nanog), the network reveals a combination of intracellular amplifying positive feedback with extracellular inhibition of Nanog in neighboring cells. This representation suggests a Turing mechanism that results in both local amplification and global inhibition (see S2 Fig). Simulations based on this Turing mechanism predict that ICMs would maintain the ratio of PrE (or EPI)/ICM, irrespective of embryo size.
At E3.0, all ICM cells are in an “undetermined state” coexpressing low levels of both Gata6 and Nanog (Fig 1), and the cell specification process is started as all these cells begin to express FGF4 [40,69]. This initial step follows the same simplified logic outlined above, as long as we assume that once specified, Gata6 cells have a lower concentration of FGF4 in their neighborhood than undetermined ICM cells.
In the model, we implement this logic by monitoring the number of nearest-neighbor cells with high FGF4 (EPI or undetermined ICM) versus low FGF4 (PrE) (as in Fig 2a). At cell division, the likelihood of a mother and a daughter cell to convert to PrE is proportional to the fraction of high FGF4 (EPI or undetermined ICM) cells in the neighborhood (P(PrE)=# of high FGF4 neighbors# of ICM neighbors), and, conversely, the likelihood to convert to EPI is P(EPI) = 1 − P(PrE). As a result of this simple rule, in our simulations, cells undergo the specification into salt-and-pepper pattern of Nanog/Gata6 cells from the initial state of unspecified ICM. In addition, all cells can potentially convert their identity between PrE and EPI later (E3.5 to E4.0), as the blastocyst grows. Although the identity of cells can be modulated in ex vivo blastocyst culture in response to FGF treatment or inhibition [27], once cell identity is established, conversion is quite rare in unmanipulated culture conditions [70]. In our simulations, the rate of conversion is low and asymmetric, which is in line with observations by Xenopoulos et al. [70].
Differential adhesion is activated once the cells specify their identity (at E3.5). It is implemented by a single change in the attraction factor for PrE precursors from S = 0.6 to S = 0.4. Biologically, this corresponds to lower adhesive properties of PrE cells. Thus, the attraction factor between 2 EPI, 2 TE cells, or an EPI cell and an ICM or TE cell remains at S = 0.6, while the attraction factor, S, between 2 PrE or PrE and any other cell type is reduced to S = 0.4. These potentials are shown in Fig 2c. TE cells only interact with their nearest neighbors (i). Limiting the range of the TE—ICM potential (ii) to about 2 cell diameters allows the model to capture the symmetry breaking event (at E3.5), with ICM and cavity forming at the opposite sides. ICM cells are assumed to interact with all the other ICM cells (e.g., by protrusions), which is implemented by a global potential without explicit cutoffs (iii). As a result, PrE precursors migrate away from the EPI core and form the PrE layer at the surface of the ICM, facing the cavity. With this rule, the model predicts that if the TE were removed at the blastocyst stage, in the isolated ICM, the EPI would end up surrounded by PrE (S1 Fig and S3 Movie). This is consistent with the experimental observation that in the cluster of mixed EPI/PrE cells, PrE cells migrate to the outer layer surrounding the EPI core [43].
At E4.5, PrE precursors in the EPI core, i.e., cells surrounded by more than 3 non-PrE precursors, undergo apoptosis to ensure that failures in lineage segregation are not incorporated into EPI development.
To compare with experimental results from 3D blastocysts, we used simple scaling relationships, converting between 2D and 3D (see Materials and methods). We have also validated our approach in 3D simulations (S2 Movie). All the major steps were the same as in 2D, with 1 modification: In 2D, TE cells would always have 2 nearest TE neighbors. We identify these 2 TE cells as nearest neighbors if they are within a certain distance. However, in 3D, this approach fails as one may obtain cell centers within a cell diameter that are not nearest but next-nearest neighbors. To account for this and to find the list of “true” nearest neighbors, we have developed a method that separates nearest from next-nearest neighbors. We evaluate if a potential nearest neighbor is closest to the given cell—and thus included as its true nearest neighbor—or if it is closer to another cell in the neighborhood and thus not counted as nearest neighbor (see Materials and methods for details). While this neighborhood assignment is necessary for the stability of the TE in the 3D model, it is not sufficient, as without polarity the TE cells would collapse into a clump, and cavity cannot be formed.
In order to quantify the importance of each of the rules, we specified the “successful” configuration of the blastocyst at E4.5 to be the one in which (i) TE cells are segregated from ICM cells and form a shell; (ii) the cavity is formed, and the ICM is positioned at one side of the cavity; (iii) ICM cells segregate in 2 distinct layers with the PrE positioned between the cavity and the EPI cells; (iv) and no isolated EPI cells are in the PrE layer nor isolated PrE in the EPI. By comparing the outcome of our simulations to the criteria above, we quantified the fraction (out of 200 simulations for each condition) of “successful” in silico blastocysts (Fig 3b). Representative screenshots from a successful simulation are shown in Fig 3a and in S1 Movie.
We found that with all 4 rules in place, the success rate is high (79%, Fig 3b, see also S2 Movie), suggesting that these rules together are sufficient for development of blastocysts up to E4.5. We also challenged these 4 rules in 3D simulations and found that they were sufficient to generate 3D blastocysts (S2 Movie). For the sake of simplicity, we will compare the impact of specific perturbations in these rules using 2D simulations. In 2D simulations, the fraction of ICM/total cells was 39 ± 12%, and the EPI/ICM fraction was 44 ± 18%, both of which are in a good agreement with experimental data [48,51,53] (Fig 3c and 3d). Of the 21% failure in our simulations, about 20% occurred when—as a result of stochastic update—the fraction of ICM cells (ICM/total cell number) was low. As a result, there were not enough cells available to close the PrE layer, resulting in a PrE error. In about 1% of the cases, the TE broke, either due to failure of maintaining contacts between the surrounding TE cells or right after polarities have been added at E3.0. This occurs if the embryo is in a “tight configuration” in which adding another ICM cell disrupts the shell of TE cells. This error, we believe, is attributed to our choice of potential and noise parameters and might happen even more rarely if the parameters are fine-tuned.
In successful cases, embryos transited through a salt-and-pepper pattern, eventually separating PrE from EPI. To what extent this pattern is salt-and-pepper, i.e., how big are the regions with the same cell types, depends on several factors. The longer the range of FGF4 signal, the larger are the patches of the same cells; on the other hand, the size of the patches is also increased if the differential adhesion molecules are expressed at the same time as cell specify (as is assumed in our model). While visually we do observe patches of different sizes in published data, validation of this aspect of our model will require single-cell quantification of 3D imaging.
To quantify the role of polar interaction, we “switched off” the polarity by setting the attraction factor for all cells to be the same as for undetermined ICM cells (S = 0.6). Without polarity (ΔPolarity case) all cells clustered together; consequently, there was no cavitation and no characteristic shell-like layer of TE cells forming (Fig 3b). These results agree well with the observations in mouse mutants and knockdowns targeting polarity pathways: homozygous mutation in downstream regulator of Yap/Taz signaling, Tead4 [60,62]; chemical inhibition of RHO-ROCK signaling (required for apical-basal polarity), knockdown of Pard6b (a component of PAR-aPKC) by RNA interference (RNAi), disturbing the apical complex aPKC/PAR6 by small interfering RNA (siRNA), downregulating aPKC/PAR3 by injecting double-stranded RNA (dsRNA), or Prickle2 mutants [72–76]—all result in severe polarity defects (including the absence of tight junctions), and all fail to form blastocoel.
Elimination of the second rule can be carried out by modulating the FGF concentration either up or down. As expected, low FGF concentration (−FGF in Fig 1 and S3 Movie) in our model resulted in no PrE formation and an ICM consisting of only EPI at E4.5. These cells were found in a clump consisting of several layers in one side of the blastocyst. This spatial configuration of EPI cells is in agreement with the observed FGF4- and FGFr2- mutants [40,77–80].
Also as expected, the maintenance of a constant FGF/ERK on state (+FGF in Fig 1 and S4 Movie) resulted in ICMs composed solely of PrE, consistent with the experimental results from introducing an excess amount of FGF [27,40,41] (Fig 3d). As a result of stronger adhesion between EPI cells compared to adhesion between PrE cells, the ICM cells clump more in the “EPI only” (low FGF) case compared to the “PrE only” (high FGF) case.
The clumping of the “PrE only” cells is in disagreement with the experimental observation by Yamanaka et al. [27] in which PrE cells are positioned on one side of the blastocyst in 1 layer lining the TE. This disagreement is likely because in our model, the difference between PrE and EPI cells is limited to differences in adhesive properties and does not include the reported apical-basal polarity of the PrE cells [18]. Adding polar interactions to the PrE layer in our model will disfavor “clumping” and make PrE cells line along the TE layer. While polarity of the PrE may add to the robustness of the blastocyst patterning, we chose not to include it into the current model as, within the criteria for success we specified, it does not seem to be necessary for the successful development of the “wild-type blastocyst.”
Noticeably, the failure rate is close to 100% when the differential adhesion between the Gata6 and Nanog positive cells is neutralized (by setting S = 0.5 for all the ICM cells) (Fig 3b, ΔDifferential adhesion case, see also S5 Movie). At E4.5, PrE progenitors remained distributed in a salt-and-pepper pattern; consequently, a considerably higher fraction of the PrE progenitors underwent apoptosis. The ICM/total cell fraction in this case was 39 ± 11%, and the EPI/ICM fraction increased to 57 ± 17%. Deletion of a number of adhesion molecules is known to produce failures in PrE and EPI segregation [47,81–84]. Inhibition of the polarity determinant aPKC [53] at the mid-blastocyst stage results in a failure of PrE/EPI segregation; the increase in inappropriately localized Gata6 cells results in an increased rate of apoptosis within this population, leading to a PrE:EPI ratio of 1:1, which is within the uncertainty of our results (Fig 3).
Deletion of the fourth rule (ΔApoptosis) resulted in 13% of embryos with a PrE precursor positioned deep within the ICM (S7 Movie). As not only misplaced PrE are likely to undergo apoptosis [42], we tested and found no significant differences in our results when we included up to 20% apoptosis in EPI cells (see S5 Fig). Despite differential adhesion and letting the system reach the equilibrium configuration, those cells were trapped in a local energy minima and could not move towards the cavity. The number of mispositioned PrE cells and, consequently, the rate of apoptosis became higher if the system did not reach equilibrium. As it is not known if the ICM cells are in equilibrium or not, our results suggest that 15% error is the lower bound estimate of how frequently differential adhesion fails to segregate PrE from EPI.
To further validate our model, we asked if it can reproduce results of classical scaling experiments [6,67,71,85,86] in which single cells from the 2-cell stage embryo were shown to develop into blastocysts, albeit of half the size and at a lower success rate. Dividing the embryo in half at any time point up to the 8-cell stage resulted in “successful” embryos at E4.5 in about 59% of cases (Fig 4a and S8 Movie). Furthermore, halved embryos were 50% smaller (66 ± 3 cells) than the unperturbed ones (132 ± 3 cells). We also observed a 20% increase in failure rate in blastocyst formation. In our simulations, that was predominantly due to PrE error as a result of a smaller ICM and the resulting fluctuations in the ratio of PrE to EPI. In cases in which the PrE/EPI ratio is smaller than in the unperturbed embryo, there are too few PrE cells to form a layer lining EPI core, and EPI cells tend to intercalate into the PrE layer resulting in a PrE error (Fig 4b). Our model also predicts that the rate of apoptosis in successful half embryos will increase, as abolishing the apoptosis rule increases the number of failures from 15% (Fig 3b) to about 23% (Fig 4b). This is related to the increase in configurations with PrE error discussed above. The model is also consistent with the recently reported scaling results from aggregating two 8-cell stage embryos [71], see Fig 5 and S9 Movie. Thus, without any parameter adjustment, our in silico results were in complete agreement with the scaling experiments and allow us to ask which of our rules is responsible for the scaling properties of the mammalian blastocyst. As 3 out of the 4 rules (Rules 2, 3, and 4) are conditional on FGF/ERK signaling—apoptosis of PrE and differential adhesion are only possible once ICM differentiated in PrE and EPI—we asked whether stage-specific signaling competence could account for scaling.
In these in silico scaling experiments, we have kept the timing of ERK activation unchanged (at E3.0), which would mean that the timing of ERK activation is set at fertilization, either based on expression of the receptor or a rate-limiting factor in the pathway. When we moved ERK activation forward in time, which in effect means delaying the salt-and-pepper pattern, the blastocyst did not fully resolve the salt-and-pepper pattern by E4.5. The model also predicted that delaying ERK signaling should decrease the fraction of PrE cells (See Fig 3D “Delay FGF4” and S10 Movie). To validate this experimentally, we have performed embryo aggregation experiments with and without a potent inhibitor of Mek (PD0325901) [36,87], henceforth referred to as Meki, the kinase that responds to FGFR activation and phosphorylates ERK (Fig 6). Because of the high level of inherent stochasticity in the model and experiments, we chose to prioritize statistically significant results. The identification of a double positive (DP) fraction (through k-mean clustering, see Materials and methods) showed large fluctuations between repeat experiments; the fraction of PrE on the other hand was very robust, so we decided to focus on quantifying this cell type as an indicator ICM patterning. To set the time of FGF/ERK competence, we cultured embryos in Meki for 24 hours and then released them from the signaling block.
In line with the model predictions, we did observe a decrease in fraction with PrE cells proportional to the duration of the ERK inhibition. Embryos were cultured for 56 hours following manipulation, to an in vitro equivalent of E4.5. During this time window, exposure to FGF/ERK signaling was manipulated in 24-hour intervals. Complete inhibition of Mek for the entire 48-hour period resulted in embryos that were entirely EPI (Figs 6b, 6c and 7b), and this is consistent with previous observations [27,71]. However, when embryos were treated for 24 hours (E2.5–E3.5) with Meki and then released from the block, PrE cells were partially recovered, but their fraction was significantly smaller than in the untreated case. Similar transient inactivation experiments have produced a variety of results [27,71] that generally support this observation but without statistical analyses of single-cell quantitation.
We also found that the duration of FGF4/ERK activation, or the point in time in which the pathway becomes competent for signaling, delimits the capacity of the ICM lineages to scale (Figs 6 and 7). Quantification of the relative level of PrE induction (Fig 6B and S4 Fig) indicates that normal aggregates maintain constant ratios of EPI/PrE and that delaying ERK activation with Meki resulted in a reduction in PrE specification. When normal and aggregated embryos are pooled, the quantitative reduction in PrE specification was statistically significant by nonparametric rank-sum test.
In addition, we observed, that based on total cell numbers, the embryos scaled but not in a perfectly linear fashion. We found that the size of the aggregated embryos increased significantly (Fig 7a), and they contained correctly proportioned ICMs, although we did observe a significantly higher ratio of TE to ICM in triplets (Fig 7c). We believe this could be a result of small differences in the relative number of founder TE cells in the triple aggregates, which proliferate at twice the rate of ICM cells [67] and could amplify these differences 2 days later. We also noticed that our embryos contain slightly lower cell numbers than the recently reported aggregation experiments [71], but we imagine these difference could be due to strain differences—in this study, we used inbred C57BL/6, whereas Saiz et al. [71] used the outbred CD1 strain.
Taken together, our attempts to delay ERK activation with Meki combined with embryo scaling experiments suggest that the accumulation of FGF4 and/or an important, limiting downstream signaling component accumulates from fertilization to the point at which this pathway can be activated. Notably, the fixed timing of cavitation, reported by Korotkevich et al. [88], suggests that the timing of TE/ICM differentiation may also start at fertilization. In our model, the timing of TE/ICM differentiation is flexible and only requires a defined time point when polarity is defined.
In this paper, we conceptualized our current knowledge of preimplantation development in a set of simple rules to capture the core modes of regulation, within and between cells, sufficient for successful embryonic development. The robustness of the early preimplantation development is conserved across species. The robustness and remarkable similarity in spatiotemporal patterning emerge despite stochasticity and species-to-species difference in core regulatory components [89,90]. Here, we have shown that 4 cell-based rules can explain how form can be conserved. As rules are not based on specific molecules, it explains how conservation can occur in the presence of variations in the expression and function of core members of the gene regulatory network.
We found that the in silico model, based on only 4 rules, could successfully reproduce a number of the nontrivial quantitative observations:
In addition to these quantitative observations, our model also suggests that specific criteria within the model may be responsible for different aspects of blastocyst development.
Without polar interactions we could not form the cavity. These results are in a good agreement with the experiments reporting on the consequences of strong polarity defects [72–76]. In earlier in silico models, which did not include polarity, the cavity was introduced by hand and was assumed to grow and create a positive pressure on TE thus driving their cell division [59,58]. In contrast, in our model, the growing cavity is a consequence of dividing TE cells, which form a shell-like layer due to polar interactions. While our results do not rule out the osmotic expansion of the blastocyst, they argue that the expansion by TE proliferation should be considered on equal footing. It is hard to delineate which of the 2 is a driving mechanism as the 2 are tightly coupled. First, even if the blastocyst expands by TE proliferation, the cavity’s osmotic pressure should be maintained at homeostasis. Second, drugs inhibiting TE ion channels do not solely act to decrease blastocyst expansion but also perturb TE metabolism [91] and proliferation. To maintain homeostasis during blastocyst expansion, it is likely that the 2 mechanisms act in tandem, feeding back on each other.
When seen from the perspective of one of the markers, e.g., Nanog in EPI cells, the reduced scheme of FGF/ERK signaling (Fig 2 and S2 Fig) can be generalized to a Turing-like patterning mechanism. This mechanism is known for patterning of animal fur, e.g., emergence of black spots in leopards, and is often summarized as “local amplification and global inhibition.” In the case of ICM cells, the “local amplification” results from Nanog intracellular positive feedbacks, whereas the “global inhibition” is realized by Nanog cells secreting FGF4 and inhibiting Nanog in neighboring cells (S2 Fig). Thus, as long as FGF4 is produced in predetermined ICM cells, the pattern of intermixed cell types will automatically emerge. However, when we sampled early stages of blastocyst formation, we found that the patterning of the early ICM (at E3.5) was similar, but not identical, to the recently reported experiments on quantification found in Saiz et al., 2016. In particular, we never detected EPI cells induced in the absence of PrE cells (S6A Fig).
Although recent single-cell RNA sequencing data [39,92] suggests that undetermined ICM cells do express FGF4, they appear to do so to a lesser extent than EPI cells, and this was not accounted for in our original model. We therefore decided to test how this may influence our modeling results; we modified the model to make undetermined ICM cells contribute half as much FGF4 as determined EPI cells. We found that the fractions of cells at E4.5 did not change (S6 Fig), and that we generated normal blastocysts, indistinguishable from our previous simulations. However, this slight modification now recovers the subpopulation of embryos with EPI but no PrE at E3.5 reported by [71] (S6A Fig). Thus, our model not only generated correctly patterned blastocysts, but now also reproduces the earliest phases of lineage segregation with higher fidelity. These simulations make a new, experimentally verifiable prediction that unsegregated ICM cells express lower, but functional, levels of FGF4 than differentiated EPI. It also further demonstrates the importance of self-regulating dynamics in patterning the blastocyst, demonstrating that the final result is not sensitive to initial conditions. The capacity to vary initial conditions without impacting on the final result also provides insight into parameters that can be manipulated in evolution.
The FGF/ERK pathway coupled to Nanog/Gata positive feedback was proposed to control PrE/EPI cell proportions in 2 other models, one exploring isolated ICM patterning [57] and the other PrE/EPI specification in embryonic stem cells (ESCs) [93]. Neither model captures the dynamic geometry of the growing embryo. Our model incorporates a conceptualized FGF/Nanog/Gata feedback circuit into embryonic development, showing that a form of this mechanism can function in the highly dynamic environment in which cells divide and move due to differential adhesion.
As a consequence of this “local amplification, global inhibition,” PrE and EPI cells in the model are capable of changing their identity at all times during the FGF/ERK competence window, i.e., the pathway remains active. While in ES cell culture and in FGF4 manipulation experiments by Bessonnard et al. [57] and Yamanaka et al. [27] there is a window in which cells are observed to change their identity, this does not seem to occur frequently under physiological levels of FGF4 in unperturbed blastocysts [70]. It is, however, not known if these changes do not occur because cells are not capable of switching or because they reach an appropriate configuration in which the switching is not required.
Our simulation suggest that the latter explanation is correct, and this explains why the cells of the blastocysts remain competent to undergo regulative transformations in response to signaling manipulation while maintaining an apparently deterministic trajectory in normal development. This, and to what extent the number of FGF4-secreting neighbors determines the fate of the cell, can be tested through targeted laser ablation of cells such as to shift the balance between FGF4-secreting and nonsecreting cells in the neighborhood.
Simulation results suggest that differential adhesion alone can often (62% of embryos) be sufficient for correct spatial arrangement of PrE and EPI cells. It is believed that the position-dependent apoptosis of Gata6 cells may play an important role in resolving the occasional positional errors: Plusa et al. [4] reports that the isolated Gata6 cells deep inside the ICM apoptose 6-fold more often than the correctly positioned PrE cells facing the blastocoel. One possible mechanism for the positional difference in apoptosis is if the concentrations of the cytokines LIF [50,51] and PDGF [52,53]—known to promote PrE survival—are lower inside the ICM than at the junction of the EPI, PrE, and blastocoel. While this still remains to be tested experimentally, the current knowledge on LIF is in line with this hypothesis. LIF secreted from TE is likely to accumulate to higher concentrations at the PrE/blastocoel boundary as there are more TE cells facing the blastocoel, and LIF produced by these cells can diffuse freely until it reaches the PrE layer.
Our observations indicate apoptosis is important for the robustness of pattern formation. The model predicts that apoptosis becomes increasingly important as the difference between adhesive properties is reduced. The difference in adhesion properties could vary in different genetic backgrounds and also must vary in time as differentiation progresses. In cases in which there is flexibility in adhesion, there would be a greater requirement for apoptosis in proofreading.
While we have shown that differential adhesion in combination with apoptosis are sufficient for proper lineage segregation, a number of other mechanisms may contribute to robustness in the segregation process. First, it has been suggested that cellular movements involve not only passive but also active mechanisms, associated with cell protrusions [42]. In line with this, in silico studies that did not consider apoptosis, showed that failures in segregation can be reduced if differential adhesion is complemented by directed cell movements [59]. Second, PrE differentiation occurs in stages that include an uncommitted and biased state within the ICM and committed, PrE lining the blastocoel. Thus, when tested in heterotrophic grafting experiments, early PrE progenitors within the ICM were competent to make EPI, while PrE progenitors that line the blastocoel cavity are only able to participate in endoderm development [29].
In our simulations—in which the segregation of the PrE and EPI is based on differential adhesion—we often, in about 50% of simulations for wild type, observe PrE forming 2 layers. While this is obviously in contrast with the observed single PrE layer in real embryos, to keep the model simple, we choose to count them as a success as long as the PrE layers seal EPI core from blastocoel. Expanding the model to include polar interactions between PrE cells lining the cavity would ensure a single PrE layer and provide a contiguous barrier between the EPI and the cavity [53].
Scaling presents a fascinating example of the robustness in embryo development, and the experimental manipulation of this phenomena served as an important validation step for the model. The close match between the experimental observations and our simulation predicted that timing of FGF/ERK signaling may be the key parameter for controlling the scaling outcome. With the 4 rules in place, the embryo would scale when divided in half or doubled prior to compaction. However, we only observe this property if cell fate specification and emergence of the salt-and-pepper pattern—attributed to the FGF/ERK signaling—take place at the same time counted from fertilization. This implies that competence for FGF/ERK signaling is primed for activation from fertilization. We have validated this prediction by showing that the ratio of PrE in the ICM decreases upon transient inhibition of FGF4 signaling both in the model and in cultured embryos.
The notion that we observe normal development with 4 rules that are largely independent of the initial gene regulatory network is particularly relevant to the current debate about the extent to which stochastic gene expression governs the initiation of blastocyst development. Our model demonstrates that initial differences in stochastic gene expression are not a necessary prerequisite for the generation of 3 distinct lineages. Instead, differentiation emerges based on the responses of a cell to its local environment, as interpreted via differential proliferation, adhesion, and gene expression. The existence of a set of rules that allow for blastocyst formation as long as a few simple conditions are satisfied could be an enabler of stochastic variation. It also could explain how mammalian development can allow for the fundamental changes in the gene regulatory network that have been observed when single-cell sequencing data has been compared between mouse and human [89,94,95].
For simplicity, we implemented our model in 2D. Cells are modeled as circles with the radius set to one.
To compare with experimental results from 3D blastocysts, we used simple scaling relationships converting between 2D and 3D. Thus, the number of TE cells, NTE3D, placed on the surface of the sphere would correspond to NTE2D=πNTE3D in 2D. Similarly, for cells in the bulk: NEPI/PrE2D=(34NEPI/PrE3Dπ)2/3.
The interaction between 2 cells is given by the following potential, V (see Fig 2c):
V(d)=exp(−d)−S exp(−d/β)
(2)
where d is the distance between the cells, S is the attraction factor given by Eq 1, and β is the parameter controlling the range of the attraction.
This potential assures repulsion at short distances, i.e., 2 cells separated by a distance less than 1 cell diameter (d < 2) will repulse from each other. On the other hand, if cells are separated by more than 1 cell diameter (d > 2), they will be attracted to each other. In the simulations, we set a distance cutoff, setting potential to 0 for all cell pairs that are further away than 5 cell radii. S and β are chosen to produce a tight packing of cells with minimal overlap (see configuration of cells in Fig 2a, with S = 0.6 and β = 5). The choice of these parameters as well as the form of the potential are not important for the model outcome as long as the condition above is satisfied. While we keep β = 5 fixed throughout the simulations, the S will capture lineage-specific differences in adhesive properties and is thus a lineage-specific parameter. Prior to first cell-lineage decision at E3.0, cells are assumed to have the same adhesive properties and thus the same strength of attraction, S = 0.6.
The motion of the cells is described by the overdamped equation of motion:
dx/dt=−dV/dx+η
(3)
where dV/dx is a x-projection of the resulting force from all of the pairwise cell—cell interactions.
To ensure that system reaches equilibrium, we add the noise term η. In the simulations, this is implemented by adding a random number from a normal distribution with a mean of 0 and a standard deviation of 10−3 at every time step. The equation is integrated numerically using Euler integration scheme. The y-position is determined in a similar way.
The polarity of the TE cells is assumed to be affected by the orientation of the polarities in neighboring TE cells such that they tend to point in the same direction. This is well described by a pairwise polarity potential Vp = −cos θi,j, where θi,j = αi − αj measures the angle between the polarities of the i,j TE neighbors.
The orientation of the polarity is described by an angle α and, similarly to the equation above, the change in polar orientation is given by dαdt=−0.1dVpdα+ηp where the prefactor of 0.1 makes the changes in polarities happen slower compared to the changes in the positions. This is necessary for the stability of the system. The noise term is implemented in the same way as above, with the only difference being that the random numbers are multiplied by π.
The TE is noise sensitive. With the chosen standard deviation 10−3 on the noise parameter, the simulation is very sensitive to the factor 1.4 in Eq 1. If this factor is increased to 1.5, the TE repulses too much when a new TE cell is added. On the other hand, if the factor is decreased to 1.3, it becomes too weak and cannot keep the TE together at E4.5. Thus, here, we apply the maximum tolerated noise to the system. Less or even no noise is also acceptable, and it allows the factor in Eq 1 to be decreased.
We started with 1 cell. In contrast to an earlier model where the growth of the blastocysts was driven by the growing blastocoel, in our simulations, the blastocyst grows as a result of cell division: A cell is randomly selected to undergo division, and the daughter cell is positioned between the mother cell and the nearest neighbor cell. In real blastocysts, prior to division, cells gradually increase in size, allowing other cells to readjust their position such that the system is near equilibrium at all times. For simplicity, we chose to keep cell size constant; however, that results in strong perturbation of the equilibrium during the simulated cell division. To assure that the configurations of the simulated blastocysts are not affected by this, we allowed enough time for the system to relax before the next cell division. In 3D, one can significantly speed up the simulations by introducing a new cell in the center of the 3 nearest neighbors, as this is closer to the minimal energetic configuration and reduces the number of relaxation steps.
During implementation of Rule 2, we first pick up random cells to divide among the undetermined ICM’s. At the division, the likelihood of a cell to convert to PrE is proportional to the fraction of high FGF4 (EPI or undetermined ICM) cells in the neighborhood (P(PrE)=# of high FGF4 neighbors# of ICM neighbors); conversely, the likelihood to convert to EPI is P(EPI) = 1 − P(PrE). After all cells have specified, a random ICM is chosen to divide, and at the division, the same rule applies as for undetermined ICM’s.
Embryos used in this study are inbred C57Bl/6NRj (Janvier Labs, France).
Mice were maintained in a 12-hour light/dark cycle in the designated facilities at the University of Copenhagen, Denmark. Embryo donor females underwent super-ovulation treatment following a standard protocol: intraperitoneal injection (IP) of 5 IU PMSG (Sigma) per female and IP injection of 5 IU hCG (Chorulon, Intervet) 47 hours later, followed by overnight mating with C57Bl/6NRj stud males. The following morning, females were monitored for copulation plug formation. Embryos were considered E0.5 on the day of plug detection. Animal work was carried in accordance with European legislation and was authorized by and carried out under Project License 2012-15-2934-00743 issued by the Danish Regulatory Authority.
Embryos were obtained at 8-cell morula stage by washing E2.5 oviducts with M2 medium (Sigma). In order to remove the zona pellucida, morulae were briefly incubated in Acid Tyrode’s solution (Sigma) at RT and then washed in M2 medium.
To generate aggregates, embryos were placed in pairs or triplets in aggregation microwells made with an aggregation needle (BLS) on Petri dishes in KSOM medium (LifeGlobal) drops. Drops were overlaid with mineral oil (Nidoil, Nidacon). Single embryos were placed alone as control. KSOM was supplemented with 0.1% BSA (Sigma) to avoid embryos adhering to the plastic. Embryos were cultured at 37°C, 5% CO2 and 90% relative humidity. For MEK inhibition treatment, 1 μM of PD 0325901 (PZ0162, Sigma) was diluted into KSOM. Wild-type embryos were generated by culture in KSOM. 24-hour and 48-hour treated embryos were generated by culture in KSOM with PD 0325901 for 24 hours and 48 hours, respectively. The data were collected over 4 repeat experiments.
Fifty-six hours after aggregation, embryos at E4.5 were fixed in 4% PFA solution for 15 minutes at room temperature. Afterwards the embryos were stained as previously described [96]. The primary antibodies used were: anti-Nanog (eBioscience, 14–5761; 1:200), anti-Cdx2 (Biogenex, MU392A-UC; 1:200), and anti-Gata6 (R&D, AF1700; 1:100). Embryos were imaged in an Attofluor chamber (ThermoFisher) on a 25-mm glass coverslip using 10x magnification on a Leica TCS SP8 confocal microscope.
We used ImageJ to manually track positions of the nuclei in single cells. Positions were saved and intensities for each fluorescence channel at each position were processed by custom-built Matlab scripts (available upon request). For each of the channels, we have used the mean intensity of the 5 x 5 x 3 voxel as a readout for single cell. To filter out the noise, cells with the Dapi intensity below 1 were removed. To differentiate between TE and ICM cells, for each of the cells, we ranked the intensities of Cdx2, Nanog, and Gata6. Cells where Cdx2 ranked first, were classified as TE cells. We validated that the identified TE cells localize to the periphery of the embryo. We classify the ICM cells as described in Saiz et al. [71]: First, we performed k-means clustering (by Squared Euclidean distance metric, Matlab built in function) on the log(Gata6) and log(Nanog) into 3 clusters with 10 repetitions on all data pooled together. Second, we classified high Gata6 and high Nanog cells as DP cells, high Gata6 and low Nanog as PrE cells, and low Gata6 and high Nanog as EPI cells. See S3 Fig for the results of the clustering.
As the distributions of the analyzed properties were clearly far from normal, we used nonparametric rank-sum test to estimate the statistical significance of the difference in medians.
|
10.1371/journal.pmed.1002920 | Risk and protective factors for child development: An observational South African birth cohort | Approximately 250 million (43%) children under the age of 5 years in low- and middle-income countries (LMICs) are failing to meet their developmental potential. Risk factors are recognised to contribute to this loss of human potential. Expanding understanding of the risks that lead to poor outcomes and which protective factors contribute to resilience in children may be critical to improving disparities.
The Drakenstein Child Health Study is a population-based birth cohort in the Western Cape, South Africa. Pregnant women were enrolled between 20 and 28 weeks’ gestation from two community clinics from 2012 to 2015; sociodemographic and psychosocial data were collected antenatally. Mothers and children were followed through birth until 2 years of age. Developmental assessments were conducted by trained assessors blinded to background, using the Bayley-III Scales of Infant and Toddler Development (BSID-III), validated for use in South Africa, at 24 months of age. The study assessed all available children at 24 months; however, some children were not able to attend, because of loss to follow-up or unavailability of a caregiver or child at the correct age. Of 1,143 live births, 1,002 were in follow-up at 24 months, and a total of 734 children (73%) had developmental assessments, of which 354 (48.2%) were girls. This sample was characterised by low household employment (n = 183; 24.9%) and household income (n = 287; 39.1% earning <R1,000 per month), and high prevalence of maternal psychosocial risk factors including alcohol use in pregnancy (n = 95; 14.5%), smoking (n = 241; 34.7%), depression (n = 156; 23.7%), lifetime intimate partner violence (n = 310; 47.3%), and history of maternal childhood trauma (n = 228; 34.7%). A high proportion of children were categorised as delayed (defined by scoring < −1 standard deviation below the mean scaled score calculated using the BSID-III norms from a United States population) in different domains (369 [50.5%] cognition, 402 [55.6%] receptive language, 389 [55.4%] expressive language, 169 [23.2%] fine motor, and 267 [38.4%] gross motor). Four hundred five (55.3%) children had >1 domain affected, and 75 (10.2%) had delay in all domains. Bivariate and multivariable analyses revealed several factors that were associated with developmental outcomes. These included protective factors (maternal education, higher birth weight, and socioeconomic status) and risk factors (maternal anaemia in pregnancy, depression or lifetime intimate partner violence, and maternal HIV infection). Boys consistently performed worse than girls (in cognition [β = −0.74; 95% CI −1.46 to −0.03, p = 0.042], receptive language [β = −1.10; 95% CI −1.70 to −0.49, p < 0.001], expressive language [β = −1.65; 95% CI −2.46 to −0.84, p < 0.001], and fine motor [β = −0.70; 95% CI −1.20 to −0.20, p = 0.006] scales). There was evidence that child sex interacted with risk and protective factors including birth weight, maternal anaemia in pregnancy, and socioeconomic factors. Important limitations of the study include attrition of sample from birth to assessment age and missing data in some exposure areas from those assessed.
This study provides reliable developmental data from a sub-Saharan African setting in a well-characterised sample of mother–child dyads. Our findings highlight not only the important protective effects of maternal education, birth weight, and socioeconomic status for developmental outcomes but also sex differences in developmental outcomes and key risk and protective factors for each group.
| Child development in early childhood lays a foundation for lifelong learning.
Risk and protective factors for child development are known to include many issues faced by children growing up in low- and middle-income countries.
Studies indicate a difference between boys and girls in terms of impact of factors influencing development, but these have not been evaluated in a sub-Saharan African context.
We assessed child development at 2 years of 734 children in the Drakenstein Child Health Study, Western Cape, South Africa.
We assessed potential risk and protective factors identified from prior literature to impact child development.
We found a number of important risk factors that contributed to poor developmental outcomes in children in this cohort.
Boys appear to be at higher risk of poor developmental performance in a high-risk environment.
Key protective factors include mothers having at least some secondary school education, better home circumstances, and healthy birth weight, and key risk factors include maternal anaemia in pregnancy, poor maternal health (such as HIV), and maternal mental health problems.
Child sex interacts with the associations between key protective and risk factors and developmental outcomes.
Understanding the related and interacting roles of factors reported in this study may inform integrated intervention policy design and implementation for supporting development in high-risk environments.
| Approximately 250 million (43%) of children under the age of 5 years in low- and middle-income countries (LMICs) are at risk of poor developmental outcomes [1]. Child development takes place as an ongoing biological and psychological process influenced by the environment, caregivers, community, and society. Key risk factors known to affect child development may be broadly grouped into those affecting (1) the wider community and environment in which the child and family live, often termed the social determinants of health [2] (poverty, lack of access to education, environmental stressors, poor water and sanitation); (2) the physical health of the caregiver [3] (maternal illness and nutrition) and the child (malnutrition, low birth weight, infections); (3) and maternal psychosocial health [4] (maternal depression, substance use, and intimate partner violence [IPV]). Conversely, protective factors are those that foster resilience and allow children to overcome adversity. These include (1) breastfeeding and good nutrition, (2) clean and safe living spaces, (3) and nurturing environments and healthy parents [3,5]. Improved understanding of key factors associated with neurodevelopmental delay and those promoting resilience is necessary to ensure children achieve their developmental potential. Children exposed to multiple risk factors have a greater likelihood of poor adult health and well-being [6].
There is increasing recognition that boys and girls may be sensitive to their environments in different ways. Previous work has further indicated an emerging theme of sex-dependent fetal programming in response to prenatal stress of various types, whereby girls may respond to challenge in more anxious and reactive ways and boys respond in less reactive but more aggressive ways [7]. More specifically, studies in high-income countries have reported a difference in developmental outcomes between girls and boys [8], and recently, a multicountry evaluation highlighted large disparities between sexes in the Asia-Pacific region, with girls performing better than boys on composite developmental score in four out of six countries [9]. This supports the hypothesis that boys and girls may be differentially affected by risk and protective factors impacting their future development [10]. However, very little of this work has been done in LMICs (and has not been explored in sub-Saharan Africa), where there is amplified potential of multiple physical, psychosocial, and environmental factors to interact in complex ways and to which boys and girls may be vulnerable in different ways.
Data are lacking from LMICs that comprehensively investigate the development of children including cognitive, language, and motor outcomes. Although recently there have been some trials examining risk factors for development [11], global estimations tend to use proxy measures, such as poverty and stunting, as measures of child development [1] because of insufficient data directly measuring these outcomes and risk factors in LMIC contexts. This gap in the literature highlights the need to report broad, multifactorial (social, clinical, psychosocial) data measuring risk and protective factors for early child development and appropriate, directly measured developmental outcomes in children living in these settings. Expanding our understanding of which risk factors lead to poor outcomes and which protective factors build resilience is critical to improving disparities, particularly in low-resource settings [12,13]. The aim of this study was to investigate the range of risk and protective factors that affect early childhood developmental outcomes and to determine sex differences in the impact of such factors in a South African birth cohort focussed on early child health and development.
The Drakenstein Child Health Study (DCHS) is a multidisciplinary population-based birth cohort study investigating the early-life determinants of child health and development. The study is located in Paarl, a periurban area, 60 km outside of Cape Town in the Western Cape of South Africa [14]. It is a stable, low-socioeconomic community comprising approximately 200,000 people, characterised by a high prevalence of a range of health risk factors such as depression, childhood trauma, IPV, and poverty [15]. The DCHS is representative of many periurban regions in South Africa, as well as in other LMICs. The majority of the population accesses healthcare in the public sector. Pregnant women were recruited from two public sector primary healthcare clinics, one serving a predominantly mixed ancestry population (TC Newman) and the other serving a predominantly black African population (Mbekweni).
Pregnant women were enrolled into the DCHS between 20 and 28 weeks’ gestation while attending routine antenatal care and are being prospectively followed through childbirth and early childhood until children are 10 years of age. Recruitment was unfiltered, and eligibility criteria included (1) attendance at one of the two study clinics, (2) being at least 18 years of age, and (3) intending to remain in the study area for at least 1 year. Between March 2012 and March 2015, 1,225 pregnant women were enrolled into the DCHS antenatally; 88 (7.2%) mothers were lost to follow-up antenatally, had a miscarriage, or had a stillbirth. In total, there were 1,143 live infants.
All measures were performed as part of the main study, and child development outcomes form a primary aim of the original DCHS [14,15]. Expanded descriptive detail and rationale may be found in the associated Methods paper [16].
Sociodemographic variables were measured using validated questionnaires administered by trained study staff at an antenatal visit at 28 to 32 weeks’ gestation. Sociodemographic variables including household factors (running water, flushing toilet, electricity in home, and household income) and maternal demographics (age at enrolment, any secondary versus only primary education, married or cohabiting with partner, employed, and whether this was the first pregnancy) were collected using an interviewer-administered questionnaire adapted from items used in the South African Stress and Health (SASH) Study [17,18].
Gestational age at delivery was calculated in weeks, based on ultrasound results when these were available and otherwise based on fundal height measurements and maternal report of last menstrual period. Prematurity was defined as <37-week gestational age. Birth weight was abstracted by trained study staff from hospital records at birth and was taken as a continuous measure. Duration of exclusive breastfeeding was derived based on maternal report of feeding practices at birth; 6, 10, and 14 weeks; and 6 and 9 months of child age. Exclusive breastfeeding was defined as occurring until the first maternal report of introduction of solid foods or formula. We included exclusive breastfeeding for 6 months, as per current recommendations by WHO [19].
Maternal haemoglobin was tested for in pregnancy. Haemoglobin levels < 10 g/dL in pregnancy were classified as moderate to severe iron deficiency anaemia as per WHO guidelines [20]. Maternal HIV status was established during routine HIV testing of women in pregnancy as per the Western Cape of South Africa guidelines for prevention of mother-to-child transmission of HIV. Stunting was investigated separately because of the known association with delayed development, but it was not included as a risk factor in the final model, because of the association with the other risk and protective factors [21]. Stunting was defined here as <−2 standard deviations below WHO z-score height for age at 2 years of age.
Alcohol use during pregnancy was assessed using a composite, dichotomous measure (exposure versus no exposure) using the Alcohol, Smoking and Substance Involvement Screening Test (ASSIST) and retrospectively collected data on hazardous alcohol use during pregnancy [22,23]. The ASSIST has shown good reliability and validity in international, multisite studies. Total scores were obtained for alcohol by summing individual items related to maternal alcohol use during pregnancy; a score of >10 indicates moderate to high levels of risk for alcohol problems (reflecting weekly or daily/almost daily alcohol use and negative consequences related to the quantity of alcohol consumed). We categorised children as alcohol-exposed versus not alcohol-exposed. We used the following inclusion criteria to define alcohol exposure in order to optimise numbers: (1) scoring greater than moderate on the ASSIST performed antenatally as per reference Myers et al. [22] and Donald et al. [23] AND/OR (2) 2 or more drinks a week on the neonatal alcohol questionnaire (to further quantify alcohol use on top of the ASSIST) AND/OR (3) 2 or more drinks a week on a retrospective alcohol questionnaire regarding alcohol use during pregnancy (as mothers may be more likely to respond to this than during/immediately after pregnancy). Tobacco exposure during pregnancy was objectively measured by maternal urine cotinine collected antenatally or at birth. Cotinine, a metabolite of nicotine measurable in the urine, was measured using the IMMULITE 1000 Nicotine Metabolite Kit (Siemens Medical Solutions Diagnostics, Glyn Rhonwy, Llanberis, United Kingdom). Mothers were categorised as active smokers during pregnancy if either the antenatal or birth cotinine levels were ≥500 ng/ml [18].
Mothers completed a battery of psychosocial measures, administered by trained study staff, at an antenatal visit between 28 and 32 weeks’ gestation. The IPV questionnaire used in this study was adapted from WHO’s multicountry study [24] and the Women’s Health Study in Zimbabwe [25]. Mothers were asked about exposure to partner behaviour and frequency of occurrence for emotional, physical, or sexual abuse behaviours; when behaviours were experienced ‘many times’, mothers were categorised as exposed to lifetime IPV. The Edinburgh Postnatal Depression Scale (EPDS) [26] was used to measure depression; this scale has been validated for use with pregnant women and in a South African population [27–29]. The EPDS consists of 10 items referring to the past 7 days; a total score was obtained by summing responses for all items, and a cutoff score of ≥13 was used to dichotomise participants into below and above threshold for depression. The Childhood Trauma Questionnaire [30] Short-Form was used to assess abuse and neglect experienced as a child. Each item was responded to on a 5-point scale; a cutoff score of >36 was used to dichotomise mothers into above or below threshold for childhood abuse. The Self Reporting Questionnaire-20 (SRQ-20) is a WHO-endorsed measure of psychological distress [31]. The SRQ-20 consists of 20 items, which assess nonpsychotic symptoms, including symptoms of depressive and anxiety disorders; items are summed, and a cutoff score of >8 was used to dichotomise mothers into below or above threshold for psychological distress [32].
Child development was assessed on all available children across the full cohort at 24 months of age using the Bayley Scales of Infant and Toddler Development (Third Edition) (BSID-III) [33]. The BSID-III is a gold-standard observational measure of development for children from 0 to 42 months. It has been validated for a South African population [34,35] and found to be culturally appropriate without modifications. However, the tool may slightly underestimate delay in this population [35]. The tool measures development by direct observation across five subscales: cognition, receptive and expressive language, and fine and gross motor [36]. These scales were measured by direct observation by a trained physiotherapist and occupational therapist blinded to the child and family risk factors, overseen by a paediatric neurologist with specialist developmental expertise [16]. Quality control and monitoring processes were implemented to ensure accuracy. All data were entered into the BSID-III scoring programme, and the data were exported to Excel. In this analysis, both the raw scores and scaled scores were used. The scaled scores are calculated using a normal US population, scaled to a mean of 10 and standard deviation of 3. We assessed poor developmental outcomes [33] by categorising the scores into ‘delay’ or ‘no delay’, defined by scoring <−1 standard deviation below the mean scaled score (using a cutoff of 7).
Ethical approval was obtained from the Faculty of Health Sciences Research Ethics Committee, University of Cape Town (401/2009), and by the Western Cape Provincial Research committee (2011RP45). Mothers gave written informed consent at enrolment and were reconsented annually for study involvement. Consent was done in the mother’s preferred language: English, Afrikaans, or isiXhosa. When significant developmental delays were identified by study staff, mothers and children were referred to local healthcare services for further assessment and management.
The analyses followed a defined approach that was decided upon prior to running the models, though no prespecified analysis plan exists. Thus, data-driven approaches to analysis were not used at any stage. Early childhood development outcomes’ scores were measured by five developmental domains using the BSID-III: cognitive development, language (receptive and expressive) development, and motor (fine and gross) development. We used a cutoff scaled score of less than 7 (<−1 SD below the mean scaled score) to define developmental delay.
BSID-III raw scores were analysed using linear regression, and developmental-delay data were analysed using logistic regression separately for each domain. Odds ratios are presented for logistic regression models, and regression coefficients are presented for linear regression models with 95% confidence intervals (CIs) for both. We used a missing-at-random approach and generated a complete-case dataset, which omitted records that were missing data in any of the covariates that we considered plus any domain (apart from gross motor, which was not analysed). Multivariable regression models were built using the complete-case dataset in a hierarchical approach. Initial models adjusted for maternal education, child age, and sex forced into the model as known and well-proven factors that impact child developmental outcomes. Subsequently, other covariates were added in blocks following the hierarchical order from sociodemographic and environmental variables to maternal and child physical variables to maternal psychosocial variables based on prior literature [37] as illustrated in Fig 1. Covariates added from each block were selected using a best-subsets variable-selection approach that aimed to minimise Akaike information criterion (AIC; using the ‘gvselect’ command in Stata 14) [38,39]. That is, the model for one outcome considered forced variables first and compared the fit of the model to the fit of the models with all possible combinations of variables from the sociodemographic block. Only the subset of these variables that reduced the AIC were retained for consideration with the next block (this includes an empty set, which would arise if no variables reduced the AIC from the previous model). Collinearity was assessed using variance inflation factors after each block, and then these steps were repeated for the next two blocks of variables. After creating a final model assessing the associations between risk and protective factors and each developmental outcome, we assessed whether including interaction terms between child sex and any variable retained from the best-subsets regression reduced the AIC further, to explore whether child sex altered the association between any of these variables and developmental outcomes. We report both the final model and the model including interaction terms. Finally, we used the same process to assess the outcome of global developmental delay in all domains consisting of delay in cognition, language (combining delay in receptive or expressive language), fine motor, and gross motor.
All analyses were performed with Stata version 14 [40].
Of 1,143 live study births, 409 (35.8%) children were lost to follow-up before reaching 2 years of age or did not complete a BSID-III assessment, because of nonattendance or moving out of the study area. Therefore, a total of 734 children completed BSID-III assessments at 2 years of age between August 2014 and September 2017. There were few differences between those who did and did not complete a BSID-III at 2 years (S1 Table), including higher maternal age (p < 0.01), less prematurity (p < 0.01), and more active smoking in pregnancy (p = 0.01) in those with a BSID-III.
Baseline clinical, demographic, and psychosocial characteristics, stratified by child sex, are presented in Table 1. Of the children with a BSID-III, 48.2% were girls, and there were no differences in sociodemographic variables between sexes (p < 0.05), apart from girls having a higher household income (65.5% had >R1,000 [approximately USD 100] per month versus 56.6% boys, p = 0.01) (Table 1). This sample was characterised by low household employment (24.9%) and low household income (39.1% earned less than R1,000 per month). The majority of households had running water (69.2%) and electricity (95.0%). A minority of mothers were married or cohabiting (40.1%), one-third of babies were born to primigravid mothers (33.0%), and 14.2% of babies were born prematurely (<37 weeks’ gestation). Of the 169 mothers (23.0%) who were confirmed HIV-infected antenatally, one child tested HIV-positive. Breastfeeding rates were low across the cohort, with only 17.1% children exclusively breastfed for 6 months. There was a high proportion of antenatal maternal substance use and other psychosocial risk factors, including alcohol exposure (14.5%), smoking (34.7%), depression (23.7%), lifetime IPV (47.3%), and history of maternal childhood trauma (34.7%).
A total of 369/731 children (50.5%) were categorised as having cognitive delay, 402/723 (55.6%) with receptive language delay, 389/702 (55.4%) with expressive language delay, 169/730 (23.2%) with fine motor delay, and 267/696 (38.4%) with gross motor delay (Table 2). Overall, 405 children from the 734 (55.2%) were classified as having delay in two or more domains; the combination of cognitive and language delay was the most common, with 304 children (41.4%). There were 75 children (10.2%) who were delayed in all four domains.
On raw scores, boys exhibited lower total scores than girls for all domains except gross motor on bivariate analyses (Table 2 and S2 Table), as well as in adjusted analyses in cognition (β = −0.74; 95% CI −1.46 to −0.03, p = 0.042), receptive language (β = −1.10; 95% CI −1.70 to −0.49, p < 0.001), expressive language (β = −1.65; 95% CI −2.46 to −0.84, p < 0.001), and fine motor (β = −0.70; 95% CI −1.20 to −0.20, p = 0.006) raw scores (Table 3). Likewise, boys were at increased risk of delay in all domains except gross motor (Table 2 and S3 Table) (p < 0.05) in bivariate analyses, and this association held when adjusting for confounders in the multivariable models (S4 Table): cognitive delay (adjusted odds ratio [aOR] 1.35; 95% CI 0.95–1.91, p = 0.098), receptive language delay (aOR 2.40; 95% CI 1.68–3.42, p < 0.001), expressive language delay (aOR 2.12; 95% CI 1.49–3.03, p < 0.001), and fine motor delay (aOR 1.92; 95% CI 1.25–2.95, p = 0.003).
At 2 years of age, 19.3% of children were classified as stunted (22.4% boys and 15.8% girls), and stunting was found to be associated with poor cognitive development (β = −1.13; 95% CI −2.10 to −0.15) and with cognitive developmental delay (OR = 1.77, 95% CI 1.18–2.65). As expected, across the group, children who were older tended to perform better in the developmental assessment, particularly in the cognitive and language subscales. However, the age window was very small, so the variance may not be meaningful in this analysis. In the gross motor domain, we investigated sex, education, and child age and we did not see any of the expected bivariate associations and therefore we did not run multiple regression analyses on this outcome variable.
Bivariate analyses (S2 and S3 Tables) and the final model multiple regression results are shown in Table 3, S4 Table, Fig 2, and S1 Fig. We first describe the final model examining associations between risk and protective factors with child developmental outcomes and then explore the interactions with child sex.
On bivariate analysis, protective factors associated with better development and reduced risk of delay comprised having higher maternal education, older child age, a primigravid mother, and better-resourced households (higher household income, flushing toilet, running water). Higher maternal education (any secondary versus only primary) was associated with increased total raw scores for all domains on bivariate analyses and in the final model for cognition (β = 1.70; 95% CI 0.45–2.95, p = 0.008) as well as lower odds of cognitive delay (aOR 0.52; 95% CI 0.28–0.99, p = 0.045) and receptive language delay (aOR 0.51; 95% CI 0.26–0.99, p = 0.045). Children of primigravid mothers had higher expressive language scores (β = 0.93; 95% CI 0.06–1.80, p = 0.036) and lower odds of expressive language delay (aOR 0.57; 95% CI 0.36–0.90, p = 0.016) in multiple regression models. In the interaction models, sex impacted the association between primigravida and expressive language outcomes. We calculated estimates of the interaction effects of primigravida in both boys and girls separately. For girls, the beta coefficient was 2.07 (95% CI 0.82–3.31, p = −0.01), and for boys, it was −2.88 (95% CI −4.64 to −1.12, p = 0.002). This suggests that primigravida status had a positive effect on expressive language for girls but a negative effect for boys in this cohort.
Flush toilets in household were associated with increased total raw scores for expressive language (β = 1.62; 95% CI 0.39–2.86, p = 0.010) in the multivariable regression. Tap water showed an interaction with child sex, and we calculated estimates of the effects of tap water in both boys and girls separately in the interaction model. For girls, the beta coefficient was −3.10 (95% CI −4.70 to −1.49, p < 0.001), and for boys, it was −3.72 (95% CI −5.41 to −2.02, p < 0.001). The effect direction is unexpected and is perhaps caused by the additional inclusion of flushing toilet in the model, although these together reduced the AIC.
On bivariate analysis, key maternal and child physical health risk factors for lower developmental scores or higher odds of delay included maternal anaemia, prematurity, maternal HIV, alcohol or tobacco use during pregnancy, and maternal depression and lifetime IPV; protective factors included higher birth weight. Increasing birth weight was associated with significantly greater scores and lower odds for delay for all domains in the total cohort in bivariate and multivariable models (Table 3 and S4 Table). There was also evidence of an interaction with child sex in cognition and language outcomes. We calculated marginal effects for the interaction of birth weight and cognition, and at a birth weight of 3 kg (mean of cohort is 3.05 kg), the marginal effect was −0.83 (95% CI −1.55 to −0.12, p = 0.022). See Fig 2 for all marginal effects. At lower birth weights, boys did less well than girls (negative coefficients indicate a lower score); and at higher birth weights, boys did better. Likewise, we calculated marginal effects for the interaction of birth weight and receptive language, and at a birth weight of 3 kg, the marginal effect was −1.17 (95% CI −1.77 to −0.56, p < 0.001). See S1 Fig for all marginal effects. As with cognition, at lower birth weights boys did less well than girls.
In the multivariable final model, maternal anaemia in pregnancy was associated with lower cognitive scores (β = −1.38; 95% CI −2.33 to −0.44, p = 0.004), receptive language scores (β = −1.12; 95% CI −1.93 to −0.31, p = 0.007), and expressive language scores (β = −1.11; 95% CI −2.18 to −0.04, p = 0.043). There was some evidence for interaction of child sex in this association. We calculated estimates of the effects of anaemia in both boys and girls separately for receptive and expressive language. For girls (receptive language), the beta coefficient was −1.78 (95% CI −2.89 to −0.67, p = 0.002), and for boys, it was −5.68 (95% CI −9.25 to −2.11, p = 0.002). For girls (expressive language), the beta coefficient was −1.99 (95% CI −3.46 to −0.52, p = 0.008), and for boys, it was −2.79 (95% CI −4.79 to −0.78, p = 0.007). This suggests that anaemia in pregnancy had a greater negative effect on receptive and expressive language for boys than for girls, although it still had an effect in the girls. Maternal anaemia was also robustly associated with increased odds of cognitive delay (aOR1.66; 95% CI 1.04–2.66, p = 0.033), receptive language delay (aOR1.77; 95% CI 1.09–2.88, p = 0.020), and expressive language delay (aOR1.83; 95% CI 1.12–2.98, p = 0.015) overall. Maternal anaemia demonstrated a stronger interaction with these outcomes in boys as above.
On bivariate analysis, maternal psychosocial risk factors associated with poorer developmental outcomes comprised maternal antenatal depression or lifetime maternal exposure to IPV. In multivariable regression, antenatal depression was associated with poorer cognitive scores (β = −1.03; 95% CI −1.04 to −0.12, p = 0.027). For maternal IPV, though the model without interactions was not significant in the adjusted analysis, there was an association with higher odds of expressive language delay, which came out in the interaction model (main effect aOR 1.88; 95% CI 1.13–3.13, p = 0.015).
In bivariate analyses, maternal education and higher birth weight were protective against developmental delay in four domains, and preterm birth was associated with poorer outcomes.
On adjusted analyses, maternal education (aOR = 0.40; 95% CI 0.17–0.94, p = 0.035) was found to be protective against developmental delay in all four domains across the total cohort (S5 and S6 Tables)
This study provides directly measured developmental data from a low- and middle-income context in a well-characterised sample of mother–child dyads and emphasises the high prevalence of developmental delay. Our findings highlight the important protective effects of maternal education, birth weight, and socioeconomic status for developmental outcomes. Key risk and protective factors impacted developmental outcomes for boys and girls in different ways, and boys consistently performed worse than girls.
The substantial prevalence of developmental delay in around half the cohort reported in this study is higher than in previous reports. UNICEF’s caregiver-reported Early Childhood Development Index found that 37% of 3- and 4-year-olds in 35 LMICs do not attain basic cognitive and socio-emotional skills [41]. In this study, boys performed uniformly poorly compared to girls with lower raw scores across cognitive, language, and fine motor domains and correspondingly had increased risk of developmental delay in any one of these domains. This pattern of findings is consistent with other studies exploring developmental performance in very young children and has recently been described in a large multicountry study of older children across South East Asia [9].
In our cohort, key factors associated with positive development at 24 months included better socioeconomic status, higher birth weight, and higher maternal educational attainment. Maternal educational attainment was the strongest protective factor for reducing the odds of four-domain developmental delay across this group of vulnerable children. The importance of social determinants on child development is well described. In a recently published monograph reflecting on the Young Lives study spanning three different LMIC countries on three continents (Vietnam, Peru, and Ethiopia), the authors demonstrate that both early child economic well-being as well as caregiver (generally mothers’) education predicts both receptive vocabulary at age 5 years as well as reading comprehension at 15 years. We found this effect is identifiable at an assessment as early as 24 months of age, which is a critical reminder of the enduring impact of these contextual factors on children’s outcomes into young adulthood. The fact that this signal is present at so young an age (typically before most children attend preschool), though an association rather than causal finding, still speaks to the importance of relational factors and that boys may be particularly sensitive to this influence. Furthermore, this is supportive of the global focus on nurturing care to improve early child development and importance of supporting young women to attend and remain in school and the caregiver–child relationship in supporting developmental outcomes [42,43].
Factors that were particularly identified as associated with risk of poor developmental outcomes include those that represent physical health, and these were the factors for which child sex appeared to have had the greatest interaction with developmental outcomes in this cohort. The interaction of male sex with birth weight and developmental outcomes was not a linear one, with boys demonstrating increased vulnerability at lower birth weights but conversely at higher birth weights demonstrating better cognitive and language outcomes. The increased vulnerability of boys to the effects of maternal anaemia and low birth weight may represent a combination of the effects of different timing of brain network development by sex or sex-related adaptation pathways of motor and sensory systems to environmental exposures [8]. Maternal anaemia, one of the pregnancy-related risk factors for which there is robust evidence for long-term developmental impact, came out most strongly as interacting with boys in associating with poor developmental outcomes, though girls were still affected. Though significant only in the unadjusted analysis (and at trend level in the multivariable regression), we mention maternal HIV infection, as it is such an important public health issue. In this cohort, maternal HIV was associated with lower cognitive and language scores in the children. Child HIV infection is known to impact neurodevelopment [44]; however, literature on the impact of maternal HIV exposure without infection is emerging [45,46], and our data indicate that HIV-exposed uninfected children may be affected at 2 years, an area that needs ongoing investigation. We found limited associations with breastfeeding in this analysis; however, it may be due to the relatively low rates of breastfeeding across the cohort, with only 17% of children being exclusively breastfed for 6 months.
Antenatal depression was associated with poorer developmental outcomes across the cohort. The impact of postnatal depression on child growth and development is well established. There is increasing evidence for the effect of prenatal mental disorders including depression on child growth and development, and common mental health disorders, such as depression, appear to increasingly be recognised and highly prevalent in women living in high-risk environments [4,47,48]. Exposure to violence within the home has, likewise, been linked to increased developmental, psychological, and behavioural problems [49–51] as well as impaired child growth both in utero and in early life. Maternal lifetime exposure to IPV increased odds of expressive language delay. Maternal exposure to IPV, depression, or distress may disrupt a mother’s ability to provide care for her child [52], and early childhood adversity has been found to impact child development in other studies [11]. A range of underlying biological mechanisms may be relevant here, including the potential for the infants’ hypothalamic–pituitary axis to have been impacted postnatally through the mother–child interaction [53]. Further work is needed to better explore potential pathways of associations between maternal exposure to violence and child health and developmental outcomes.
Dropout rate remains an important limitation to consider in a study of this type. Of the total live births in the DCHS cohort, approximately 35.8% of children did not attend the 24-month developmental assessment visit (representing 26.8% of the children still in the cohort at 24 months). Although every effort was made to minimise study dropout, the inevitable loss of children from long assessments such as the developmental assessment visit tended to be clustered amongst the slightly younger mothers, who were more likely to be first-time parents and possibly less sensitised to the importance of developmental outcomes. Ex-premature children were also less likely to attend the BSID-III visit. We hypothesise that this may be due to mothers of children being born prematurely having a greater focus on the visits that related to physical health, given the vulnerability of ex-premature infants to intercurrent infections and other physical health problems. It is difficult to interpret the reason for the lower proportion of smoking mothers who chose not to attend the BSID-III visit, although the prevalence of active smoking was high in both groups. Despite these group differences, our remaining sample is large enough and adequately representative of the population in the region for us to be confident that our findings can be meaningfully interpreted. Additionally, bias was reduced by using a population-based cohort study design—choosing a community sample and enrolling consecutive pregnancies when eligible—and performing a complete case–based analysis approach. Finally, the BSID-III uses US population–normative means to create scaled scores, as currently there are no South African norms. Although the BSID-III has been validated for use in South Africa, this limitation means it may not be generalisable to sub-Saharan Africa and may underestimate developmental delay; however, the use of raw scores alongside in this study adds validity to our outcomes.
In conclusion, with the increasing global focus on early child development, population studies directly measuring developmental outcomes are needed to complement the global estimates that use poverty and stunting as proxy measures [41]. This will aid tracking progress towards the Sustainable Development Goals and enable appropriate early child development programmes that are being developed to appropriately target key factors that impact outcomes across the life span. Threats to development, which have been laid out in this manuscript, support the current framework for intervention that targets services for the components of nurturing care [54]. Given the high prevalence of developmental delay in this population, the risk and protective factors identified in this study provide valuable focus for intervention policy design and implementation in this critical area. Public health policy needs to work along the continuum of prepregnancy, pregnancy, and early childhood in the context of families, concentrating on health and nutrition, and encouraging safety, security, and early learning in the context of nurturing care.
|
10.1371/journal.ppat.1002012 | Metabolite Cross-Feeding Enhances Virulence in a Model Polymicrobial Infection | Microbes within polymicrobial infections often display synergistic interactions resulting in enhanced pathogenesis; however, the molecular mechanisms governing these interactions are not well understood. Development of model systems that allow detailed mechanistic studies of polymicrobial synergy is a critical step towards a comprehensive understanding of these infections in vivo. In this study, we used a model polymicrobial infection including the opportunistic pathogen Aggregatibacter actinomycetemcomitans and the commensal Streptococcus gordonii to examine the importance of metabolite cross-feeding for establishing co-culture infections. Our results reveal that co-culture with S. gordonii enhances the pathogenesis of A. actinomycetemcomitans in a murine abscess model of infection. Interestingly, the ability of A. actinomycetemcomitans to utilize L-lactate as an energy source is essential for these co-culture benefits. Surprisingly, inactivation of L-lactate catabolism had no impact on mono-culture growth in vitro and in vivo suggesting that A. actinomycetemcomitans L-lactate catabolism is only critical for establishing co-culture infections. These results demonstrate that metabolite cross-feeding is critical for A. actinomycetemcomitans to persist in a polymicrobial infection with S. gordonii supporting the idea that the metabolic properties of commensal bacteria alter the course of pathogenesis in polymicrobial communities.
| Many bacterial infections are not the result of colonization and persistence of a single pathogenic microbe in an infection site but instead the result of colonization by several. Although the importance of polymicrobial interactions and pathogenesis has been noted by many prominent microbiologists including Louis Pasteur, most studies of pathogenic microbes have focused on single organism infections. One of the primary reasons for this oversight is the lack of robust model systems for studying bacterial interactions in an infection site. Here, we use a model co-culture system composed of the opportunistic oral pathogen Aggregatibacter actinomycetemcomitans and the common oral commensal Streptococcus gordonii to assess the impact of polymicrobial growth on pathogenesis. We found that the abilities of A. actinomycetemcomitans to persist and cause disease are enhanced during co-culture with S. gordonii. Remarkably, this enhanced persistence requires A. actinomycetemcomitans catabolism of L-lactate, the primary metabolite produced by S. gordonii. These data demonstrate that during co-culture growth, S. gordonii provides a carbon source for A. actinomycetemcomitans that is necessary for establishing a robust polymicrobial infection. This study also demonstrates that virulence of an opportunistic pathogen is impacted by members of the commensal flora.
| The survival of pathogens in the human body has been rigorously studied for well over a century. The ability of bacteria to colonize, persist and thrive in vivo is due to an array of capabilities including the ability to attach to host tissues, produce extracellular virulence factors, and evade the immune system. Invading pathogens must also obtain carbon and energy from an infection site, and specific carbon sources are required for several pathogens to colonize and persist in the host [1]. Although mono-culture infections provide interesting insight into pathogenesis, many bacterial infections are not simply the result of colonization with a single species, but are instead a result of colonization with several [2], [3], [4], [5]. The mammalian oral cavity is an excellent environment to study polymicrobial interactions as it is persistently colonized with diverse commensal bacteria as well as opportunistic pathogens. Our lab has utilized a two-species model system composed of the opportunistic pathogen Aggregatibacter actinomycetemcomitans and the common commensal Streptococcus gordonii to provide mechanistic insight into how specific carbon sources impact disease pathogenesis in polymicrobial infections [6], [7].
A. actinomycetemcomitans is a Gram-negative facultative anaerobic bacterium that inhabits the human oral cavity and is a proposed causative agent of localized aggressive periodontitis [8]. A. actinomycetemcomitans is found between the gums and tooth surface in the subgingival crevice [9], [10], an area restricted for O2 depending on tissue depth [11] and irrigated by a serum exudate called gingival crevicular fluid (GCF). GCF not only contains serum proteins such as complement and immunoglobulin [12], but also glucose from 10 to 500 µM in healthy patients [13] and as high as 3 mM in patients with periodontal infections [14]. L-lactate is produced by host lactate dehydrogenase in GCF [15], [16] and resident oral streptococci. Together glucose and L-lactate represent two of the small number of carbon sources that A. actinomycetemcomitans is able to catabolize [17]. A. actinomycetemcomitans has been proposed to primarily inhabit the aerobic [9] “moderate” pockets (4 to 6 mm in depth) of the gingival crevice as opposed to deeper anaerobic subgingival pockets [18].
In addition to A. actinomycetemcomitans, the subgingival crevice is home to a diverse bacterial population, including numerous oral streptococci [19], that reside in surface-associated biofilm communities [20]. Oral streptococci, aside from Streptococcus mutans, are typically non-pathogenic and depending upon the human subject and method of sampling, comprise approximately 5% [21] to over 60% [22] of the recoverable oral flora. Through fermentation of carbohydrates to L-lactate and sometimes H2O2, acetate, and CO2, oral streptococci such as S. gordonii have been shown to influence the composition of oral biofilms [19], [20], [23], [24]. Additionally, S. gordonii-produced H2O2 influences interactions between A. actinomycetemcomitans and the host by inducing production of ApiA, a factor H binding protein that inhibits complement-mediated lysis [7], [25]. Thus, streptococcal metabolites are important cues that influence the growth and population dynamics of oral biofilms and how oral bacteria interact with the host.
A. actinomycetemcomitans preferentially catabolizes L-lactate over high energy carbon sources such as glucose and fructose in multiple strains, despite the fact that this bacterium grows more slowly with L-lactate [6]. Given this preference for a presumably inferior carbon source and the observation that A. actinomycetemcomitans resides in close association with oral streptococci [26], [27], we hypothesize an in vivo benefit exists for A. actinomycetemcomitans L-lactate preference. To test this hypothesis, we investigated the importance of A. actinomycetemcomitans L-lactate catabolism during mono-culture and co-culture with S. gordonii in vitro and in a murine abscess model of infection. Our results reveal that co-culture with S. gordonii enhances colonization and pathogenesis of A. actinomycetemcomitans, and the ability to utilize L-lactate as an energy source is essential for these co-culture benefits. Surprisingly, inactivation of L-lactate catabolism had no impact on mono-culture growth in vitro and in vivo suggesting that A. actinomycetemcomitans L-lactate catabolism is only critical for establishing co-culture infections. Taken together, these results provide compelling mechanistic evidence that the metabolic properties of human commensals such as S. gordonii can alter the course of pathogenesis in polymicrobial communities.
Within the gingival crevice, host-produced glucose and L-lactate are present [13], [14], [15], [16], [28] and likely serve as in vivo carbon sources for A. actinomycetemcomitans. However in contrast to glucose, L-lactate is also produced by the oral microbial flora, primarily oral streptococci [20]. Indeed, the ability of A. actinomycetemcomitans to catabolize streptococcal-produced L-lactate has been demonstrated previously [6], and it was proposed that A. actinomycetemcomitans consumes streptococcal-produced L-lactate during co-culture. To assess the importance of A. actinomycetemcomitans L-lactate catabolism in polymicrobial communities in vitro, we examined the metabolic profile during catabolism of L-lactate and glucose under aerobic and anaerobic conditions. Aerobically, A. actinomycetemcomitans primarily produced lactate and acetate from glucose (Fig. 1A) while acetate was the sole metabolite produced by L-lactate-grown bacteria (Fig. 1C). It was intriguing that lactate was produced, but not consumed, by A. actinomycetemcomitans during aerobic catabolism of glucose. We hypothesized that the lactate produced by A. actinomycetemcomitans was likely D-lactate, which is not catabolized by A. actinomycetemcomitans [29]. Using an enzymatic assay [30], we were able to verify that >99% of the lactate produced by A. actinomycetemcomitans was indeed D-lactate.
Anaerobically from glucose, A. actinomycetemcomitans primarily produced the mixed acid fermentation products formate and acetate along with lactate, succinate, and trace amounts of ethanol (Fig. 1B). Surprisingly, A. actinomycetemcomitans was unable to catabolize L-lactate anaerobically (Fig. 1C), even if the potential alternative electron acceptors nitrate or dimethyl sulfoxide were added, suggesting that L-lactate oxidation was O2 dependent. This is distinct from other oral bacteria including members of the genus Veillonella [24], [31], in which L-lactate is an important anaerobic carbon and energy source. If O2 respiration was indeed required for A. actinomycetemcomitans growth with L-lactate, we hypothesized that elimination of the terminal respiratory oxidase, which is required for aerobic respiration, would abolish L-lactate utilization by A. actinomycetemcomitans aerobically. To test this hypothesis, cydB, which encodes a component of the sole putative A. actinomycetemcomitans respiratory oxidase, was insertionally inactivated. The cydB mutant was unable to catabolize L-lactate aerobically supporting the hypothesis that L-lactate oxidation requires O2 respiration (Fig. 1C). Interestingly when grown with glucose aerobically, the cydB mutant doubled much slower (6.6 hr) than the wt (1.9 hr) and cell suspensions produced a metabolite profile that differed from the wt (Fig. 1A) indicating that while not required for aerobic growth on glucose, O2 respiration is the primary means by which glucose is catabolized by wt A. actinomycetemcomitans. As expected, the cydB mutant exhibited identical growth rates anaerobically on glucose (not shown) and produced similar metabolites as the wt (Fig. 1B). Collectively, these data indicate that O2 respiration is required for L-lactate oxidation in A. actinomycetemcomitans.
As the ultimate goal of this study is to assess the importance of A. actinomycetemcomitans L-lactate catabolism for establishing co-culture with oral streptococci, it was important to assess whether eliminating the ability of A. actinomycetemcomitans to utilize L-lactate affected growth with glucose. To examine this, we examined growth and metabolite production in an A. actinomycetemcomitans strain in which the catabolic L-lactate dehydrogenase LctD, which is present in all strains sequenced to date [32], [33], was insertionally inactivated [29]. LctD oxidizes L-lactate to pyruvate and is required for A. actinomycetemcomitans growth with L-lactate as the sole energy source [29]. As expected, the lctD mutant was unable to catabolize L-lactate aerobically or anaerobically (Fig. 1C); however, metabolite production from glucose was not affected (Fig. 1A&B) nor was the growth rate with glucose (not shown). These data indicate that L-lactate catabolism can be eliminated in A. actinomycetemcomitans without affecting growth and metabolite production with glucose.
Because A. actinomycetemcomitans preferentially catabolizes L-lactate in lieu of hexose sugars [6], we hypothesized that L-lactate cross-feeding was important for establishing co-culture with oral streptococci grown on glucose. To test this hypothesis, we examined growth of glucose-grown A. actinomycetemcomitans and S. gordonii during in vitro co-culture aerobically and anaerobically. Aerobically, wt A. actinomycetemcomitans co-culture cell numbers were similar to those observed in mono-culture while the A. actinomycetemcomitans lctD mutant exhibited an approximate 25-fold decrease in cell number during co-culture with S. gordonii (Fig. 2). Anaerobically, both wt A. actinomycetemcomitans and A. actinomycetemcomitans lctD- cell numbers diminished nearly 10-fold in co-culture compared to mono-culture (Fig. 2), likely due to the inability to catabolize S. gordonii-produced L-lactate.
Examination of aerobic metabolic end products of the A. actinomycetemcomitans lctD-/S. gordonii co-culture revealed high levels of lactate, reminiscent of S. gordonii mono-cultures, indicating that as expected, the A. actinomycetemcomitans lctD mutant is unable to catabolize L-lactate in co-culture (Fig. 3A). Additionally, metabolite concentrations in anaerobic co-cultures were similar to S. gordonii mono-culture (Fig. 3B). It should be noted that these metabolites were measured from growing cells, not cell suspensions as in Fig. 1. These data provide strong evidence that the inability to use L-lactate, even when glucose is present, significantly inhibits A. actinomycetemcomitans growth and survival in co-culture.
Interestingly, an approximate 7-fold increase in S. gordonii cell numbers were observed in the presence of A. actinomycetemcomitans aerobically, indicating that A. actinomycetemcomitans enhances S. gordonii proliferation under these co-culture conditions even when A. actinomycetemcomitans is unable to utilize L-lactate (Fig. S1 in Text S1). Importantly, the pH of the medium used in these experiments remained at neutrality; thus changes in cell numbers were not due to alterations in pH.
The observation that L-lactate catabolism is critical for A. actinomycetemcomitans to establish co-culture with S. gordonii in vitro provides new insight into this model polymicrobial community; however, whether the requirement for this catabolic pathway extended to in vivo co-culture was not known. To examine the role of A. actinomycetemcomitans L-lactate catabolism for in vivo growth in mono- and co-culture, we used a mouse thigh abscess model. This model has relevance as A. actinomycetemcomitans causes abscess infections outside of the oral cavity in close association with other bacteria [34] and has been used as a model system to examine pathogenesis of several oral bacteria [35], [36]. Using this model, bacterial survival and abscess formation was assessed for wt A. actinomycetemcomitans and A. actinomycetemcomitans lctD- during mono- and co-culture with S. gordonii (Fig. 4).
Unexpectedly, wt A. actinomycetemcomitans and the lctD mutant established similar infections in terms of cell number (Fig. 4A) and in abscess weight (Fig. 4B) indicating that host-derived L-lactate is not an important in vivo nutrient source during mono-culture infection. Interestingly, wt A. actinomycetemcomitans displayed a 10-fold increase in cell number when co-cultured with S. gordonii, while cell number of the lctD mutant declined >100-fold compared to the wild-type providing evidence that the ability to catabolize L-lactate is crucial for A. actinomycetemcomitans co-culture survival in vivo. These data also indicate that while not critical for mono-culture growth, L-lactate is an important energy source during co-culture infection. Unlike the in vitro experiments (Fig. S1 in Text S1), S. gordonii numbers were not statistically different in monoculture or in co-culture abscesses (2.7×107 and 1.3×107 CFU/ml respectively; p = 0.15 via Mann-Whitney test) indicating that S. gordonii does not receive a benefit, at least in regard to cell number, from co-culture with A. actinomycetemcomitans. As a control, in vivo growth of the A. actinomycetemcomitans apiA mutant, which is hypersusceptible to killing by innate immunity, was examined. As expected, the apiA mutant exhibited a >250-fold decrease in mono-culture in vivo survival, which was unchanged in the presence of S. gordonii (Fig. 4A).
Microbes within polymicrobial infections often display synergistic interactions that result in enhanced colonization and persistence in the infection site [5], [34], [36], [37], [38], [39], [40]. Such interactions have been particularly noted in oral polymicrobial infections, although the molecular processes controlling these synergistic interactions are not well defined. Detailed mechanistic studies of the interactions required for enhanced persistence in vivo is a critical step towards a more comprehensive understanding of natural polymicrobial infections. In this study, we used a model polymicrobial infection [6], [7] to determine the importance of metabolic cross-feeding for establishing co-culture infections. Cross-feeding in polymicrobial populations has been reported in numerous studies [24], [41], [42], but its importance for establishing co-culture infections has not been investigated in depth. The methodology used in this study began with detailed studies of the metabolic pathways required for growth with the in vivo carbon sources glucose and L-lactate, followed by examination of the importance of specific catabolic pathways for establishing co-culture infections.
It is relevant to discuss the rationale for two in vivo experimental parameters: using a ‘smooth’ strain of A. actinomycetemcomitans in lieu of a ‘rough’ strain; and using a murine abscess model in lieu of a rat periodontal infection model [43], [44]. A “smooth” strain of A. actinomycetemcomitans, which displays impaired surface attachment, was used in this study [45], [46]. As we were not investigating attachment or biofilm development, we opted to utilize a smooth strain that had undergone robust metabolic characterization, and feel this decision is justified as this bacterium clearly causes abscess infections in this model (Fig. 4). The murine abscess model was used for several reasons. First, in addition to periodontal infections, A. actinomycetemcomitans causes abscess infections outside of the oral cavity that resemble, from a gross morphological standpoint, the abscess model infection [34]; thus the abscess model has clinical relevance. Second, the abscess model avoids complications arising from the normal flora, which are not completely eradicated in the periodontal rat infection models, and whose presence would make interpretation of metabolic interactions extraordinarily complex. Third, the abscess model allows direct, controlled inoculation with a finite number of cells that can be quantified throughout the infection by assessing colony forming units after removal of the entire abscess [37], [47], [48]. Finally, although the abscess model has primarily been used to study anaerobic pathogens [35], [36], it is also relevant for studying aerobic pathogens, demonstrated by the large abscesses [48] formed by the strict aerobe Acinetobacter baumanii [17], [49]. The presence of aerobic microenvironments in the abscess is also supported by our observations that the S. gordonii spxB mutant is significantly impaired for abscess formation (Fig. S2 in Text S1). The spxB gene encodes pyruvate oxidase which utilizes O2 for biosynthesis of the virulence factor H2O2 [50]; thus its importance is limited to aerobic infections.
The observation that A. actinomycetemcomitans requires O2 to catabolize L-lactate was surprising, as many oral bacteria grow on L-lactate anaerobically [24], [31]. These results also solve an apparent contradiction in the literature. It was reported by multiple sources [17], [51] that A. actinomycetemcomitans does not catabolize L-lactate, yet we recently provided evidence that several strains of A. actinomycetemcomitans grow aerobically with L-lactate as the sole energy source [6], [29]. Interrogation of the previous growth environments revealed that A. actinomycetemcomitans was grown under very low or O2 free conditions; thus it is not surprising that significant growth was not observed in these studies. The O2 dependency of L-lactate oxidation also highlights another facet of our in vivo data. In the murine abscess model, the A. actinomycetemcomitans wt and lctD mutant grew equally well in mono-culture (Fig 4). However, in co-culture only the survival of the lctD mutant was impaired. This result is reminiscent of our in vitro data (Fig. 2) suggesting that O2 dependent metabolism occurs in our model polymicrobial infection.
The observation that the terminal oxidase CydB is required for aerobic growth with L-lactate allows development of a new model for L-lactate consumption in A. actinomycetemcomitans (Fig. 5). Since L-lactate dehydrogenase (LctD) is necessary for lactate oxidation and does not use NAD+ as an electron acceptor [29], anaerobic fermentation pathways that regenerate NAD+ cannot act as electron acceptors for L-lactate oxidation. The model predicts that A. actinomycetemcomitans instead donates electrons directly to the quinone pool which in turn is re-oxidized by CydAB [52]. It should be noted that this does not rule out an unknown electron carrier between LctD and the membrane associated quinone.
The most exciting observation from these studies is that L-lactate catabolism is likely an important factor for A. actinomycetemcomitans to establish a polymicrobial, but not mono-culture, infection in a murine abscess model (Fig. 4). These data indicate that host-produced L-lactate is not a vital energy source for A. actinomycetemcomitans in mono-culture abscesses, but when S. gordonii is present, L-lactate catabolism becomes critical. We speculate that in the absence of S. gordonii, carbohydrates such as glucose are present in the infection site for A. actinomycetemcomitans growth. When S. gordonii is introduced, competition for these carbohydrates increases, and A. actinomycetemcomitans is likely at a disadvantage due to its relatively slow growth and catabolic rates compared to S. gordonii [6]. Thus, the ability to preferentially utilize L-lactate, the primary metabolite produced by S. gordonii, allows A. actinomycetemcomitans to avoid competition with S. gordonii for carbohydrates and consequently enhance its survival in the abscess. This model (Fig. 6) suggests that the importance of individual carbon catabolic pathways is dependent on the context of the infection, specifically if oral streptococci are present.
Our work demonstrates that metabolic pathways required for A. actinomycetemcomitans proliferation during mono-culture infection are distinct from those required for co-culture infection with a common commensal. This study provides strong evidence that simply because elimination of a catabolic pathway does not elicit a virulence defect in mono-species infection does not preclude it from being important in polymicrobial infections. Since metabolic interactions can potentially occur in virtually any polymicrobial infection, our results suggest that in some cases, the ability to cause infection will be as dependent on metabolic interactions as it is on known immune defense mechanisms and classical virulence factors. Our observations also have therapeutic implications, as development of small molecule inhibitors of metabolic pathways, particularly pathways restricted to prokaryotic pathogens, have promise as new therapeutic targets. Based on this study, efforts to develop such therapeutics will require a detailed understanding of how polymicrobial cross-feeding affects colonization and persistence in an infection site.
This study was carried out in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health. The protocol was approved by the Institutional Animal Care and Use Committee of Texas Tech University Health Sciences Center (Protocol Number: 09039).
A. actinomycetemcomitans strains VT1169 [53], Streptococcus gordonii strain Challis DL1.1 (ATCC 49818), S. gordonii spxB- [50], Escherichia coli DH5α-λpir, and E. coli SM10-λpir were used in this study. A. actinomycetemcomitans and S. gordonii were routinely cultured using Tryptic Soy Broth + 0.5% Yeast Extract (TSBYE). For resting cell suspension A. actinomycetemcomitans metabolite analysis, a Chemically Defined Medium (CDM) [6] lacking nucleotides, amino acids, pimelate and thioctic acid (to eliminate further cell growth) containing either 20 mM glucose or 40 mM L-lactate was used. For co-culture experiments, complete CDM with 3 mM glucose was used. Aerobic culture conditions were 37°C in a 5% CO2 atmosphere shaking at 165 RPM, and anaerobic culture conditions were static growth at 37°C in an anaerobic chamber (Coy, USA) with a 5% H2, 10% CO2 and 85% N2 atmosphere. E. coli strains were grown on Luria-Bertani (LB) medium at 37°C. Where applicable, antibiotics were used at the following concentrations: chloramphenicol, 2 µg/ml for A. actinomycetemcomitans and 20 µg/ml for E. coli; spectinomycin, 50 µg/ml for selection and 10 µg/ml for maintenance for A. actinomycetemcomitans and E. coli and 100 µg/ml for selection and maintenance for S. gordonii spxB-; kanamycin, 40 µg/ml for selection and 10 µg/ml for maintenance; naladixic acid, 25 µg/ml; streptomycin, 50 µg/ml for selection and 20 µg/ml for maintenance. For quantifying CFU/ml in co-culture assays, vancomycin (5 µg/ml) was added to agar plates to enumerate A. actinomycetemcomitans and streptomycin (100 µg/ml) was added to agar plates to enumerate S. gordonii.
DNA and plasmid isolations were performed using standard methods [54]. Restriction endonucleases and DNA modification enzymes were purchased from New England Biolabs. Chromosomal DNA from A. actinomycetemcomitans was isolated using DNeasy tissue kits (Qiagen), and plasmid isolations were performed using QIAprep spin miniprep kits (Qiagen). DNA fragments were purified using QIAquick mini-elute PCR purification kits (Qiagen), and PCR was performed using the Expand Long Template PCR system (Roche). DNA sequencing was performed by automated sequencing technology using the University of Texas Institute for Cell and Molecular Biology sequencing core facility.
Allelic replacement of apiA (AA2485) was carried out by double homologous recombination. For construction of the knockout construct, 856 bp and 842 bp DNA fragments flanking apiA were amplified and combined with the aphA gene (encoding kanamycin resistance) from pBBR1-MCS2 [55] by overlap extension PCR [56]. The construct was prepared so that aphA was positioned between the upstream and downstream regions. Primers used were: Kan-5′ (ATGTCAGCTACTGGGCTATCTG) and Kan-3′ (ATTTCGAACCCCAGAGTCCCGC) for the 1074 bp aphA-containing fragment; ApiA-UF (CCGATAACAGTAAGATCTTCTAC) and ApiA-UR (CAGATAGCCCAGTAGCTGACATCCTTTTCGGCTTGAATTTATACC) for the upstream apiA fragment; and ApiA-DF (GCGGGACTCTGGGGTTCGAAATGCGGTCAGAATTTTAGGTGTTTT) and ApiA-DR (CGAAACCAACGAACTCTTTATTC) for the downstream apiA fragment. Underlined sequences indicate overlapping DNA sequences between the apiA fragments and aphA. The overlap extension product was TA-cloned into the pGEM-T Easy vector (Promega, USA) and excised by EcoRI digest. The EcoRI fragment containing the overlap extension product was ligated into the unique EcoRI site within the λpir-dependent suicide vector pVT1461 [57]. The cloned construct, pVT1461-apiA-KO, was first transformed into E. coli DH5α-λpir then into E. coli SM10-λpir for conjugation into A. actinomycetemcomitans. Conjugation was performed as described [53] and potential mutants were plated onto TSBYE agar plates containing kanamycin to select for recombinant A. actinomycetemcomitans and nalidixic acid to kill the E. coli donors. Kanamycin resistant, spectinomycin sensitive double recombinants were selected and verified by PCR. Enhanced susceptibility of the apiA mutant to serum was verified as described previously [7].
Insertional mutagenesis of the cydB gene was performed by single homologous recombination using a 543 bp internal piece of the cydB (AA2840) gene amplified using the primers cydB-KO5′ (GAAGATCTTTATGATTAATACTATCGCGCCG) and cydB-KO3′ (GAAGATCTCAAAACCATCTTTGAAAGATAACCA). Underlined sequences represent BglII restriction sites. The internal cydB fragment was digested with BglII and ligated into the A. actinomycetemcomitans suicide vector pMRKO-1 (see below) to generate pMRKO-cydB. pMRKO-cydB was transformed into E. coli SM10-λpir and conjugated into A. actinomycetemcomitans. A. actinomycetemcomitans recombinants were grown anaerobically on TSBYE agar containing spectinomycin and naladixic acid. Colonies were subcultured anaerobically on liquid medium at the same antibiotic concentrations and insertion into cydB was verified by PCR.
The spectinomycin resistance gene from pDMG4 [58] was amplified by PCR using the primers: 5′Spec-cass-NotI (ATAAGAATGCGGCCGCCGATTTTCGTTCGTGAATACATG) and 3′ Spec-cass-EcoRI (CGGAATTCCATATGCAAGGGTTTATTGTTT), digested with NotI-EcoRI and ligated into NotI-EcoRI digested pmCherry (Clontech) underlined sequences indicate NotI and EcoRI restriction sites. The 3105 bp region containing the pUC origin of replication, plac:mCherry and the spectinomycin resistance gene were PCR amplified using the primers: 5′pMcher-trunc (GAAGATCTGACCAAGTTTACTCATATATACT) and 3′ Spec-cass-EcoRI (CGGAATTCCATATGCAAGGGTTTATTGTTT). Underlined sequences indicate BglII and EcoRI restriction sites. This fragment was digested with BglII and EcoRI and ligated into the 2780 bp fragment from BglII-EcoRI digested pVT1461. The resulting plasmid (pMRKO-1, submitted to Genbank) is a suicide vector for A. actinomycetemcomitans and contains oriT, mob, and tra genes from pVT1461 along with the pUC origin of replication, mCherry expressed from plac, and a spectinomycin resistance cassette.
A. actinomycetemcomitans was grown in CDM overnight either aerobically or anaerobically in the presence of 20 mM glucose or 40 mM L-lactate. Bacteria were then subcultured in 30 ml of medium and exponential phase cells (OD600 = 0.4) were collected by centrifugation (5,000 x g for 15 min) at 25°C. Cell pellets were resuspended in an equal volume of CDM lacking nucleotides, amino acids and any carbon source. Cells were incubated at 37°C aerobically or anaerobically depending on the test conditions for 1 hr. Cells were collected again by centrifugation as described above and resuspended to an OD600 of 2 in 3 ml of CDM without nucleotides, amino acids, pimelate and thioctic acid containing either 20 mM glucose or 40 mM lactate. Cells were incubated for 4 h at 37°C either aerobically or anaerobically. After incubation samples were stored at −20°C for HPLC analysis.
D-lactate assays were performed as described [30] with modifications. Glycylglycine buffer was replaced with an equal concentration of Bicine (Fisher, USA) buffer and enzymatic assays were monitored by spectrophotometry at 340 nm for 4 hours.
A. actinomycetemcomitans and S. gordonii were grown overnight in CDM containing 3 mM glucose. 3 mM glucose was used to ensure that the medium was limited for catabolizable carbon. Cells were diluted 1∶50 in the same medium and allowed to grow to exponential phase (OD600 of 0.2). Cells were then diluted 1∶100 (2×106 S. gordonii/ml and 1×107 A. actinomycetemcomitans/ml) as mono-cultures or co-cultures in 3 ml CDM containing 3 mM glucose. Cultures were allowed to grow for 10 h aerobically or 12 h anaerobically, after which cells were serially diluted, plated on either TSBYE agar + vancomycin for A. actinomycetemcomitans enumeration or TSBYE agar + streptomycin for S. gordonii enumeration. Colonies were counted after incubation at 37°C for 48 h. An aliquot of the culture was also stored at −20°C for HPLC metabolite analysis.
Metabolite levels were quantified using a Varian HPLC with a Varian Metacarb 87H 300×6.5 mm column at 35°C. Samples were eluted using isocratic conditions with 0.025 N H2SO4 elution buffer and a flow rate of 0.5 ml/minute. A Varian refractive index (RI) detector at 35°C was used for metabolite enumeration by comparison with acetate, ethanol, formate, glucose, L-lactate, D-lactate, pyruvate and succinate standards.
Murine abscesses were generated essentially as described previously [37]. Briefly, 6–8 week-old, female, Swiss Webster mice were anesthetized with an intraperitoneal injection of Nembutal (50 mg/kg). The hair on the left inner thigh of each mouse was shaved, and the skin was disinfected with 70% alcohol. Mice were injected subcutaneously in the inner thigh with 107 CFU A. actinomycetemcomitans, S. gordonii or both. At 6 days post- infection, mice were euthanized and intact abscesses were harvested, weighed and placed into 2 ml of sterile PBS (or water for pH measurements). Tissues were homogenized, serially diluted and plated on Brain Heart Infusion (BHI) agar + 20 µg/ml Na2CO3 + vancomycin for A. actinomycetemcomitans enumeration or BHI agar + 20 µg/ml Na2CO3 + streptomycin for S. gordonii enumeration, to determine bacterial CFU/abscess. Experimental protocols involving mice were examined and approved by the Texas Tech University HSC Institutional Animal Care and Use Committee.
|
10.1371/journal.ppat.1005853 | Overexpression of Differentially Expressed Genes Identified in Non-pathogenic and Pathogenic Entamoeba histolytica Clones Allow Identification of New Pathogenicity Factors Involved in Amoebic Liver Abscess Formation | We here compared pathogenic (p) and non-pathogenic (np) isolates of Entamoeba histolytica to identify molecules involved in the ability of this parasite to induce amoebic liver abscess (ALA)-like lesions in two rodent models for the disease. We performed a comprehensive analysis of 12 clones (A1–A12) derived from a non-pathogenic isolate HM-1:IMSS-A and 12 clones (B1–B12) derived from a pathogenic isolate HM-1:IMSS-B. “Non-pathogenicity” included the induction of small and quickly resolved lesions while “pathogenicity” comprised larger abscess development that overstayed day 7 post infection. All A-clones were designated as non-pathogenic, whereas 4 out of 12 B-clones lost their ability to induce ALAs in gerbils. No correlation between ALA formation and cysteine peptidase (CP) activity, haemolytic activity, erythrophagocytosis, motility or cytopathic activity was found. To identify the molecular framework underlying different pathogenic phenotypes, three clones were selected for in-depth transcriptome analyses. Comparison of a non-pathogenic clone A1np with pathogenic clone B2p revealed 76 differentially expressed genes, whereas comparison of a non-pathogenic clone B8np with B2p revealed only 19 differentially expressed genes. Only six genes were found to be similarly regulated in the two non-pathogenic clones A1np and B8np in comparison with the pathogenic clone B2p. Based on these analyses, we chose 20 candidate genes and evaluated their roles in ALA formation using the respective gene-overexpressing transfectants. We conclude that different mechanisms lead to loss of pathogenicity. In total, we identified eight proteins, comprising a metallopeptidase, C2 domain proteins, alcohol dehydrogenases and hypothetical proteins, that affect the pathogenicity of E. histolytica.
| The pathogen Entamoeba histolytica can live asymptomatically in the human gut, or it can disrupt the intestinal barrier and induce life-threatening abscesses in different organs, most often in the liver. The molecular framework that enables this invasive, highly pathogenic phenotype is still not well understood. In order to identify factors that are positively or negatively correlated for invasion and destruction of the liver, we used a unique tool, E. histolytica clones that differ dramatically in their pathogenicity, while sharing almost identical genetic background. Based on comprehensive transcriptome studies of these clones, we identified a set of candidate genes that are potentially involved in pathogenicity. Using ectopic overexpression of the most promising candidates, either in pathogenic or in non-pathogenic Entamoeba clones, we identified genes where high expression reduced pathogenicity and only one gene that increased pathogenicity to a certain extend.
Taken together, the current study identifies novel pathogenicity factors of E. histolytica and highlights the observation that various different genes contribute to pathogenicity.
| The protozoan parasite Entamoeba histolytica is responsible for approximately 50 million cases of invasive amoebiasis per year, resulting in an annual death toll of 40,000–100,000 [1]. The parasite life cycle is relatively simple, comprising infectious cysts that can survive outside the host and vegetative trophozoites that proliferate in the human gut. After infection, E. histolytica trophozoites can asymptomatically persist for months or years in its human host [2].
Under as yet unknown circumstances, E. histolytica escapes from the gut lumen, either by penetrating the intestinal mucosa and inducing colitis, or by disseminating to other organs, most commonly the liver, where it induces abscess formation. The factors that determine the clinical outcomes of E. histolytica infections are not well understood. Possible factors comprise genetic make-up of the parasite and/or host, the immune response mounted by the host, concomitant infections and host diet. Identification of E. histolytica pathogenicity factors is a major topic in the field. Recently, research dealing with E. histolytica pathogenicity factors has mainly focused on a triad of protein families, namely, galactose/N-acetyl d-galactosamine–inhibitable Gal/GalNAc-lectins, cysteine peptidases (CPs) and amoebapores. Results obtained using transgenic amoebae supported the hypothesis that these molecules are involved in amoebic liver abscess (ALA) formation [3–6]. Nevertheless, homologues of the majority of these potential pathogenicity factors are also present in the non-pathogenic sister species Entamoeba dispar, a commensal protozoan that is genetically closely related to E. histolytica. Therefore, it remains to be shown whether one of these factors or their combination is responsible for amoeba pathogenicity or whether additional factors are involved. Thus, the mechanisms and processes enabling E. histolytica to penetrate host tissues and induce colitis and/or liver abscesses are still not understood. One straight-forward approach of identifying pathogenicity factors is a direct comparison of pathogenic and non-pathogenic E. histolytica isolates that has been performed using comparative microarray and proteome approaches [7–10]. Unfortunately, these studies used two isolates with completely different genetic backgrounds (pathogenic isolate HM-1:IMSS and non-pathogenic isolate Rahman). This rendered the straight-forward identification of pathogenicity factors almost impossible. In addition, an in-depth phenotypical characterisation of the Rahman isolate revealed a number of genomic defects that presumably interfere with its virulence capacity [10].
Recently, we identified two cell lines that were both derived from the clinical E. histolytica isolate HM-1:IMSS but which significantly differ in their pathogenicity. Whereas cell line HM-1:IMSS-A completely lost its ability to induce ALAs in gerbils (Meriones unguiculatus) and mice (Mus musculus), cell line HM-1:IMSS-B is highly aggressive and induces large ALAs. Comparative transcriptomic and proteomic studies of these cell lines have already been performed [11]. The studies revealed 31 differentially (≥3-fold) expressed genes [12] and 31 proteins with differential abundances in HM-1:IMSS-A and HM-1:IMSS-B [11]. However, an overlap of only two molecules was found between the proteomic and transcriptomic approaches.
Until now, neither pathogenic nor non-pathogenic cell lines have been cloned, and therefore it is possible that they consist of a mixture of cells. Thus, in the present study, both cell lines were cloned to obtain homogenous cell populations, allowing analyses of the pathogenicity factors in a straight-forward fashion. Twelve clones derived from the non-pathogenic cell line HM-1:IMSS-A (A1–A12) and 12 clones derived from the pathogenic cell line HM-1:IMSS-B (B1–B12) were generated. All 24 clones were analysed for their ability to induce ALAs in gerbils and for their specific CP activity. One non-pathogenic A-clone (A1np), one pathogenic B-clone (B2p) and one non-pathogenic B-clone (B8np) were analysed in more detail, including their time course of ALA formation, growth, size, motility, haemolytic activity, erythrophagocytosis, cytopathic activity and transcriptome profiles. Furthermore, transfectants overexpressing genes that were identified as differentially expressed between the pathogenic and non-pathogenic clones were tested for their involvement in ALA formation.
To analyse whether the E. histolytica cell lines consisted of a mixture of different cell types with different pathogenic phenotypes, the cell lines were cloned by limited dilution method. This resulted in 12 clones derived from cell line HM-1:IMSS-A and 12 clones derived from cell line HM-1:IMSS-B. The ALAs-generating ability of the different clones was analysed using the gerbil model. The animals were sacrificed 7 days post infection, and ALA sizes were determined. The results clearly indicated that the HM-1:IMSS-A cell line consists of a homogenous cell population. Except a few cases of small ALA formation, the majority of animals infected with different A-clones showed no ALA formation (Fig 1). The results were more divergent for clones derived from the HM-1:IMSS-B cell line. Eight clones, B2–B7, B9 and B10, showed a pathogenic phenotype comparable with the original cell line HM-1:IMSS-B. However, although clones B1, B8, B11 and B12 were derived from the pathogenic cell line, their ability to induce abscess formation was significantly reduced. This was especially evident for clone B8 that did not induce any abscess formation 7 days post infection (Fig 1).
The non-pathogenic clones A1np and B8np and pathogenic clone B2p have been continuously cultivated for more than 5 years, without any change of the respective phenotypes. The pathogenic phenotype of clone B2p remained especially stable over the years without the need for animal passaging.
To ensure that the observed phenotypes were indeed stable and uniform, the non-pathogenic clone B8np and the highly pathogenic clone B2p were sub-cloned. All sub-clones showed the same phenotype as the respective mother clone. All five sub-clones derived from the non-pathogenic clone B8np were unable to induce ALAs, whereas all five sub-clones of the pathogenic clone B2p produced large abscesses (Fig 2).
Recently, it was shown that CP activity of the pathogenic cell line HM-1:IMSS-B is approximately ten times greater (110 ± 25 mU/mg) than in the non-pathogenic cell line HM-1:IMSS-A (15 ± 5 mU/mg) [11]. A similar difference has been measured for the non-pathogenic clone A1np (15 ± 10 mU/mg) and the pathogenic clone B2p (123 ± 60 mU/mg) [3]. To investigate if the observed correlation between CP activity and pathogenicity is generally valid, the activities of all A- and B-clones were determined and correlated with ALA formation (Fig 3A). In general, the clones derived from HM-1:IMSS-B had a significantly higher CP activity (85 ± 50 mU/mg) than clones derived from HM-1:IMSS-A (18 ± 11; p < 0.0001). However, a direct correlation between the CP activity and ALA formation was not observed. This was obvious especially for clones B1 and B12. Although these clones have a high CP activity (198 ± 52 mU/mg and 139 ± 89 mU/mg, respectively), they only induce small ALAs (Figs 2 and 3A). CPs EhCP-A1, EhCP-A2, EhCP-A4, EhCP-A5 and EhCP-A7 can be visualised by substrate gel electrophoresis [3]. Here, the results from substrate gel experiments clearly indicated that different CP activities of the various A- and B-clones are not linked to the expression of a single peptidase (Fig 3B).
To identify the underlying mechanisms of different virulence phenotypes, three clones were selected for in-depth analyses. These were as follows: a non-pathogenic clone derived from HM-1:IMSS-A (clone A1np), pathogenic clone B2p that induced the largest ALAs among the B-clones, and clone B8np that completely lost its ability to produce ALAs. Both B-clones have been derived from HM-1:IMSS-B.
Magnetic resonance imaging (MRI) was employed to follow post-infection abscess formation over time in more detail. Infection time course was analysed over 10 days in gerbil and mouse ALA models using the three clones, A1np, B2p and B8np (Fig 4A and 4B). Our findings clearly confirmed the results of animal experiments described above (Fig 1). On day 7 post infection, no or very small ALAs were detected in both animal models with the two non-pathogenic clones A1np and B8np, whereas unequivocal abscess formation was observed with clone B2p (Fig 4A and 4B). Nevertheless, it became apparent that clone A1np was also able to induce abscess formation initially, as lesions were detected on day 3 post infection. However, these ALAs were smaller compared with ALAs seen during clone B2p infection and were more rapidly resolved. In contrast to clone A1np, the pathogenicity of clone B8np was almost completely abolished. No ALAs were detected in gerbils infected with this clone, while, in mice, ALAs on day 3 post infection were significantly smaller compared with ALAs induced by clone B2p (Fig 4A and 4B).
Clones A1np, B2p and B8np were then phenotypically characterised. This included determination of size, growth rate, haemolytic activity, erythrophagocytosis and cytopathic activity.
Microscopic analyses indicated that trophozoite sizes of clone B2p (818 ± 235 μm2) and clone B8np (782 ± 193 μm2) differed significantly and were larger in comparison with cells from clone A1np (591 ± 154 μM2) (Table 1).
Clone A1np grew significantly slower in comparison with clone B2p and clone B8np. The doubling time of clone A1np was approximately 12 ± 2.7 h, whereas it was approximately 8 ± 1.6 h for clone B2p and 9.5 ± 2.2 h for clone B8np (Table 1).
With an accumulated distance of 228 (±156) μm/10 min the amoebae of clone A1np move significantly slower in comparison to amoebae of clone B2p and clone B8np (376 ±172 μm/10 min (p < 0.0001) and 453 ±184 μm/10 min (p < 0.0001), respectively). While clone B8np moved significantly faster than B2p (p < 0.0013) (Table 1).
Clone A1np and clone B2p were able to lyse erythrocytes, but no haemolytic activity was detected for clone B8np. In addition, no correlation of haemolytic activity with pathogenicity was observed, since the non-pathogenic clone A1np had the highest activity (Table 1). By contrast, clone B8np displayed the highest erythrophagocytosis rate, followed by clone A1np and clone B2p.
Cytopathic activity (percentage of monolayer disruption) was highest for clone A1np followed by clone B2p. Clone B8np was unable to disrupt a cell monolayer (Table 1).
RNAseq experiments were performed to identify differences in gene expression profiles of clones A1np, B2p and B8np. Comparison of the non-pathogenic clone A1np and pathogenic clone B2p revealed 76 differentially expressed genes (threshold ≥ 3-fold, p-value adjusted (padj) < 0.05). Some genes (46) were expressed more highly in clone A1np and some (30) in clone B2p (Tables 2 and 3, S2 Table). From the 46 genes with higher expression levels in clone A1np 10 code for surface proteins (EHI_015290, EHI_082070, EHI_118130, EHI_169280, EHI_074080, EHI_075660 EHI_164900, EHI_039020, EHI_006170, EHI_086540) [13]. Amongst them are 2 members of the C2 domain protein family and 2 members of the Rab family. Since the analysis of the surface proteome referred to was performed with trophozoites of cell line A, it was not surprising that genes with higher expression levels in clone B2p could not be identified as surface associated [13].
The greatest differences, with expression fold changes ~20–200, concerned genes encoding three C2 domain proteins, three Rab family GTPases, cell surface protease gp63 and one hypothetical protein (EHI_074080), whose expression was higher in clone A1np; and five genes encoding hypothetical proteins EHI_127670, EHI_144490, EHI_169670, EHI_050490 and EHI_062080, with higher expression in clone B2p. The majority of the identified genes showed 3-4-fold differential expression (Tables 2 and 3). The detected differential expression was verified by quantitative real-time PCR (qPCR) for 26 genes. Results of next generation sequencing were confirmed for all the analysed genes, with the exception of EHI_056490 (threshold ≥ 2 fold, Table 4).
Most of the identified genes encode proteins with unknown function. Some of these hypothetical proteins contain functional domains, e.g., C2 domains, phosphatase domains, tyrosine kinase domains, RecF/RecN/SMC domains and lecithin:cholesterol acyltransferase domains. Proteins with known functions up-regulated in clone A1np, when compared with clone B2p, mainly included Rab family proteins, peptidases and heat shock proteins. Putative function could be assigned to only 6/30 genes identified as more highly expressed in clone B2p in comparison with clone A1np (e.g., tyrosine kinase, methionine gamma-lyase, thioredoxin) (Tables 2 and 3).
Comparisons of the two B-clones, B2p and B8np, revealed only 19 differentially expressed genes. Twelve genes were expressed at higher levels in clone B8np in comparison with clone B2p, and seven genes were expressed more highly in B2p in comparison with B8np (Tables 5 and 6, S3 Table). The corresponding proteins assigned to EHI_039020 and EHI_088020 were found to be part of the surface proteome of E. histolytica [13]. Fold change ≥ 10 was detected for only three genes. These genes encoded two hypothetical proteins and a leucine-rich repeat-containing protein. All these genes were more highly expressed in B2p in comparison with B8np. As observed in the A1np/B2p comparison, the majority of the identified genes showed 3-4-fold differential expression (Tables 5 and 6). The majority (11/19) of the genes encoded hypothetical proteins. The remaining genes were annotated as galactose-inhibitable lectin 35 kDa subunit, phosphoserine aminotransferase, actobindin, alcohol dehydrogenase, AIG1 family protein and methionine gamma-lyase. However, only galactose-inhibitable lectin 35 kDa subunit and methionine gamma-lyase have been biochemically characterised in E. histolytica [14–17].
Interestingly, only six genes in the two non-pathogenic clones A1np and B8np were similarly regulated vs. pathogenic clone B2p. Genes EHI_026360, EHI_039020 and EHI_056490 were up-regulated, and genes EHI_127670, EHI_144490 and EHI_144610 were down-regulated, in clones A1np and B8np in relation to clone B2p. This suggests different mechanisms accounting for the inability of clones A1np and B8np to induce ALAs.
In total 89 genes were found to be differentially expressed between clone A1np and clone B2p and/or between clone B8np and clone B2p. In a previous study, the transcriptomes of the non-clonal cell lines A and B were compared using a microarray approach [12]. Here, in total 31 genes were differentially expressed (threshold ≥3-fold). Of the 12 genes with higher expression levels in cell line B in comparison to cell line A, 7 genes had also higher expression levels in clone B2p in comparison to clone A1np. Out of the 19 genes with higher expression levels in cell line A in comparison to cell line B, 11 genes showed also higher expression in clone A1np in comparison to clone B2p (S4 Table). Only 8 of the identified 89 genes encoded for proteins containing a signal peptide and 15 genes encoded for proteins containing between 1–7 transmembrane domains (S4 Table). In a previous study in which the surface proteome of cell line A was analysed, 693 putative surface-associated proteins were identified [13]. Out of them 11 showed differential expression between the different clones (S4 Table).
From the 89 identified genes, 32 encode hypothetical proteins, where no homology to other proteins or protein domains could be identified.
Furthermore, 3 genes encode for proteins of the C2 superfamily, 4 genes encode for members of the small GTPase superfamily, 10 genes encode for heat shock proteins, 6 genes encode for AIG1 family proteins, 3 genes encode for kinases 2 genes encode for cysteine synthases and 2 genes encode for proteases. Additional 18 genes encode for proteins with other known functions (S4 Table).
To investigate whether the differentially expressed genes play a role in ALA formation, their respective overexpressing transfectants were generated. For genes that were more highly expressed in clone A1np in comparison with clone B2p, 13 overexpressing transfectants of clone B2p were generated. These included eight genes that displayed the highest differential expression (>20-fold) and three genes that were also up-regulated in clone B8np (EHI_026360, EHI_056490, EHI_039020) (Tables 2 and 5). We were unable to generate transfectants overexpressing genes EHI_169280 (ehrab7e), EHI_074080, EHI_187090 (ehrab7g) and EHI_075660 (ehcaax). For all other genes, a relative 2.4–235-fold overexpression was obtained (S5 Table). The pathogenic phenotype of the majority of B2p transfectants overexpressing genes that were more highly expressed in the non-pathogenic clone A1np vs. pathogenic clone B2p was unaffected. These also included transfectants that overexpressed three genes regulated in the same manner in clones A1np and B8np. All these B2p transfectants induced ALA formation in mice. Four genes were identified whose overexpression had a dramatic impact on ALA formation. When EHI_015290 (ehc2-3), EHI_059860 (ehc2-5), EHI_042870 (ehmp8-2) or EHI_075690 were overexpressed in clone B2p, these clones lost their pathogenic phenotype and produced significantly smaller ALAs than the respective controls (Fig 5A).
B2p transfectants were generated for 9/12 genes that were expressed at higher levels in clone B8np in comparison with clone B2p. They showed 4–300-fold increased expression in comparison with the control (S5 Table). Overexpression of genes EHI_058920, EHI_088020 and EHI_160670 significantly reduced pathogenicity of clone B2p (Fig 5A). In silico analyses indicated that nucleotide sequences of EHI_088020 and EHI_160670 were identical and that the genes encode an alcohol dehydrogenase. However, the first 480 nucleotides of the EHI_088020 coding region were missing from the EHI_160670 sequence. This may be because the E. histolytica genome is not yet fully annotated (AmoebaDB, http://amoebadb.org/amoeba/). Regardless, increased expression of the full-length or truncated gene impacted abscess formation (Fig 5A).
For genes expressed more highly in clone B2p in comparison with clone A1np, five gene-overexpressing clone A1np transfectants were generated, including three genes with an initially detected >90-fold differential expression. Overexpression of EHI_169670 was unsuccessful; however, for all other genes a relative 4–490-fold expression was obtained (S6 Table). Overexpression of these genes did not significantly affect ALA formation. However, strikingly, 4/9 mice infected with EHI_127670-overexpressing A1np transfectant produced large ALAs (Fig 5B).
Seven genes were more highly expressed in clone B8np in comparison with clone B2p. Two, EHI_144610 and EHI_057550, were identical and encode methionine gamma-lyase. Four B8np transfectants were generated, and all of them showed 3–70-fold overexpression in comparison with the control (S7 Table). Similarly to A1np transfectants, no significant influence on ALA formation was observed. Interestingly, large ALAs were detected 7 days post infection in 6/9 mice infected with B8np transfectant overexpressing EHI_127670 (p = 0.0589), as was observed for infections with A1np-EHI_127670 transfectants (p = 0.292) (Fig 5B and 5C).
Recent studies aiming to identify the differences between the non-pathogenic isolate HM-1:IMSS_A and the closely related but pathogenic isolate HM-1:IMSS_B using proteomic and transcriptomic analyses revealed a surprisingly small overlap between the two approaches [11, 12]. To eliminate the potential experimental limitations associated with cell population heterogeneity, we cloned both mother lines resulting in 12 clones derived from cell line HM-1:IMSS_A (A1–A12) and 12 clones derived from cell line HM-1:IMSS_B (B1–B12). As expected, all clones derived from cell line HM-1:IMSS_A were unable to form ALAs. Surprisingly, the situation was more complex for clones derived from the HM-1:IMSS_B cell line. Here, only 8/12 clones analysed displayed the pathogenic phenotype of the original cell line. To understand the phenotypically different outcomes, three clones were selected for in-depth analyses. These were pathogenic clone B2p and non-pathogenic clones A1np and B8np.
Interestingly, and in contrast to other pathogenic isolates, the pathogenicity of clone B2p remained stable for years without animal passage. Usually, trophozoites become less virulent after long-term culture and pathogenicity retainment requires regular animal passages or at least the addition of cholesterol to the culture medium [18–22].
Although the mouse model for ALA, has some limitations regarding the artificial route of infection, it is well established [23–25] and allows in this study to discriminate between pathogenic and less pathogenic clones. Accordingly, the high reproducibility of the differences in the recovery time of the liver between „non-pathogenic”and „pathogenic”E. histolytica clones opens a stable time frame that enables studying pathogenicity factors involved in liver pathology on a significant level. However, no conclusion can be drawn concerning the process of invasion into the intestinal mucosa, induction of amoebic colitis and immune evasion [26, 27].
CPs have been described as major pathogenicity factors of E. histolytica. In several studies, a direct correlation between CP activity and ALA formation was observed [19, 28–31]. In addition, ALA formation can be inhibited by specific cysteine peptidase inhibitors, and overexpression and silencing of individual E. histolytica cp genes can alter the ALAs-inducing ability of amoebae [3, 6, 32–36]. Furthermore, several studies indicate that especially EhCP-A5 is involved in the invasion process into the intestinal mucosa [37–39]. In this context, it was shown that EhCP-A5 triggers the production and release of human matrix metalloproteinases (MMPs) through inflammatory cytokine induction. Moreover it cleaves pro-MMP-3 and converts it into active MMP-3. Together, these processes are involved in the ECM remodelling required for tissue invasion [38]. Furthermore, it was shown that EhCP-A5 abrogates the MUC2 protective function by cleavage of the MUC2 C-terminus and that EhCP-A5 also plays an role in contact-dependent mucin hypersecretion during intestinal amebiasis [39]. Correlation between CP activity and ALA formation was described for the non-pathogenic cell line HM-1:IMSS_A (CP activity, ~15 mU/mg) and the pathogenic cell line HM-1:IMSS_B (CP activity, ~110 mU/mg), as well as clones B2p (CP activity, ~120 mU/mg) and A1np (CP activity, ~15 mU/mg) [3, 11]. However, determination of CP activities, especially in B-clones, did not support these observations. Prominent examples of this discrepancy are clones B1 and B12, which are non-pathogenic but have the highest CP activity of all the clones analysed (~150–200 mU/mg). Substrate gel electrophoresis experiments indicated that these different levels of CP activity result from altered abundances of all major CPs rather than from changes of individual CPs. Such lack of correlation between CP activity levels and ALA formation was reported in only one other publication, where Montfort and colleagues also used pathogenic and non-pathogenic cultures of E. histolytica isolate HM-1:IMSS [21]. However, the missing correlation between CP activity and ALA formation is not necessarily inconsistent with previous results. Therefore, it can’t be excluded that a high CP activity alone is not sufficient to induce ALAs. Other factors may be involved that, in combination with the CPs, lead to ALA formation. Furthermore, it was recently shown, that the different CPs have different impact on ALA formation. It was shown that overexpression of ehcp-a5, one of the major expressed ehcps, but also of the very low expressed ehcps, ehcp-b8, -b9, and -c13 restored the pathogenic phenotype of the non-pathogenic clone A1np, whereas overexpression of various other peptidase genes including the major expressed ehcp-a1 and ehcp-a2 had no effect on pathogenicity [3]. In addition, in the present study the expression level of the genes under culture conditions were compared between the different clones. However it was recently shown, that the expression level of some ehcp genes (ehcp-a3, -a4, -a5, -a6, -a10, -b8, -b9, and -c13) increased during ALA formation [3]. Therefore, it can be speculated, that the non-pathogenic amoebae lost their ability to regulate the expression of the peptidases under altered environmental conditions.
In addition to CP activity, other frequently used in vitro pathogenicity markers are erythrophagocytosis, haemolytic activity and cytopathic activity. In contrast with most reports that described a correlation between erythrophagocytosis and ALA formation [22, 40–43], Monfort and colleagues and Tsutsumi and colleagues, as well as this study, did not confirm this correlation [21, 44]. In the present study, non-pathogenic clone B8np showed the highest erythrophagocytosis rate, followed by clones A1np and B2p. Recently, a mechanism described as amoebic trogocytosis was identified, where amoebae ingest “bites” of host cells [45, 46]. Interestingly, only living cells were ingested by trogocytosis, whereas dead cells were phagocytosed in total. So far, we have no hint if amoebic trogocytosis correlates with the different ability of clone A1np, B2p and B8np to from ALAs. However, none of the molecules with known function in amoebic trogocytosis of erythrocytes were found to be differentially expressed in either of the three clones [45].
Haemolytic activity has also been reported as related to pathogenicity [22, 47]. However, similarly to what was described decades ago by Keller and colleagues [48], we did not observe any correlation of this trait with virulence. We showed that non-pathogenic clone B8np was unable to lyse erythrocytes, while non-pathogenic clone A1np had a higher haemolytic activity than pathogenic clone B2p. Furthermore, it was recently shown that haemolytic activity of the non-pathogenic cell line HM-1:IMSS_A is significantly higher than that of the pathogenic cell line HM-1:IMSS_B [11].
It is indisputable that cytopathogenicity is an important feature for E. histolytica pathogenicity [49, 50]. However, to the best of our knowledge, no study has correlated cytopathogenicity with virulence. A comparison of the abilities of clones A1np, B2p, and B8np to disrupt a CHO cell monolayer showed that clone A1np had the highest cytopathic activity, whereas clone B8np had no cytopathic activity. Therefore, even when cytopathic activity was taken into consideration, no correlation with ALA formation could be found.
The non-pathogenic clone A1np exhibits a higher haemolytic and cytopathic activity in contrast to clone B8np, while the motility of clone B8np is higher compared to clone A1np. From these phenotypical observations we speculate that pathogenicity of the parasite might require i) the ability to destroy and ii) phagocytose host cells and iii) to exhibit a certain motility, parameters which we find to be combined in clone B2p. Finally we conclude that the ability of E. histolytica to destroy liver tissue involves complex processes, from both, the parasite and the host side [24].
The genome of E. histolytica comprises ~8400 genes. Matching the transcriptomes of clones A1np, B2p and B8np to one another revealed that only a minority of genes were differentially transcribed. Comparing the transcriptomes of non-pathogenic clone A1np and pathogenic clone B2p revealed 46 genes that were more highly expressed in clone A1np and 30 genes that were more highly expressed in clone B2p (≥3-fold). The expression of 60% [18/31] of the genes that were differentially expressed in mother cell lines HM-1:IMSS_A and HM-1:IMSS_B [12] was also significantly different for clones A1np and B2p. These included genes encoding Rab family GTPases (EhRab7D, EHI_082070; EhRab7E, EHI_169280; EhRab 7G, EHI_187090), C2 domain-containing protein (EhC2-2, EHI_118130) and cell surface protease gp63 (EhMP8-2, EHI_042870). Comparison of the pathogenic clone B2p with the non-pathogenic clone B8np, both derived from pathogenic cell line HM-1:IMSS_B, revealed only 19 differentially expressed genes (≥3-fold). Of these, 12 were more highly expressed in clone B8np and seven were more highly expressed in clone B2p. Only six genes were regulated in the same manner in the two non-pathogenic clones A1np and B8np (EHI_026360, EHI_056490, EHI_039020, EHI_127670, EHI_144490 and EHI_144610).
Since the identified differentially expressed genes encode proteins of divergent function or hypothetical proteins, it is speculative whether these genes correlate with pathogenic amoeba phenotype. To clarify this issue, we generated overexpression transfectants for a set of candidate genes. Six genes up-regulated in clone A1np, six genes up-regulated in clone B8np and three genes up-regulated in clones A1np and B8np, in comparison with B2p, were overexpressed in B2p. Significant reduction in abscess size in comparison with the control was determined for 7/15 B2p transfectants. The respective overexpressed genes encoded two C2 domain proteins (EhC2-3, EHI_15290; EhC2-5, EHI_05980), cell surface protease gp63 (EhMP8-2, EHI_042870), two alcohol dehydrogenases (EHI_088020, EHI_160670) and two hypothetical proteins (EHI_075690, EHI_058920).
Within the E. histolytica genome, four genes (EHI_069320, EHI_118130, EHI_015290, EHI_059860) were identified as encoding C2 domain proteins that have 60–75% identity. Three of these were more highly expressed in clone A1np compared with clone B2p, and for all three the respective B2p transfectants were generated. Overexpression of EHI_015290 (EhC2-3) and EHI_059860 (EhC2-5) significantly reduced pathogenicity of clone B2p. In general, C2 domains are involved in targeting proteins to cell membranes. Thus far, only one (EHI_069320, EhC2-1) of the C2 domain proteins has been characterised. It mediates anchoring of the transcription factor URE3-BP to the amoebic plasma membrane [51].
Two cell surface protease gp63 homologues (EHI_200230, EhMP8-1; EHI_042870, EhMP8-2) are encoded in the E. histolytica genome. EhMP8-2 was more highly expressed in clone A1np in comparison with clone B2p. This was also true for the two cell lines HM-1:IMSS_A and HM-1:IMSS_B [12]. However, no differential expression was observed for EhMP8-1. Both metalloproteases belong to the M8 family zinc metalloproteases with homology to leishmanolysin, a protein essential for virulence of Leishmania [52, 53]. They contain a zinc-binding HEXXH catalytic site motif and a putative transmembrane domain, and have 34% identity with each other. EhMP8-1 is localised on the trophozoite surface, and further characterisation revealed involvement in adherence, mobility, cytopathogenic activity and phagocytosis [54]. There is no indication that EhMP8-2 exhibits similar functions, since its expression levels in different clones did not correlate with cytopathogenic activity and phagocytosis.
The expression of EHI_075690 was five times higher in A1np than in B2p. The gene EHI_075690 encodes a 218-amino acid hypothetical protein. In silico analysis revealed that the protein consists of four transmembrane domains with homology to tetraspanin family proteins. Until now, six tetraspanins were identified in the E. histolytica genome; however, their function is mostly unknown [55]. In general, tetraspanins are known to be involved in cell proliferation, adhesion, signalling and migration [56]. Recently, it was shown that the tetraspanin TvTSP8 of Trichomonas vaginalis is involved in parasite-parasite communication [57].
EHI_058920 was more highly expressed in clone B8np in comparison with clone B2p, and overexpression reduced the pathogenicity of clone B2p. The gene encodes a protein of 316 amino acids. No homologues in other organisms and no conserved domains were identified within the protein.
EHI_088020 and EHI_160670, both more highly expressed in clone B8np in comparison with clone B2p, encode alcohol dehydrogenases with the highest homology to Fe-dependent dehydrogenases of Gram-negative obligatorily anaerobic prokaryotes. Therefore, it may be assumed that they were incorporated into the amoebal genome by lateral gene transfer. As mentioned above, in silico analyses indicated that amino acid sequences of proteins encoded by EHI_088020 and EHI_160670 are identical; however, the first 160 amino acids of EHI_088020 are missing from the EHI_160670 sequence. Since EHI_160670 is located at the 5′-end of the published contig DS571485, this ‘deletion’ may be explained by a not-fully annotated status of the E. histolytica genome (AmoebaDB, http://amoebadb.org/amoeba/). However, we were unable to identify the sequence upstream of EHI_160670. Nevertheless, ectopic expression of the full-length or truncated gene affected abscess formation, as it significantly reduced the pathogenicity of clone B2p. At least ten genes encoding alcohol dehydrogenases are found in the genome of E. histolytica. Two of them show 79% (EHI_192470) and 70% (EHI_198760) amino acid sequence identity with EHI_088020. The expression levels of both genes were similar in different clones. Interestingly, EHI_198760 (EhADH3) was described to be present in lower amounts in the non-pathogenic isolate Rahman, as well as in E. dispar in comparison to E. histolytica HM-1:IMSS [10, 58]. However, no correlation between EhAH3 amount and pathogenicity was observed [58].
Interestingly, none of the three genes (EHI_026360, EHI_056490, EHI_039020) up-regulated in the two non-pathogenic clones A1np and B8np vs. pathogenic clone B2p affected ALA formation during infection with their overexpressing B2p transfectants.
Of the genes that were more highly expressed in clone B2p in comparison with clones A1np or B8np, the only one impacting ALA formation when overexpressed was EHI_127670. EHI_127670 was one of the genes that was more highly expressed in clone B2p than in the two non-pathogenic clones A1np and B8np. Transfectants of either clone ectopically expressing this gene were able to induce ALA formation. However, since the pathogenic phenotype was only observed in 4/9 animals infected with A1np-EHI_127670 transfectants and 6/9 animals infected with B8np-EHI_127670 transfectants, these results were not statistically significant. Interestingly, if the results of both non-pathogenic clones overexpressing EHI_12760 were summarised the effect on ALA formation became significant. EHI_127670 encodes a putative protein of 111 amino acids. No homologues and no conserved domains within the protein were identified.
In this study we analysed the influence on ALA formation for 20 out of the 89 differentially expressed genes identified. However, it was not possible to overexpress 4 rab protein encoding genes (EHI_082070/ehrab7d, EHI_169280/ehrab7e, EHI_187090/ehrab7g, EHI164900), which are highly expressed in clone A1np and very low expressed in clone B2p. This differential expression was also observed comparing the non-clonal cell lines A and B [12]. Rab GTPases are essential for the regulation of vesicular trafficking in the endocytic and exocytic/secretory pathways of eukaryotic cells [59]. The genome of E. histolytica contains more than 90 rab genes, including nine of the Rab7 isotype. Therefore, E. histolytica seems to be an organism with extremely diverse and complex Rab functions [60, 61]. One of the Rab7 isotypes, namely EhRab7A, is involved in transport of CPs to phagosomes and in recycling of a CP receptor from the phagosomes to the trans-Golgi network [62–64]. EhRab7A and EhRab7B are involved in lysosome biogenesis [61]. There is additional evidence that all EhRab7 isotypes are sequentially and coordinately involved in phagosome biogenesis [61]. However, so far it remains elusive whether the differential expression of the 4 rab genes indeed influences the ALA formation.
In this study, no correlation was found between the ability of E. histolytica clones to produce amoebic liver abscesses and their cysteine protease, haemolytic, erythrophagocytosis, or cytopathic activities, or their sizes or growth characteristics. However, the clones showed different expression profiles. We conclude that different mechanisms result in the loss of E. histolytica pathogenicity, because only a few genes were found to be differentially regulated in the same way when either of the two non-pathogenic clones A1np and B8np were compared with the pathogenic clone B2p. However, overexpression of seven different genes, encoding a metallopeptidase, C2 domain proteins, alcohol dehydrogenases, and hypothetical proteins in the pathogenic clone B2p correlated with reduced ability of E. histolytica to produce amoebic liver abscesses. Only one gene was identified whose overexpression transformed a non-pathogenic phenotype into a pathogenic one.
Animal experiments were carried out in accordance with the guidelines from the German National Board for Laboratory Animals and ARRIVE guidelines (https://www.nc3rs.org.uk/arrive-guidelines) and approved by the review board of the State of Hamburg, Germany (Ministry of Health and Consumer Protection/Behörde für Gesundheit und Verbraucherschutz—ethical permits 145/13, 20.01.2014)
E. histolytica trophozoites were cultured axenically in TYI-S-33 medium in plastic tissue culture flasks [65]. E. histolytica cell lines HM-1:IMSS-A and HM-1:IMSS-B were derived from the isolate HM-1:IMSS and both were originally obtained from the American Type Culture Collection (ATCC) under the catalogue number 30459[11]. HM-1:IMSS was originally isolated from a colonic biopsy of rectal ulcer from an adult male patient with amoebic dysentry in 1967 (Mexico City, Mexico). The monoxenic cultured HM-1:IMSS isolate was passed from Margarita de la Torre to Louis S. Diamond who adapted it to axenic cultivation. Thereafter, this axenically cultivated HM-1:IMSS isolate was transferred to the ATCC library. Cell line A was sent to us in 2001 by Barbara Mann (Charlottesville, University of Virginia), as a batch of cells from the same culture that was used for DNA preparation to sequence the E. histolytica genome [66]. The pathogenic cell line B was obtained directly from ATCC in 1991. Since then, the ability of cell line B to induce liver pathology remained stable.
Both cell lines were cloned by limited dilution. For this, a dilution of 120 amoebae/24 ml TYI-S-33 medium was prepared and 200 μl of this dilution was added to each well of a 96-well plate. The presence of only one amoebae/well was analysed microscopically and the trophozoites were cultivated under anaerobic conditions using Anaerocult (Merck) for one week. Afterwards the clones were transferred for further cultivation to tissue culture flasks.
For individual experiments, 1 × 106 trophozoites were cultivated for 24 h in 75 mL culture flasks. Subsequently, after chilling on ice for 5 min, trophozoites were harvested by sedimentation at 430 × g at 4°C for 5 min. The resulting cell pellets were washed twice either in phosphate-buffered saline (PBS; 6.7 mM NaHPO4, 3.3 mM NaH2PO4, 140 mM NaCl, pH 7.2) or in incomplete TYI-S-33 medium (medium without serum). To prepare amoeba extracts, cells were lysed over four freeze-thaw cycles in CO2/ethanol and sedimented by centrifugation (9000 × g at 4°C for 15 min).
Animal infections were performed with 10- to 12-week-old female gerbils obtained from JANVIER LABS (Saint Berthevin Cedex 53941 France) or with 10- to 12-week-old C57BL/6 male mice bred in the animal facility of the Bernhard Nocht Institute for Tropical Medicine, Hamburg, Germany. All animals were maintained in a specific pathogen-free environment. Animal experiments were approved by the review board of the State of Hamburg, Germany (Ministry of Health and Consumer Protection/Behörde für Gesundheit und Verbraucherschutz, (145/13, 20.01.2014) and conducted in accordance with institutional and ARRIVE guidelines (https://www.nc3rs.org.uk/arrive-guidelines).
For gerbil infections, 1 × 106 trophozoites in 50 μL of incomplete TYI-S-33 medium (without serum) were injected into the left liver lobe, as described previously [67]. For mice infection experiments, 1.25 × 105 trophozoites in 25 μL of incomplete TYI-S-33 medium were injected into the liver, as described by Lotter and colleagues [23].
To analyse ALA formation of the various amoeba clones, gerbils were sacrificed at 7 days post infection and the extent of the abscessed liver area was measured manually using a caliper and determined as size in mm2. For each E. histolytica clone, ALA formation was analysed in at least four animals. Significance (p-values) was established using the Mann-Whitney U test.
MRI was performed to analyse the time course of ALA formation using a small animal 7 Tesla MR scanner (ClinScan, Bruker Biospin GmbH, Ettlingen, Germany). For these experiments, gerbil and mouse livers were imaged in vivo on days 3, 5, 7 and 10 after intrahepatic injection of E. histolytica. Anaesthesia was performed as described by Ernst and colleagues [68]. Images were acquired using T2-weighted fast spin echo (T2w FSE) sequences for high-resolution anatomical reference. Total abscess volume was calculated by measuring the region of interest (ROI) in each abscess-containing slice, using transversal sections of the abdomen and the OsiriX Imaging Software DICOM Viewer (Open-source version 32-bit 4.1.1). Significance (p-values) was established using an unpaired t test.
E. histolytica trophozoites (1 × 106) were cultivated in 75 mL culture flasks for 24 h. The cells were harvested after chilling on ice for 5 min and sedimented at 200 × g for 5 min at 4°C. Cell pellets were washed twice with PBS. To isolate total RNA, trophozoites were treated with TRIzol reagent (Thermo Fisher Scientific, Schwerte, Germany) following the manufacturer’s instructions. Extracted RNA was further purified using the RNeasy mini kit (Qiagen, Hilden, Germany) but without β-mercaptoethanol in buffer RLT, and DNA was digested with DNase I (Qiagen, Hilden, Germany). Total RNA for transcriptomics analyses was purified using the mirVana miRNA isolation kit (Ambion-Thermo Fisher Scientific, Schwerte, Germany), according to the manufacturer’s instructions.
cDNA synthesis was obtained using the SuperScript III Reverse Transcriptase system (Thermo Fisher Scientific, Schwerte, Germany). Briefly, RNase-free and DNase-treated total RNA (1 μg in a 20 μL final volume) was mixed with 5 × First-Strand buffer, 1 mM dNTPs, 500 nM OdT-T71 (5′-GAG AGA GGA TCC AAG TAC TAA TAC GAC TCA CTA TAG GGA GAT24), 2 mM DTT, 0.5 mM MgCl2, 40 U RNaseOut (Thermo Fisher Scientific, Schwerte, Germany) and SuperScript III (200 U/μL). cDNA was synthesised for 1 h at 42°C.
For qRT-PCR experiments, sense and antisense primers were designed to amplify 100–120 bp fragments of the respective genes (S8 and S9 Tables). Quantitative amplifications were performed in Rotor-Gene PCR apparatus (Qiagen, Hilden, Germany) using RealMasterMix SYBR ROX Kit (5PRIME, Hilden, Germany). cDNA (1 μL) was mixed with 2.5 × RealMasterMix/20 × SYBR and 5 pmol/μL appropriate sense and antisense primers, to a final 20 μL volume. Amplification conditions were as follows: 40 cycles of 95°C for 15 s, 58°C for 20 s and 68°C for 20 s, and an adjacent melting step (67–95°C). Two biological replicates were analysed in duplicate. Relative differences in gene expression were calculated using the 2-ΔΔCT method with Rotor-Gene software [69]. Depending on the experiment, clone A1np, clone B2p, or clone B8np was used as the calibrator (= 1), and actin was used as the house-keeping gene for normalisation.
RNA for RNA-Seq library preparation was purified as described above. RNA quantity and quality were evaluated spectrophotometrically (NanoDrop 2000, Thermo Fisher Scientific, Schwerte, Germany) and with an Agilent 2100 Bioanalyzer with RNA 6000 Pico Assays kit (Agilent Technologies, Waldbronn, Germany).
Samples were Turbo DNase-treated using TURBO DNA-free Kit (Ambion-Thermo Fisher Scientific, Schwerte, Germany). After quality control, rRNA was depleted using RiboZero Magnetic Gold kit (Human/Mouse/Rat; Epicentre-Illumina, Munich, Germany) and Agencourt RNAClean XP kit (Beckmann Coulter, Krefeld, Germany), according to the manufacturers’ protocol. RNA-Seq libraries were then generated using ScriptSeq v2 kit (Epicentre-Illumina, Munich, Germany) according to the manufacturer’s instructions. Each library was indexed with Illumina-compatible barcodes to allow multiplexing. The individual libraries were assessed using Qubit dsDNA high sensitivity kit and Bioanalyzer DNA HS chips to ascertain the concentration (4 nM) and fragment size distribution, respectively, prior to library multiplexing. Libraries were denatured and diluted to final concentration of 8 pM for sequencing on the MiSeq platform following the manufacturer’s instructions. Reads were aligned to E. histolytica transcriptome (AmoebaDB 28, released 30 March 2016) using Bowtie 2 version 2.2.3 [70] and differential expression was analysed using DESeq [71] version 1.18.
To determine amoeba size, the circumference of 80 trophozoites of each clone was measured using a BZ9000 Keyence microscope (Keyence, Neu-Isenburg, Germany). To determine growth rate, 500 trophozoites of each clone were seeded into a 24-well plate and the cells were counted every 24 h over 72 h. The growth rate was determined three times in triplicate for each clone. The movement of the amoebae was directly filmed over a time period of 10 min using Evos FL Auto microscope from Life Technologies. A picture was taken every 5 sec. The movement of 80 amoebae/ clone was analysed manually using ImageJ version 2.0.0-rc-43/1.51d with plugins for manual tracking and chemotaxis. Significance (p-values) was established using the Mann-Whitney U test.
CP activity was measured using the synthetic peptide Z-Arg-Arg-pNA (Bachem, Bubendorf, Switzerland) as substrate [72]. One unit of enzymatic activity is defined as the amount of protein that catalyses the generation of 1 μmoL p-nitroaniline per min.
Substrate gel electrophoresis was performed as described previously [73]. Briefly, amoebic extract (2 μg) was incubated in Laemmli buffer with 20 mM DTT, for 5 min at 37°C. For the substrate gel, 12% SDS-polyacrylamide gel was co-polymerised with 0.1% gelatine. After protein separation and incubation in buffer A (2.5% Triton X-100; 1 h) and buffer B (100 mM sodium acetate, pH 4.5, 1% Triton X-100, 20 mM DTT; 3 h), at 37°C, the gel was stained with Coomassie blue.
Erythrophagocytosis assay was performed as described by Biller and colleagues [11]. Human 0+ erythrocytes were provided by the blood bank of the University Medical Center Hamburg-Eppendorf (UKE)–Transfusion Medicine–Germany. Human erythrocytes and trophozoites were washed twice with serum-free TYI-S-33 medium. Erythrocytes and amoebae were mixed at a 1000:1 ratio (2 × 108 erythrocytes, 2 × 105 amoebae), to a final volume of 400 μL, in serum-free TYI-S-33 medium and incubated in parallel at 37°C for 30 min. To stop phagocytosis and lyse non-phagocytosed erythrocytes, 1 mL of distilled water was added, twice. Trophozoites were washed twice with PBS. Average numbers of ingested erythrocytes were quantified by measuring the absorbance at 397 nm after trophozoite lysis in 90% formic acid. The experiment was performed three times in triplicate. Significance (p-values) was established using the Mann-Whitney U test.
Haemolytic activity assay was performed as described by Biller and colleagues [11]. Human erythrocytes and trophozoites were washed three times with PBS. The assay was performed by mixing trophozoites and erythrocytes in a 1:2000 ratio (2 × 105 amoebae with 4 × 108 erythrocytes per mL of PBS), followed by incubation for 1 h at 37°C. After incubation, the cells were sedimented for 1 min at 2000 × g. Haemoglobin released into the supernatant was measured at 570 nm in a spectrophotometer. Separately incubated erythrocytes and trophozoites were used as negative controls. To determine 100% haemoglobin release, 4 × 108 erythrocytes were lysed in 1 mL of water. The experiment was performed three times in triplicate. Significance (p-values) was established using the Mann-Whitney U test.
Interaction of trophozoites and Chinese hamster ovarian (CHO) cells was determined by a modified method of Bracha and Mirelman [74]. CHO cells defective in glycosaminoglycan biosynthesis (CHO-745; American Type Culture Collection No. CRL-2242) were used. CHO cells (1 × 105 per well) were grown for 24 h in 24-well plates in Ham’s F12 (with l-Glutamine) medium supplemented with 10% fetal calf serum (FCS) and penicillin-streptomycin. After washing the CHO cells with preheated (37°C) Ham’s medium, 500 μL of Ham’s medium was added. E. histolytica trophozoites (1 × 105) were washed twice with serum-free TYI-S-33 medium, resuspended in 500 μL of ABS-free TYI-S-33 and added to the CHO cells. The mixture was incubated for 20 min at 37°C under 5% CO2. Cells were washed with 1 mL of ice-cold PBS and treated with 0.5 mL of 4% paraformaldehyde in PBS for 2 min. After another PBS wash, the cells were stained with 500 μL of 0.1% methylene blue for 2 min. Finally, the cells were washed with 0.01% methylene blue and PBS. Cells were lysed with 1 mL of 0.1 M HCl for 30 min at 37°C. Samples were photometrically analysed at 660 nm. As a control, methylene blue concentration was determined for CHO cells that had not been co-cultivated with trophozoites (i.e., no destruction of cell monolayer). Experiments were performed three times in sextuplicate. Significance (p-values) was established using the Mann-Whitney U test.
All plasmids used for E. histolytica trophozoite transfections are derivatives of the expression vector pEhNEO/CAT (pNC) [75, 76]. Genes of interest were amplified by PCR using genomic E. histolytica DNA as a template, cloned into TOPO TA vector, sequenced and cloned into pNC using KpnI and BamHI restriction sites (S9 Table). For overexpression, coding sequences of the genes of interest were flanked by 485 bp 5′-untranslated sequence of the E. histolytica lectin gene and 600 bp 3′-untranslated region of the actin gene. Neomycin phosphotransferase was used as a selectable marker.
Transfections were performed by electroporation as described previously [76]. Two days post transfection, cells were transferred to a selection medium containing 10 μg/mL G-418 sulphate, for approximately 2 weeks. Subsequently, the cells were cloned by a limited dilution method and cultivated in the presence of 20 μg/mL G418. Successful overexpression of at least four clones was checked by qRT-PCR. For infection experiments, trophozoites were cultivated for 24 h in the absence of G418.
|
10.1371/journal.pntd.0000987 | Evaluation of Mammalian and Intermediate Host Surveillance Methods for Detecting Schistosomiasis Reemergence in Southwest China | Schistosomiasis has reemerged in China, threatening schistosomiasis elimination efforts. Surveillance methods that can identify locations where schistosomiasis has reemerged are needed to prevent the further spread of infections.
We tested humans, cows, water buffalo and the intermediate host snail, Oncomelania hupensis, for Schistosoma japonicum infection, assessed snail densities and extracted regional surveillance records in areas where schistosomiasis reemerged in Sichuan province. We then evaluated the ability of surveillance methods to identify villages where human infections were present. Human infections were detected in 35 of the 53 villages surveyed (infection prevalence: 0 to 43%), including 17 of 28 villages with no prior evidence of reemergence. Bovine infections were detected in 23 villages (infection prevalence: 0 to 65%) and snail infections in one village. Two common surveillance methods, acute schistosomiasis case reports and surveys for S. japonicum-infected snails, grossly underestimated the number of villages where human infections were present (sensitivity 1% and 3%, respectively). Screening bovines for S. japonicum and surveys for the presence of O. hupensis had modest sensitivity (59% and 69% respectively) and specificity (67% and 44%, respectively). Older adults and bovine owners were at elevated risk of infection. Testing only these high-risk human populations yielded sensitivities of 77% and 71%, respectively.
Human and bovine schistosomiasis were widespread in regions where schistosomiasis had reemerged but acute schistosomiasis and S. japonicum-infected snails were rare and, therefore, poor surveillance targets. Until more efficient, sensitive surveillance strategies are developed, direct, targeted parasitological testing of high-risk human populations should be considered to monitor for schistosomiasis reemergence.
| Schistosomiasis has reemerged in China in regions where it was previously controlled. As reductions in schistosomiasis, a water-born parasitic infection, prompt consideration of schistosomiasis elimination, surveillance strategies that can signal reemergence and prevent further lapses in control are needed. We examined the distribution of Schistosoma japonicum, the species that causes schistosomiasis in China, in 53 villages. The villages were located in regions of Sichuan province where schistosomiasis reemergence had been documented by public health authorities. We tested three key reservoirs, humans, cows and water buffalo, and freshwater snails for S. japonicum infection in an effort to identify high-risk populations and evaluate their ability to signal reemergence. Human and bovine infections were common, detected in 35 villages and 23 villages, respectively, but infected snails were rare, found in only one village. Two commonly used surveillance methods, hospital reports of acute schistosomiasis and surveys for S. japonicum-infected snails, grossly underestimated the number of villages where human infections were present. Schistosomiasis was widespread in the region we studied, highlighting the danger reemergence poses to disease elimination programs. Surveillance systems that monitor high-risk populations such as older adults or bovine owners should be considered to promote detection of reemergence.
| The success of disease control programs in reducing schistosomiasis infections and morbidity have prompted consideration of the elimination of human schistosomiasis [1], [2]. Dramatic declines in Schistosoma haematobium and S. mansoni have been observed following widespread distribution of the antihelminthic drug, praziquantel, in six countries in sub-Saharan Africa [3]–[5]. Disease control efforts in China, including a ten-year partnership with the World Bank to promote treatment, have led to the interruption of S. japonicum transmission in 5 of 12 endemic provinces and 60% of endemic counties [6], [7]. Currently, China is aiming to eliminate schistosomiasis, setting an initial goal of reducing human and bovine infection prevalence below 1% in every endemic region by 2015 [8]. If successful, China's program may serve as a model for schistosomiasis control elsewhere.
However, schistosomiasis has reemerged in previously controlled regions, highlighting the challenges of sustaining reductions in infections. In Sichuan, China, schistosomiasis was identified in 8 of 46 counties that had met Chinese Ministry of Health criteria for transmission control, which require the reduction of human and bovine infection prevalence below 1% in every endemic village [9]. Nationwide, 38 counties that have met transmission control criteria have been reclassified as reemerging [7]. In the absence of a vaccine or lasting immunity, and with at least forty competent mammalian reservoirs, S. japonicum reemergence remains a threat in controlled regions [10]. Little is known about the epidemiology of reemerging schistosomiasis, including how infections are distributed across human populations, other mammalian reservoirs and intermediate host snails.
Surveillance systems that can identify areas where lapses in control have occurred are an essential component of disease elimination strategies [11], [12]. They can enable timely treatment of infected populations and interventions to prevent further spread of infections. In China, surveillance for schistosomiasis in controlled regions includes hospital-based surveillance for acute schistosomiasis, surveys for the intermediate host snail, Oncomelania hupensis, and direct testing of the human population [13]. Acute schistosomiasis is triggered by the migration of the parasite through the body shortly after infection, leading to rapid onset of symptoms including high fever, myalgia and eosinophilia [14]. Due to the quick and severe onset, acute schistosomiasis, which is a reportable disease in China, can serve as a sentinel event, signaling the reemergence or emergence of schistosomiasis, as occurred in the Yangtze River valley following flooding events and in Sichuan province [9], [15], [16]. But acute schistosomiasis is rare, comprising less than 1% of all schistosomiasis cases [17]. It is possible for schistosomiasis to reemerge more quietly, as schistosomiasis typically induces chronic morbidity [18]–[20], leading to uncertainties about the sensitivity of surveillance methods that rely on acute schistosomiasis case reports. Similarly, surveillance for schistosome-infected snails and children is recommended in regions approaching schistosomiasis elimination, but how well these surveillance targets can identify areas where human infections are present remains uncertain [21].
In an effort to inform surveillance for S. japonicum reemergence, we examined the distribution of S. japonicum infections in human, bovine and snail populations in regions where schistosomiasis had reemerged. We then evaluated the ability of active and passive surveillance methods, including acute schistosomiasis case reporting, surveys for the presence of O. hupensis, and surveys for S. japonicum infections in O. hupensis, cows, water buffalo, and high-risk human populations, to identify villages where human infections were present.
Three of the eight counties where schistosomiasis reemergence was detected in Sichuan province following transmission control were selected for inclusion in this study [9]. Counties were selected based on the availability of surveillance records and the willingness of the control station personnel to collaborate on this project. Due to the sensitive nature of conducting infection surveys in regions where schistosomiasis transmission control criteria have officially been met and to promote candid reporting, the names and exact locations of the counties and study villages have been withheld.
County surveillance records were examined in March 2007 in order to identify all villages where S. japonicum had been detected after the attainment of transmission control. We examined reemergence at the smallest community unit, the natural village or production group (referred to here as, simply, village) which generally includes 100 to 200 residents. Reemergence criteria were based on typical post-control surveillance. This includes passive monitoring for acute schistosomiasis through health providers, and active surveillance for S. japonicum-infected snails and humans through surveys conducted at least once every three years in formerly endemic areas. Human surveys in controlled areas often focus on children, as S. japonicum infections in children who were born after transmission control was attained indicate renewed transmission. Records were examined from the year that transmission control was attained through March 2007.
We classified villages as historically reemerging (HR) if schistosomiasis was endemic prior to transmission control and acute human schistosomiasis, a S. japonicum-infected snail or a S. japonicum infection in a child younger than 12 years was detected after transmission control. County surveillance records were supplemented by provincial surveillance data when available. We identified 112 HR villages, including 109 villages identified through county surveillance records and three villages identified through provincial records (Table 1). Infected snails were the most frequent indicator of reemergence. Acute schistosomiasis and infected children were less commonly detected. Evidence of reemergence was first detected in County 1 five years after transmission control criteria were met, the same year as reemergence was detected in County 2, which borders County 1 and had attained transmission control 15 years earlier.
Twenty-five villages were selected from the HR villages. Villages were stratified by county and how reemergence was indicated and a sample was selected from each stratum. The selected villages include 11 villages with acute schistosomiasis, 19 with infected snails and one village with infected children (five villages had multiple indicators), detected from 1997 through 2005. For comparison, 28 villages were selected from formerly endemic villages in the same counties with no history of reemergence (NHR) since transmission control was achieved. In each of the 53 selected villages, village residents were interviewed to describe demographic and household characteristics, and human, bovine and snail infection surveys were conducted.
In June 2007, all residents aged six years and older in the 53 selected villages were invited to complete a brief survey about their age, sex, occupation, highest level of schooling, travel and schistosomiasis treatment history. The head of each household was also asked to complete a detailed questionnaire describing household agricultural practices, ownership of domestic animals and socioeconomic indicators. Given the challenges of estimating income in agrarian regions, household socioeconomic status was assessed based on household assets [22]. The head of each household was asked if any member of his or her household owned a car, tractor, motorcycle, computer, television, washing machine, air conditioner or refrigerator. A household asset score was assigned based on the number of assets owned. In 2008, attempts were made to interview any participant in the human infection survey missing household or individual interview data from 2007.
Questionnaires were pilot-tested to ensure questions were appropriate for the study region. Interviews were conducted by trained staff at the Institute of Parasitic Diseases (IPD), Sichuan Center for Disease Control and Prevention and the county Anti-schistosomiasis Control Stations fluent in Sichuan dialect, which is spoken by the study population. Household and individual interviews were scanned using optical mark recognition software (Remark Office OMR, Gravic Corporation, Malvern, PA). Approximately 10% of scanned questionnaires were checked against paper records to ensure data accuracy.
In November and December 2007, all residents aged 6 to 65 years were invited to submit three stool samples from three consecutive days which were analyzed using the miracidium hatching test and the Kato-Katz thick smear procedure [23], [24]. Of 3,009 participants, three stool samples were collected from 2,504 participants (85%), two samples from 7% and one sample from 8%. Samples were collected from villages daily and brought to a central laboratory in each county where they were stored out of direct sunlight until processing (90% were processed within one day of collection).
The miracidium hatching test was used to examine each stool sample. Approximately 30 g of stool were suspended in aqueous solution, strained with copper mesh to remove large particles, then strained with nylon mesh to concentrate schistosome eggs. This sediment was re-suspended and left in a room with ambient temperatures between 28 and 30°C. Two, five and eight hours after preparation, samples were examined for the presence of miracidia for at least two minutes each time.
Using the Kato-Katz thick smear procedure, three slides were prepared using 41.7 mg of homogenized stool from the first sample submitted by each participant. Three slides were prepared for 97% of infection survey participants (two slides were prepared for 21 participants, one slide was prepared for 21 participants and no slides were prepared for 55 participants). Each slide was examined using a dissecting microscope and if any S. japonicum eggs were detected, the species and number of eggs was confirmed by a second reader. Infection intensity, expressed in eggs per gram of stool (EPG), was calculated as the total number of S. japonicum eggs divided by the total sample weight. A person was classified as infected if the miracidium hatching test was positive or at least one egg was detected using the Kato-Katz technique.
The domestic bovines, water buffalo (Bubalus bubalis) and cows (Bos taurus), were tested for S. japonicum infection at the same time as the human surveys. Attempts were made to collect three stool samples from all bovines in study villages by keeping the animal in a pen or tied until stool was produced on three separate days. Samples were collected shortly after defecation and from the center of fresh stool samples in order to minimize potential contamination. Three samples were collected from 68% of the 537 bovines tested, two samples from 18% and one sample from 14%. Each sample was examined using the miracidium hatching test as described above. Due to the rapid hatching and short survival of miracidia in bovine stool, samples were examined one, three and five hours after preparation and 99% of samples were processed within one day of collection. Bovines were classified as infected if at least one hatching test was positive for S. japonicum.
Bovines with at least one positive hatching test were subsequently examined using an adaptation of the Danish Bilharziasis Laboratory (DBL) method in order to estimate infection intensity [25]. Briefly, 5 g of homogenized stool were washed through a series of three sieves (mesh size: 400 µm, 100 µm and 45 µm). The material in the 45 µm sieve was washed into a sedimentation tube, two drops of formalin were added and the suspension was left in the dark to sediment. The solution was centrifuged and the top half of the liquid gently decanted. The remaining sediment was re-suspended, adding enough water to create 10 mL of solution, re-centrifuged, and the top 80% of the solution gently decanted in order to obtain 2 mL of solution. A thick smear approach was used to count the eggs. Infection intensity, expressed in EPG, was calculated as the total number of eggs divided by the sample weight.
In April 2007, all irrigation ditches in the study villages were surveyed for O. hupensis. Teams of trained IPD and county Anti-Schistosomiasis Control Station staff with extensive experience conducting snail surveys collected samples at 10 m intervals along irrigation ditches. At each location, a square frame (kuang) measuring 0.11 m2 was placed at the waterline and all O. hupensis snails within the frame were collected. In addition, snails were sampled from 10 terrace walls per village (or all terrace walls if there were fewer than 10 terraces), since the lower portions of terrace walls accumulate moisture and may provide suitable habitat for snails. The sampling frame was placed at the base of the terrace walls in three locations: the middle and both ends of the terrace and all snails within the frame were collected. Collected snails were deposited in paper envelopes and brought to the laboratory where they were crushed between two glass slides and inspected for cercariae using a dissecting microscope.
All adult participants provided written, informed consent before participating in this study. All children provided assent and their parents or guardians provided written, informed permission for them to participate in this study. The research protocol was approved by the Sichuan Institutional Review Board and the University of California, Berkeley, Committee for the Protection of Human Subjects. Each person who tested positive for S. japonicum was provided treatment with 40 mg per kg of praziquantel tablets by the county Anti-Schistosomiasis Control Station. Because bovine stool samples were collected after they were excreted, the Animal Care and Use Committee at the University of California, Berkeley determined the protocol was exempt from review. All bovines testing positive were referred to the county veterinary station for treatment with praziquantel. The reporting of this cross-sectional study was evaluated using the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) checklist (Checklist S1).
In order to identify subpopulations at high risk of S. japonicum infection in reemerging areas, human infection prevalence and intensity were examined across 11 demographic variables: age, sex, occupation, educational attainment, socioeconomic status, time spent out of the village, bovine ownership and whether members of the household plant rice, corn, wheat or rapeseed. Similarly, bovine infection prevalence was examined across species, sex and age (as reported by owner). Variables were selected based on their potential to affect exposure to S. japonicum, and the ease with which surveillance teams could identify individuals based on the selected characteristics. For each demographic variable, we estimated an odds ratio (OR) and 95% confidence interval (CI) using logistic regression, modeling infection status as a function of the demographic variable, adjusting for county of residence and village HR status. For humans, we also tested the hypothesis that, among the infected, infection intensity is predicted by these same 11 demographic variables. We estimated the arithmetic mean EPG for each subpopulation. We modeled infection intensity, in EPG, as a negative binomial distribution because even among the infected, infection intensity was overdispersed as is common for helminthic infections (note that 55% of those who tested positive for S. japonicum had zero detectable eggs, testing positive by the miracidium hatching test only). Logistic and negative binomial models accounted for correlation of infections within villages using generalized estimating equations (GEE) with exchangeable correlation and with inference from robust variance estimates [26], [27].
We evaluated five types of schistosomiasis surveillance methods for their ability to identify villages where human infections were present: acute schistosomiasis case reports, surveys for the presence of O. hupensis, and surveys for S. japonicum infections in O. hupensis, bovines and high-risk human populations. The sensitivity and specificity of each method was calculated using the human infection survey results as the gold standard: villages where at least one person tested positive for S. japonicum were classified as true positives, villages where all people tested negative for S. japonicum were classified as true negatives. Bootstrapping was used to estimate the variability around each point estimate [28]. We drew 1,000 random samples of the 53 villages with replacement, estimated the sensitivity and specificity for each sampled population, and the 2.5th and 97.5th percentile values were used to generate nonparametric 95% CIs. The sensitivity and specificity of testing high-risk human populations were estimated for characteristics that were significant predictors of human infection status and that defined less than 50% of the population.
The village selection procedure involved an oversampling of villages where surveillance records indicated reemergence generally and where acute cases had been reported, in particular. Failure to account for this selection procedure can lead to biased estimates of the sensitivity and specificity of acute case reporting. We therefore applied differential weights in estimating the sensitivity and specificity of acute case reporting. Villages where acute cases were present were weighted by where n is the total number of villages formerly endemic for schistosomiasis and na is equal to 25, the number of villages where one or more acute schistosomiasis cases were reported in the three counties. Villages where acute cases were not present were divided into two subgroups: HR villages where no acute cases were found, which were weighted using , and NHR villages, weighted using where np is equal to 112, the number of formerly endemic villages where surveillance records indicated reemergence. Both np and na were directly measured through our examination of surveillance records. The number of previously endemic villages, n, is not known precisely and was estimated, based on conversations with county schistosomiasis control officers, to be 1,000. Given the uncertain value of n, we conducted a sensitivity analysis, estimating sensitivity and specificity setting n to lower and upper bound estimates (500 and 5,000 villages, respectively).
Tests of statistical significance were conducted setting α = 0.05. Data analysis was conducted using Stata, version 10 (StataCorp, College Station, TX, USA) and R, version 2.9.2 (www.r-project.org).
Beginning in June 2007, we enrolled 4,399 study participants in 53 villages and 1,784 heads of household completed the household questionnaire. There were, on average, 83 participants per village (range: 30, 169).
Most adults (≥18 years) were farmers (96.5%) as shown in Table 2. Rice and corn were the principal crops planted in the summer months (typically, May to September), whereas rapeseed and wheat were the major crops planted during the winter growing season (typically, October to April). Socioeconomic status was modest: 61.3% of individuals reported their household owned no more than two of eight specified household assets. The most commonly reported asset was a television, owned by 94.4% of households, followed by a washing machine (55.5%), motorcycle (33.1%) and refrigerator (16.5%). Tractors (2.8%), air conditioners (2.0%), cars (1.7%) and computers (0.5%) were rare.
There were few teenagers and young adults among the study population relative to other age groups (Figure 1). Residents reported many people, particularly younger populations, had left their villages to find work in urban areas. There is also a dip in the population corresponding with the birth years 1959 to 1961, the years of the Chinese famine. Travel outside of the village was common among study participants. Most young adults (aged 18–29 years) reported spending at least one month out of their village in the past year, as did 40% of adults aged 30–39 years and 26% of adults aged 40–49 years. Most of these individuals left to work as laborers. Some teenagers (aged 12–17 years) also reported living outside of their village for more than one month in the past year (21%), primarily to attend school.
Schistosomiasis control was ongoing in the study regions and occurred in both HR and NHR villages. Praziquantel treatment in 2006 or 2007 (prior to the infection survey) was reported by 45% of residents in HR villages and 40% of residents in NHR villages. County records indicated that the molluscicide, niclosamide, was applied to snail habitats in 23 of the 25 HR villages and 25 of the 28 NHR villages in 2006.
We tested 3,009 people from the 53 villages for S. japonicum infection. Participation in the infection surveys ranged from 45% to 97% by village and varied by county, age, occupation and travel out of the village (Table 2). Participation was lowest for individuals aged 12–29 years.
In 2007, 195 people (6.5%) tested positive for S. japonicum, including 159 who tested positive using the miracidium hatching test and 88 who tested positive using the Kato-Katz thick smear procedure. Human infections were detected in 35 villages, including 18 of the 25 HR villages, and in 17 of the 28 NHR villages. Human infection prevalence ranged from 0 to 42.9% by village (Figure 2), and varied by county and HR status (Table 3). Mean infection intensity was 1.6 EPG, and the maximum village infection intensity was 10.6 EPG. S. japonicum eggs were clustered in a few individuals: 24% of all eggs detected were excreted by two individuals.
We identified several groups at high risk of infection on the basis of demographic characteristics (Table 3). Infection prevalence was highest in adults, aged 50 years and older, individuals living in households planting rapeseed, individuals in households with one or more bovines and adults with no formal schooling beyond elementary school. These variables were significant predictors of infection status, controlling for county and village HR status. Because most residents in our study region lived in households planting rapeseed (86%) and were adults with less than an elementary school education (58%), only bovine ownership and older age were examined as target populations for human surveillance (comprising 37% and 39% of the population, respectively). Human S. japonicum infections did not vary by sex, occupation, household asset score or time spent out of the village.
Among the infected, infection intensity did not vary significantly by any of the demographic characteristics examined. Infection intensity did increase with age but was highest in adults aged 40 to 49 years, rather than the oldest, most frequently infected age group.
We identified 821 cows and water buffalo present in 50 of the 53 villages. We tested 537 (65.4%) bovines from 44 villages for S. japonicum infection. The six villages where bovines were present but infection surveys were not conducted had fewer bovines (mean 5.8 bovines, range: 3, 8) compared to villages where infection surveys were conducted (mean 17.9 bovines, range: 2, 42).
Bovine infections were detected in 23 villages, including 15 of the 19 HR villages where bovines were tested and 8 of the 25 NHR villages where bovines were tested. Mean bovine infection prevalence was 13.4%, with a maximum village infection prevalence of 65.4% (Figure 2). Like human infection prevalence, bovine infection prevalence varied by county and village HR status (Table 4). We measured infection intensity in 67 of the 72 bovines positive by the miracidium hatching test: 11 had detectable eggs. While bovine infection prevalence was higher than human infection prevalence, mean bovine infection intensity was lower, ranging from 0 to 0.11 EPG by village.
Infection prevalence was modestly but not significantly higher in cows compared to water buffalo, adjusting for county and village HR status (Table 4). Bovine sex did not predict infection status. Water buffalo infections declined with age whereas cow infections increased with age (Figure 3).
We surveyed a total of 15,054 locations along irrigation ditches and 2,498 locations at the base of terrace walls for snails. O. hupensis were present in 38 of the 53 villages surveyed, found along irrigation ditches in 34 villages and terrace walls in 17 villages. Mean snail density by village was 2.3 snails/m2 in irrigation ditches (range: 0, 24.8 snails/m2) and 0.6 snails/m2 in terrace walls (range: 0, 8.5 snails/m2). The density of snails in terraces was not strongly correlated with snail densities in irrigation ditches (Spearman's rho 0.302).
Of the 7,325 snails collected from irrigation ditches, one infected snail was found. None of the 190 snails collected from terrace walls were infected.
Acute case reporting and surveys for S. japonicum-infected snails yielded very low sensitivity: 1% and 3%, respectively (Table 5). Setting n to 500 and 5,000 yielded sensitivities of 3% and 0.3%, respectively for acute case reporting. The presence of snails in irrigation ditches yielded higher sensitivity (69%) and sensitivity was improved further when terraces were also sampled (74%); however specificity was low (44% and 33%, respectively). Surveys for S. japonicum-infected bovines had modest sensitivity (59%) and specificity (67%).
Testing only individuals who belong to high-risk groups provided the most accurate indicator of the presence of human infections in a village. Testing individuals aged 50 years and older or individuals in households that own bovines correctly identified 77% and 71% of villages where human infections were present, respectively. Specificity was 100%, as a single infected human was sufficient to designate a village as infected.
The use of only one of the two human infection testing methods led to decreased sensitivity. Testing all village residents using three hatching tests from three stool samples yielded a sensitivity of 97%. Testing all village residents by preparing three Kato-Katz slides from a single stool sample yielded a sensitivity of 80%.
Human S. japonicum infections were present in 35 of the 53 villages surveyed, with infection prevalence exceeding 20% in six villages, indicating widespread reemergence of schistosomiasis in areas where it had previously been controlled. Schistosomiasis was also detected in 13% of bovines, suggesting non-human mammalian reservoirs may play a role in the reemergence of schistosomiasis. Two key surveillance strategies, acute schistosomiasis reporting and surveys for S. japonicum-infected snails, grossly underestimated the number of villages where schistosomiasis had reemerged, misclassifying over 90% of villages where human infections were present. Alternative surveillance strategies, including surveys for the presence of the intermediate host snail or S. japonicum infections in bovine populations, yielded modest sensitivity and specificity. Testing high-risk human populations for S. japonicum infection appears to be the most reliable currently available strategy for monitoring the reemergence of human schistosomiasis, apart from testing all at-risk populations.
The reemergence of schistosomiasis documented here, in Sichuan province, and elsewhere underscores the challenge of sustaining reductions in human infections. Local elimination of human infections has proven difficult: a rise in S. japonicum infection prevalence was detected in endemic regions of China in the 2004 national infection survey [29]. The dramatic declines in S. haematobium and S. mansoni infection after widespread administration of praziquantel in West Africa were followed by an increase in infections in some areas [30]. China has developed a rigorous process for certifying reductions in infection, a process that accounts for spatial heterogeneity in transmission, the clustering of infections in few individuals and temporal variations in infection prevalence. To attain transmission control every endemic village in a county seeking certification must demonstrate that human infection prevalence is below 1% by testing at least 95% of the at-risk population. Villages are randomly selected for re-testing to confirm first-round results. All of the counties in this study had previously met transmission control criteria and therefore, every village in this study had, at one point, reduced infection prevalence below 1%. The fact that human infections were detected in 35 of these villages, at prevalences reaching 43%, highlights the instability of S. japonicum control, even when very low infection thresholds are attained, and the need for ongoing monitoring for reemergence.
The presence of at least 40 competent mammalian S. japonicum hosts, the potential for the parasite to be transported from endemic to controlled areas and the absence of a vaccine or lasting immunity contribute to the difficulty of local schistosomiasis elimination [10]. Bovine infections were present in 25 villages and bovine infection prevalence exceeded human infection prevalence in 18 of these villages, suggesting non-human reservoir species may be an important source of S. japonicum eggs in reemerging regions. The importance of different reservoir hosts appears to vary regionally. While previous studies have not found bovines to be important reservoirs in the endemic, hilly regions of China [31], [32] or in the Philippines [33], [34], in the hilly regions we studied, they may contribute to reemergence risk. Current efforts to replace bovines with tractors may reduce the risks posed by bovines [8], but, as rodents and dogs have been shown to play an important role in transmission in regions where bovines are absent [31], bovine removal may lead to a shift in the importance of other mammalian reservoirs. Connections between villages can sustain endemic schistosomiasis transmission and may promote the reintroduction of the parasite through social or hydrological networks [35]. In this study, schistosomiasis reemergence was detected the same year in two adjacent counties, and several villages with high human infection prevalence were located near each other on either side of the county line, suggesting infections may have spread across village and county boundaries. In addition, 34% of residents reported traveling outside of their villages in the past year, providing a possible mechanism for parasite import and export.
Treatment of human populations with praziquantel and the application of molluscicides were ongoing in the study region, likely prompted by the discovery of reemergence in these areas. This may explain why no human infections were detected in 7 of the 25 villages with historical evidence of reemergence and why snail populations were lower than observed elsewhere in Sichuan province [32]. However, the number of human and bovine infections detected in the region, despite ongoing treatment, underscores the need for improved surveillance.
Given the ongoing threat of schistosomiasis reemergence in controlled areas, timely and accurate detection of lapses in control are necessary to sustain disease control. Several types of surveillance methods have been used to monitor emerging and reemerging infections generally. First, infections may be detected through reporting of an acute presentation of a disease by health care providers. This has been a central strategy in monitoring poliomyelitis: acute flaccid paralysis, like acute schistosomiasis, is severe, rare – seen in approximately 1% of those infected with poliovirus – and occurs shortly after infection, providing a timely marker of renewed transmission [36]. China has a sophisticated disease reporting system [37], making surveillance for acute schistosomiasis appealing as it requires little additional capital or labor, although underreporting remains a concern. However, acute schistosomiasis reporting yielded poor sensitivity, identifying only 1% of villages where human S. japonicum infections were present.
Second, infections may be detected by monitoring non-human reservoirs, vectors or intermediate host populations. Non-human surveillance methods present the opportunity to detect the return of the parasite before human infections have occurred, a benefit that has been recognized in the design of surveillance systems to monitor emerging arboviruses [38], [39]. Snails, water buffalo, cows and other mammalian reservoirs have the potential to provide a key source of S. japonicum in reemerging regions and therefore may serve as sentinel populations. In our study regions, infected snails were rare, and therefore were a poor indicator of reemergence. Because we found only one infected snail out of over 7,000 examined in April 2007, the snail survey was repeated in half of the villages in September 2007. Again, only one infected snail was detected. Surveys for the presence of O. hupensis yielded modest sensitivity, but the method yielded numerous false positives. Given the low specificity, surveys for the presence of O. hupensis could be the first step in a two-part screening method that uses a more sensitive method to screen villages where snails are detected. The performance of surveys for the presence of O. hupensis may be different in regions where transmission has been controlled and snail control activities have been halted. Monitoring for the presence of S. japonicum infections in bovines offered modest sensitivity and specificity.
Third, surveillance may involve direct testing of human populations at high-risk of infection. Focusing human testing on a high-risk sample of the population as is currently being done to monitor the progress of lymphatic filariasis elimination [40], can increase the efficiency of surveillance in human populations, reducing the number of samples that need to be collected and tested. In this study, we identified two sentinel populations based on characteristics that can be readily identified by village leaders: older age and bovine ownership; that composed less than 40% of the total population. As this is the first population-level study of human S. japonicum infection in reemerging areas, research in other regions is needed to assess the generalizability of the high-risk characteristics we identified. However, our finding that infection prevalence was highest in older age groups is consistent with other studies of S. japonicum in China and the Philippines [29], [41], and supports our conclusion that children make poor surveillance targets for S. japonicum. In contrast, S. haematobium and S. mansoni infections typically concentrate in children and teenagers [42] suggesting that sentinel age groups may vary across schistosome species. Bovines have been shown to play a key role in endemic schistosomiasis transmission in the lakes and marshland regions of China [43], [44], suggesting bovine ownership may also indicate human reemergence risk in other regions, but this may not be the case in the Philippines [33]. Schistosomiasis is strongly associated with poverty, thus it is not surprising that higher educational attainment appeared protective against S. japonicum infection [45], [46]. However, as with rapeseed planting, the highest risk groups comprised over 50% of the population, leading to minimal gains in efficiency.
High-risk human population monitoring was the most accurate alternative to testing all at-risk people, outperforming bovine and snail-based surveillance as well as acute schistosomiasis reporting. Longitudinal studies, currently underway, will aid in identifying appropriate sampling intervals, as will mathematical models of reemergence over time and space. An analysis of the costs-effectiveness of different surveillance strategies can further aid in the development of an effective and feasible surveillance protocol in areas approaching schistosomiasis elimination.
Sentinel human populations may also be defined by geographic and environmental characteristics. Even within a relatively homogeneous region in terms of demographic characteristics and disease control measures, infection prevalence and intensity may vary widely between villages, as observed in this study and elsewhere [32], [47]. Local variations in factors such as rainfall, sanitation and intermediate host habitat may promote or impede the acquisition of human parasitic infections and ultimately, the reemergence of human infections [48], [49]. The characterization of local environments that are at high risk of schistosomiasis reemergence can further refine the definition of human surveillance targets.
As schistosomiasis infections decline, highly sensitive diagnostic tests will be needed to identify remaining infections. We used two stool-based testing methods, the Kato-Katz thick smear procedure and the miracidium hatching test to identify human infections. These methods are highly specific but, due to variability in egg excretion by infected individuals, the sensitivities of these tests decline when infection intensity is low [50], [51]. The use of multiple diagnostic tests and the collection of stool samples on multiple days increase the likelihood of detecting infections. Nonetheless, the infection prevalences detected here are likely lower bound estimates. PCR-based methods to detect schistosome eggs in the stool of human and other mammalian hosts show promise and may provide the sensitivity needed to diagnose very low intensity S. japonicum infections in regions approaching schistosomiasis elimination [52], [53]. Similarly, PCR-based methods to identify the presence of cercariae in water may aid in the identification of environments where the parasite remains endemic [54]. Pooled analysis of snails, water samples or even mammalian stool for S. japonicum may increase the efficiency of population-based surveillance at low infection intensities [55].
The sensitivity and specificity estimates presented here are calculated for the methods as we performed them. For example, surveys for the presence of O. hupensis in irrigation ditches were conducted by sampling all irrigation ditches in a village at 10 m intervals – if fewer locations are sampled, the probability of finding the snail host, and therefore the sensitivity of the test, may be diminished. Similarly, immunoassays are often used to screen humans for S. japonicum infection in China, and, due to the low specificity of this assay, individuals with a positive immunoassay result are then asked to provide a stool sample for examination using the miracidium hatching or Kato-Katz test, as was done in the 2004 national schistosomiasis survey [29]. As individuals with a negative immunoassay are not screened using coprological methods, this two-step screening produces false negatives from both diagnostic methods, yielding an overall sensitivity lower than that of the immunoassay alone. Targeted sampling of high-risk populations, using an immunoassay based method will yield different sensitivities than those calculated here. Similarly, relative distributions of S. japonicum infections in bovine, snail and human populations may vary regionally, a factor that should be considered when adopting post-control surveillance plans.
We attempted to test all humans and bovines in the 53 selected villages for infection, but approximately 30% of the population did not participate. Estimating the true population of each study village, and therefore the true infection survey participation percentage, is difficult. Due to residency requirements, government population registers in rural areas often include families that have moved to urban areas without registering such moves. Conversations with village leaders suggest almost all residents who spent most of their time in the village were captured by the demographic and household surveys. Participation in the infection surveys was lowest among people who spent time out of their village in the past year and young adults. Among those tested, neither young adults nor people who left their village had high infection prevalence. Nonetheless, it is possible that some infections were missed, leading some villages where infections were present only among non-participating residents to be misclassified. The number of villages with human or bovine infections may be greater than reported here.
The dramatic reduction of schistosomiasis in China and elsewhere has prompted consideration of the next phase of schistosomiasis control, motivating public health leaders to look beyond morbidity control toward the elimination of human schistosomiasis [1], [8], [21]. While this transition marks progress in controlling schistosomiasis, new challenges arise when approaching elimination. The reemergence of schistosomiasis as documented here highlights the transience of reductions in schistosomiasis in some areas. Before the introduction of praziquantel, schistosomiasis control focused on environmental modifications to reduce snail habitat and improve sanitation [2], [56]. The long-term interruption of schistosomiasis transmission in China and elsewhere will require the integration of praziquantel treatment and alterations to local environments to reduce their potential to sustain the parasite lifecycle [49], [57]. In addition, it will require a surveillance system that can detect the reemergence of infections with sufficient speed and accuracy to allow interventions to halt renewed transmission and prevent the further spread of infections.
|
10.1371/journal.pcbi.1003005 | Sparse Coding Can Predict Primary Visual Cortex Receptive Field Changes Induced by Abnormal Visual Input | Receptive fields acquired through unsupervised learning of sparse representations of natural scenes have similar properties to primary visual cortex (V1) simple cell receptive fields. However, what drives in vivo development of receptive fields remains controversial. The strongest evidence for the importance of sensory experience in visual development comes from receptive field changes in animals reared with abnormal visual input. However, most sparse coding accounts have considered only normal visual input and the development of monocular receptive fields. Here, we applied three sparse coding models to binocular receptive field development across six abnormal rearing conditions. In every condition, the changes in receptive field properties previously observed experimentally were matched to a similar and highly faithful degree by all the models, suggesting that early sensory development can indeed be understood in terms of an impetus towards sparsity. As previously predicted in the literature, we found that asymmetries in inter-ocular correlation across orientations lead to orientation-specific binocular receptive fields. Finally we used our models to design a novel stimulus that, if present during rearing, is predicted by the sparsity principle to lead robustly to radically abnormal receptive fields.
| The responses of neurons in the primary visual cortex (V1), a region of the brain involved in encoding visual input, are modified by the visual experience of the animal during development. For example, most neurons in animals reared viewing stripes of a particular orientation only respond to the orientation that the animal experienced. The responses of V1 cells in normal animals are similar to responses that simple optimisation algorithms can learn when trained on images. However, whether the similarity between these algorithms and V1 responses is merely coincidental has been unclear. Here, we used the results of a number of experiments where animals were reared with modified visual experience to test the explanatory power of three related optimisation algorithms. We did this by filtering the images for the algorithms in ways that mimicked the visual experience of the animals. This allowed us to show that the changes in V1 responses in experiment were consistent with the algorithms. This is evidence that the precepts of the algorithms, notably sparsity, can be used to understand the development of V1 responses. Further, we used our model to propose a novel rearing condition which we expect to have a dramatic effect on development.
| Simple cells in the mammalian primary visual cortex (V1) are among the cells in the brain that are best functionally characterised [1]–[3]. They have also been used as a key model system for studying the complex interplay of intrinsic and extrinsic factors, i.e., nature and nurture, in controlling development. For instance, there is ample evidence that receptive field structure exists prior to eye-opening [e.g. 4]–[6], being significantly present in dark-reared animals [7], [8]. Yet numerous studies, many taking advantage of the fact that simple cells are the earliest in the visual pathway to encode input from both eyes [9], have demonstrated that receptive field properties are modified by visual experience during development [e.g. 10]–[20].
Developing a general theory of sensory coding has been an important goal of computational neuroscience. One famously powerful idea, Barlow's efficient coding hypothesis, is that early sensory coding attempts to remove redundancy by representing input in informationally optimal ways [21]. Among other achievements, this hypothesis has provided compelling explanations for the characteristics of retinal receptive fields [22]. However, redundancy reduction may be only a first step in sensory processing [23]. For instance, V1 is many times overcomplete in its representation of input [24], a fact that, on the surface at least, increases rather than decreases the redundancy in the encoding of the input [25].
One possibility is that V1 is attempting to code visual input sparsely [26]. Many variants of sparse coding have been mooted [27]–[35], and, when tailored for natural scene input, almost ubiquitously lead to units with response properties similar to V1 simple cells. Other work has extended sparse coding models of V1 to complex cells [36], the dimension of time [37] and color [38], [39] [reviewed in 33]. Sparse learning schemes often trade off the amount of sparsity and the error in the encoding. The justification for sparse coding has ranged from the energetic grounds of being metabolically efficient [40], to the statistical grounds of exposing underlying latent structure in the input [24], [31].
The boldest claim of the sparse coding hypothesis is that it offers more than just an interpretation of simple cell receptive fields, but rather that it can account for the outcome (if not necessarily the time-course) of cortical plasticity. Showing this would offer a more stringent response to criticisms about the utility of these forms of unsupervised learning models for understanding visual development [20], [41], and also license applications of the same principles at more advanced stages of sensory processing. However, bar some notable exceptions [e.g. 35], [42], models based on precepts such as sparse coding have typically been applied to the development under normal input, for which the role of nurture can be questioned, rather than under abnormal input, for which it cannot. Furthermore, apart from notable exceptions such as Hoyer and Hyvärinen [39], the models have typically focused on monocular rather than binocular receptive fields, thus not addressing many of the most important experimental conditions.
Here we tested whether receptive field changes in six abnormal rearing conditions applied to cats (table 1) can be captured by binocular versions of sparse coding models of receptive field development. The cat was chosen as the model organism to match since all the conditions have been examined experimentally by several different groups, leading to broad agreement in the results. The more limited range of experiments that have been conducted in other species, notably macaques, have led to similar results, as in [43].
To ensure the outcomes were due to the general principle of sparsity, rather than the specifics of a particular algorithm, we used three different generative models for learning sparsity: product of experts [44], k-means clustering and independent component analysis [45]. We found that all three models qualitatively reproduced the receptive field changes observed in experiment in every rearing condition considered, and provided a good quantitative match in cases in which there was sufficiently ample sampling of receptive fields in the relevant experiments. This agreement provides evidence that receptive fields are indeed optimized during development in response to input statistics. Further, we used our models to design a novel rearing condition that we propose offers a strong test of the explanatory power of sparse coding. This involves presenting white noise with sparsity greater than that of natural scenes such that, even when augmented with natural input, it is still expected to lead to the development of highly localized receptive fields that are quite different from those of normal simple cells.
Overall, we suggest that examining abnormal rearing conditions will offer tests of functional accounts of development that are revealing and stringent, and look forward to the prospect of their application to higher visual areas and other sensory modalities.
We examined whether simple unsupervised learning models could capture the receptive field structures observed in abnormally reared animals. The models learn sparse responses which are conditionally independent given the input. We considered normal rearing, along with six abnormal rearing conditions (summarised in table 1).
All three unsupervised learning models gave qualitatively similar results across the rearing conditions. Figures shown in the main text, starting from the bottom rows in each column in Figure 1 which show sample receptive fields for each rearing condition, are for the results found using the product of experts model [44]. In Text S1/S2 we provide the same figures with the results for independent components analysis/k-means clustering. Where there are notable differences between the models we mention this is the main text.
To facilitate a direct comparison between models and experiment, in each of the subsequent sections we first provide a brief literature review of the relevant experimental work for that condition, and then present the results of our models. Although some conditions required specialised additional comparisons, changes were observed in every condition in receptive field binocularities (figure 2), the fraction of oriented receptive fields devoted to each eye (figure 3) and the joint orientation-binocularity distribution (figure 4). These figures are referred to in each section. In sum, all three models are able to match the changes in receptive field properties observed across all the conditions considered.
As expected from previous work in the monocular case [27], [28], [31], [32], the binocular receptive fields learned based on normal input were Gabor-like edge detectors (figure 1A). This property broadly survived the modified rearing conditions, up to some degradation and broadening. Given normal input, receptive fields were distributed over the full range of orientations (figure 4A) with primarily binocular responses (figure 2A). Note that the quantification of binocularity in some early experimental results is somewhat subjective, which complicates quantitative comparison. For example, most experimental groups classify monocular cells as ones which respond solely to one eye, which is difficult to define theoretically, since learned receptive fields are never entirely empty.
One salient feature of the orientation distribution is the over-representation of cardinal orientations. There is evidence that some degree of cardinal over-representation is present in normally reared animals [46] and in the visual environment [47], although it may be accentuated in our work due to the pixel representation of the training images (see later for further discussion and references).
An additional feature of note is the relationship observed in the normal rearing condition between orientation and binocularity. Li and Atick [48] examined 2nd-order correlations in visual input and predicted that vertically oriented receptive fields should be more monocular than horizontally oriented ones due to the asymmetry in inter-ocular correlations with horizontal disparity. This asymmetry in encoding can be seen in the normal case (figure 4A), with significantly more monocular receptive fields for vertical orientations. We are not aware of detailed experimental investigation of this phenomena (see Discussion).
One concern we examined for the case of normal input was the robustness of the models to training set size. This is particularly important since, for computational reasons, we trained our models with training examples, which is approximately a factor of four less than the number of degrees of freedom of the overcomplete models. By inspection, receptive fields appeared similar for different training set sizes. To test this more quantitatively, we examined the dependence on size of a key statistic of the receptive fields, namely orientation selectivity (figure 5). We found no dependence. Thus, the sparse constraints of the model result in a fit that is robust to training set size, even in the under-constrained regime.
Stripe rearing refers to the condition in which animals are raised with visual experience consisting primarily of a single orientation. This can be achieved by the use of cylindrical lenses, lenses painted with stripes, or rearing chambers with striped walls. Early electrophysiological studies on the effects of stripe rearing conflict, with some reports of a complete absence of receptive fields responsive to the unexperienced orientations [49], [50], while others found no effect on receptive field distribution [51], and Freeman and Pettigrew [52] found a more limited over-representation of the experienced orientation and reduced selectivity to the unexperienced orientations. Later experiments found significant over-representation of the experienced orientation [53], [54] and reduced orientation selectivity for unexperienced orientations. Unlike the other studies, Freeman and Pettigrew [52] and Blasdel et al. [53] reported a reduction in binocular responses; however Freeman and Pettigrew [52] attributed this to misalignment in the oriented lines between the two eyes. Blakemore [55] found that stripe reared animals have normal levels of binocularity.
More recently, optical imaging techniques have allowed simultaneous characterisation of large regions of V1. Using optical imaging, Sengpiel, Stawinski and Bonhoeffer [15] found a roughly 60% increase in the cortical area devoted to the over-represented orientation, with no change in the orientation selectivity between the experienced and unexperienced orientations. Tanaka et al. [56], [57] used exclusive goggle rearing with cylindrical lenses and found a much more dramatic 3–6 fold over-representation and increased selectivity (and reduced variance) of the exposed orientation. Tanaka and colleagues also noted that older animals exhibited reduced over-representation despite continued goggle rearing, and that increased dark exposure limited the effect of the goggle rearing.
There are several possible explanations for some of the differences between studies. The method of over-representation varied: some studies used stripe-cylinders, while others used goggles containing lines or strong cylindrical lenses. The age and duration of exposure also varied, and some early studies may have suffered from sampling biases. However, there is broad agreement between studies that stripe rearing leads to increased binocularity and significant increases in over-representation of the exposed orientation. Further, several studies found increased selectivity of the exposed orientation.
We modeled stripe-rearing by filtering the input using oriented Gaussian blurring designed to attenuate the power of off-axis orientations sharply. To maintain stability of the algorithm, 10% normal images were included in the training mixture (as in [42], see Discussion). The output of the models was consistent with the experimental observations. Receptive fields trained on striped input showed increased binocularity (figure 2B, 8% monocular cells in the stripe-reared condition compared with 17% monocular cells in normal condition, two-sided t-test). As in the experiments, there was a slight reduction in the number of orientation selective responses (figure 3), and the over-exposed orientation (in our case, horizontal) had sharper tuning curves (figure 6) with a smaller variance in their tuning. The size of these changes was dependent on the strength of the input filtering (data not shown), which is another possible explanation for the differing effect sizes seen between groups using different rearing techniques. These changes collectively meant that many receptive fields were less Gabor-like, although orientation selectivity was largely preserved. Since we could not find empirical studies employing methods such as reverse correlation against which to compare our results, it is difficult to determine how faithful this result is. However, the loss of structure of some receptive fields is at least qualitatively consistent with the experimental finding of reduced orientation selectivity in stripe rearing.
Orthogonal rearing is a binocular extension of stripe rearing in which the two eyes are exposed to orthogonal orientations. Hirsch and Spinelli [58], [59] found that orthogonally reared animals had reduced binocularity and almost exclusively monocular responses for cells with well-defined orientations. They also found an almost perfect correlation between receptive field ocular and orientation preferences along with an overall reduction in the fraction of oriented responses. Subsequent groups found similar results although the quantitative changes reported varied, possibly due again to differences in the type of filtering and the strength of the manipulation between experiments. Freeman and Pettigrew [52] noticed reduced binocularity (25–35% binocular cells for orthogonally reared animals compared with 85% for normally reared animals) and a strong correlation between orientation and ocular preference along with broader orientation tuning away from the over-exposed orientations. Blakemore [55] also noted reduced binocularity (53% monocular responses).
Leventhal and Hirsch [60] observed a difference between horizontal and vertical orthogonal rearing and orthogonal rearing with oblique angles. In the cardinal case, they found a strong correlation between ocular and orientation preference. With oblique angles they found a continued dominance of horizontal and vertical orientations, but little response to the non-exposed non-cardinal orientation. In all cases they noted a reduction in binocularity. Stryker et al. [54] also found a reduction in the number of oriented responses (50% of cells were not responsive or not selective to orientation) and strong correlation between eye and orientation preference. They also observed over-representations of approximately two-fold for the exposed orientations and almost no binocular cells. More recently, albeit as yet only in abstract form, Tani and Tanaka [61] confirmed the over-representation of the exposed orientations using optical imaging.
As in the stripe reared case, we modelled orthogonal rearing by oriented Gaussian blurring. However, in this condition, the left eye viewed horizontally filtered images while the right eye viewed vertically filtered images. This led to similar results as in the experiments. When trained on this orthogonally filtered input, model receptive fields were significantly more monocular (figure 2C, 31% monocular compared with 17% in the normal case, , two-sided paired ttest), although this effect was not as pronounced as reported experimentally. Responses showed a strong correlation between ocular and orientation preference (figure 4C). Additionally, there was a reduced fraction of oriented responses compared with the normal case (figure 3, 37% compared with 63% in the normal case, ). The results were similar when oblique rearing orientations are considered, although in this case there was also a smaller, cardinal over-representation effect (data not shown). This cardinal over-representation is presumably driven by the same causes as in normal case (discussed further below): cardinal over-representation in the input and the square pixel representation.
Monocular deprivation, in which one eye is deprived of visual input, is perhaps the best-studied manipulation. There is substantial variation in deprivation length and daily visual exposure between different studies. We only considered results based on experiments with no recovery period with normal visual input. We examine the recovery of binocular fields later in the section on partial monocular rearing.
Early work by Wiesel and Hubel [62], [63] using electrophysiology found almost no response to the deprived eye (1/84 cells responded). Similarly, Hubel and Wiesel [64] observed only 7% of cells responsive to the deprived eye after 3 months of deprivation (all cells were classified as having ocular dominance values of 6 or 7 on a scale of 0–7). Blakemore and Van Sluyters [65] also demonstrated almost complete domination of V1 by the deprived eye, with normal levels of orientation selectivity in the non-deprived eye. Using autoradiography, Shatz, Lindstrom and Wiesel [12], Stryker [14] and Shatz and Stryker [13] found shrinkage of the deprived eye's territory with only 22–25% cortical area labelled by the deprived eye. Olson and Freeman [66] considered the effects of shorter periods of deprivation, finding pronounced decreases in binocularity after just 2.5 days of deprivation, and almost total loss of responsiveness to the deprived eye after longer periods, compared with 80% binocular responses in the normal animals. Similarly, Peck and Blakemore [67] found that, with just 20 hours of monocular deprivation, all oriented cells had ocular dominances in the range 4–7. Only a small number of unoriented responses remained exclusive to the deprived eye. Schechter and Murphy [68] noted 86% of cells responded exclusively to the open eye, 3% to the deprived eye, 3% had binocular responses and 7% were unresponsive. Kratz and Spear [69] observed a reduced number of orientation selective cells (65% compared with 85% in normal) and reduced direction selectivity (70% compared with 90% in normal). Blasdel and Pettigrew [70] found that 3 weeks of molecular deprivation led to most cells having ocular dominances of 7, with a small number reported as being 5 and 6. Olson and Freeman [71] also noted that only 3% of cells responded to the deprived eye (87% in normal) after 10 days of monocular deprivation. Singer et al. [72] found that the majority of cells were responsive only to the open eye.
Some variations have added insight about the changes occurring during monocular deprivation. Blakemore [55] and Wilson, Webb and Sherman [73] demonstrated that there was little difference between monocular deprivation with the nictitating membrane or the full eye-lids, showing that the loss of spatial patterns, rather than the change in luminance, is the critical component of deprivation. Blakemore and Hillman [74] showed that the open eye dominated whether optically stimulated, or driven electrically. Olson and Freeman [75] interspersed dark-reared intervals during the deprivation and continued to find almost no response to the deprived eye. Tumosa, Tieman and Hirsch [76] used behavioral assays to show that the animals had no functional visual acuity in the deprived eye. Mitchell [77] demonstrated that recovery was improved when the non-deprived eye was occluded during the recovery period. More recent experiments have used optical imaging, and found only 14–18% of cortex responded to the occluded eye [16] and confirmed that little functional visual acuity remains in the deprived eye [78].
We used the observation from Wilson, Webb and Sherman [73] that it is the spatial pattern of the input associated with the deprived eye that matters rather than the overall power to realize a stringent test of the model. We simulated this by using an extremely low-pass boxcar filter on the right eye's input, so that almost all contrast was destroyed. The model reproduced the experimental findings, producing primarily cells with ocular dominance values of 2, and none greater than 4 (figure 2D), indicating no cell responded more strongly to the deprived eye than the open eye. No oriented response was assigned to the deprived eye (figure 3).
In alternating monocular rearing, animals are monocularly deprived, but which eye is deprived is alternated regularly (every few hours of visual experience). This removes all inter-ocular correlations, while not favouring the development of either eye. Hubel and Wiesel [10] first performed this experiment and found that 91% of the resulting cells had monocular responses, evenly distributed between the two eyes. Behaviourally, the animals appeared to have normal spatial acuity in each eye. Blake and Hirsch [79] also measured normal acuity in each eye but observed defects in stereopsis. They also noted an almost complete absence of binocular responses even after animals were reared with normal input for a year after the critical period. Blakemore [55] found alternating monocular rearing resulted in 55% of neurons responding monocularly, similar to a strabismic animal. Blasdel and Pettigrew [80] also found reduced binocularity ( binocularity) except when they used a mechanised device to reduce the alternation interval to less than 1 second. They also observed a low correlation in orientation tuning between the two eyes ( versus in normal animals). Tumosa, Tieman and Hirsch [76] used behavioural assays and found normal visual acuity in each eye, and equal cortical coverage devoted to each eye [81]. In all of these experiments, alternating monocular rearing had a similar effect to strabismic rearing: reduced binocularity while retaining an equal number of neurons devoted to each eye.
We simulated this condition in a similar way to monocular rearing, with the blind eye having its input low-pass filtered so that little contrast remained. The difference from monocular rearing was that the eye that was blind was alternated for each patch. This resulted in quantitative agreement with experiment. The PoE model predicted (figure 2E) 89% monocular responses (compared with 17% in the normal case, , two-sided paired t-test) and a symmetrical ocular dominance distribution. There was also a reduced fraction of oriented responses (figure 3).
Monocular rearing leads to almost complete loss of function in the deprived eye. There has thus been substantial interest in the question as to what features of the input are necessary for preventing this. Partial-monocular rearing, in which animals are exposed to a small fraction of binocular experience, allows the amount of normal input needed for the maintenance of responses to both eyes to be determined.
Olson and Freeman [75] monocularly deprived kittens for 4 hours while providing 14–20 hours of binocular experience per day. They found this limited deprivation had little effect. Similarly, Kind et al. [16] examined the effect of monocular deprivation for 10 days (at 5 weeks old) followed by binocular exposure for 14 days. Again, these kittens had nearly normal visual development, although non-aligned binocular input (i.e. strabismic) led to only 34% coverage for the deprived eye, demonstrating that correlated visual input may be important for recovery. Conversely, Malach and Van Sluyters [82] found that strabismic binocular input did lead to recovery of binocular responses, but this may be because the animals were dark-reared for 18 hours per day, a manipulation that is known have a protective effect [56].
Follow-up experiments interleaved binocular experience with monocular deprivation. Mitchell et al. [78] found that even 0.5 hours of binocular experience with 6.5 hours monocular deprivation preserved moderate spatial acuity in the deprived eye and 2 hours binocular experience with 5 hours monocular experience resulted in normal acuity. Again, they found the inter-ocular correlations were vital, as binocular experience in which artificial strabismus had been induced by prisms resulted in poor recovery of visual acuity. Mitchell et al. [83] found that splitting the binocular experience into discontiguous blocks impeded recovery.
Later experiments explored the neural basis of these changes. Schwarzkopf et al. [17] found normal cortical coverage of both eyes for animals with more than 30 minutes daily binocular experience, whether this was matched with 3.5 or 7 hours of monocular deprivation. Vorobyov et al. [18] examined interocular phase selectivity (a measure of disparity tuning) and found reduced phase selectivity in the partial monocularly reared animals. Mitchell et al. [19] demonstrated that, while partial monocularly reared animals recovered normal levels of spatial acuity in each eye, most had severe deficits in binocular depth perception (unlike normal animals, their depth estimates did not improve when they were allowed to use both eyes). Mitchell et al. [84] showed that animals developed normal spatial acuity provided they received at least 30% binocular experience (regardless of total exposure length) [85]. These results broadly agree that a significant response to the deprived eye recovers with as little as binocular input, and that normal levels of spatial acuity occur with 30% binocular input. However, significant deficits in binocular integration remain even with 30% binocular experience.
We simulated this condition by including a fraction of normal visual input along with the same boxcar filtered input used for monocular rearing. This resulted in significant recovery of deprived eye responses with just 10% binocular input (figure 7B), and recovery of equal representation of each eye with 40% binocular input (figure 7E) with the PoE model. However, as in the experiments, recovered responses tended to be monocular, with fewer (67% compared with 83% in normal, two-sided paired ttest) strongly binocular responses. The deprived eye rapidly recovered responses to the full range of spatial frequencies (figure 8) which corresponds well with the behavioural experiments.
This experimental condition was the only one which showed qualitative differences between the different unsupervised learning models. With the ICA model (Text S1) the recovery from monocular deprivation is much weaker than with the PoE (figure 6) or kmeans clustering (Text S2). Unlike the other models the ICA model is not overcomplete: it has half as many receptive fields to allocate, and may therefore be more susceptible to allocating receptive fields only to the majority input statistics. As we discuss later, there is much evidence that biological V1 is substantially overcomplete.
This recovery appears counterintuitive as it seems the normal input is exerting a disproportionate effect on receptive field development. One explanation may be that, since the sparse coding model strongly penalises representations which are insufficiently sparse, only a small amount of binocular experience is necessary before a significant number of receptive fields are allocated to the deprived eye. As we discuss later, this result suggests that a preferential mechanism for normal input may not be required to explain the recovery observed in partial monocular rearing: rather, it may be a natural consequence of development prior to patterned input.
Strabismus is of both clinical interest [86], as a condition which affects a significant fraction of the population, and theoretical interest, as it lowers inter-ocular correlations. The effects of convergent or divergent stabismus are similar [87], [88]. Hubel and Wiesel [10] used electrophysiology to show that kittens raised with divergent strabismus develop a majority of neurons which are responsive to only a single eye. This finding was reinforced by Shatz, Lindstrom and Wiesel [12] who used histology to demonstrate that strabismic kittens have more bimodal distribution of ocular dominance columns. Chino et al. [89], using animals with strabismus greater than 10°, found no neuron in ocular dominance category 4 (strongly binocular). Unlike other investigators, they also found reduced spatial acuity in the deviating eye and a reduced contrast response. Yinon and Auerbach [87] found that approximately 70% of cells were monocular, and noted an increased number of unresponsive neurons. Similarly, Blakemore [55] measured 76% monocular responses. In a follow-up experiment they observed no strong binocular response when neurons were stimulated optically, and only a small fraction with direct electrophysiological stimulation [74]. Van Sluyters and Levitt [90] used prisms rather than surgical manipulation, which allowed them to create symmetric strabismus. In both divergent and convergent conditions, they found a loss of binocularity with the majority of neurons being assigned ocular dominance categories 1, 2, 6 or 7.
Later results confirmed these findings. Levitt and Van Sluyters [91] found that kittens raised with strabismus during the critical period (2–4 weeks) had cells that were primarily monocular. Grunau [92] measured 80% binocular responses in normal animals and 26% in strabismic. Berman and Murphy [93] also noted a loss of binocularity ( binocular simple cells, compared with 65% for controls), and also observed increased receptive field sizes. Kalil, Spear and Langsetmo [94] found only 7% binocular cells in strabismus. They also reported that animals with divergent strabismus had equal representation of each eye while convergent strabismus resulted in a slight reduction in the representation of the periphery of the deviating eye. Eschweiler and Rauschecker [95] and Schmidt, Singer and Galuske [96] both confirmed that the majority of neurons were monocular. Schmidt Singer and Galuske [96] also found similar orientation preference map characteristics between normal and strabismus (and between maps measured in the deviant and normal eye). Vorobyov et al. [18] also noted a significant decrease in binocular responses compared with control in strabismic animals.
In sum, there is substantial agreement about the effects of strabismus. With 10–20° deviance, whether divergent or convergent, animals develop 80% monocular responses, no strongly binocular response at all, and a reduced number of responsive cells overall. There are mixed reports regarding preferences for the non-deviating eye.
We simulated the effect of strabismus in the models by choosing visual scene patches independently in each eye (focal points were still identical in each). This disrupts inter-ocular correlations, as each eye views different parts of the scene, and led to similar results to those found in the experiments. With the PoE model, only 2% of cells were in ocular dominance category 4, with the majority being in categories 2 and 6 (figure 2G). Additionally, there was a significant reduction in the total number of orientation selective cells responding to either eye (figure 3). The full range of orientation preferences continued to be expressed (figure 4G).
We have shown above that sparse coding provides an explanation for the RFs that result from several different abnormal rearing conditions. We therefore considered whether there was a novel experiment which could directly address the importance of sparsity in RF envelopment. To do this we exploited that fact that advances in experimental technology have made it possible to rear animals with visual experience which is almost entirely computer-driven (e.g. [97]). This provides nearly unconstrained scope for modification of visual experience after eye-opening.
We sought a stimulus for which sparsity was central, and that we would predict would lead to markedly different receptive fields from normal animals when used as input. It was important that the difference would persist even if some more naturalistic input was additionally provided, for instance from retinal waves present before eye-opening, or from small amounts of natural visual input that cannot be completely controlled in a practical experiment [98]. We also required the stimulus to be statistically stationary in space, so that it would not be necessary to track eye position in presenting the stimulus.
We constructed stimuli by independently sampling each pixel of the patch from a sparse distribution (student-t with 2 degrees of freedom). This is intentionally close to the distribution of coefficients (rather than pixels) in natural scenes, so as to push the model towards learning the Cartesian basis. As a control, we also considered stimuli with a uniform or Gaussian distribution of coefficients. We examined the receptive fields predicted by the model when trained on mixtures of natural scenes and these noise stimuli. All input data were normalized to have the same mean and variance before combination.
We found that sparse noise provoked a disproportionately strong response in the receptive field development of the models (PoE results in figure 9). Even when trained on a mixture of 50% natural scenes, sparse noise resulted in strongly localized and distinctive receptive fields (figure 10). This effect was particular to sparse noise, as Gaussian or uniform noise had substantially less influence on receptive field development (figure 10).
In this one instance, the k-means clustering model results (text S1) contained some deviation from the results of the other two models. In particular, the development of orientation selectivity was impeded more strongly than in the other models by Gaussian and flat noise, and the mixtures of sparse noise and natural scenes developed a high proportion of low-frequency spatial fields. The exact reasons for these deviations are unclear. However, even with k-means clustering the modification to receptive fields trained with sparse input is large, and robust to the inclusion into the training set of a substantial fraction of natural scene input.
We have shown that sparse coding models can simulate the structure of the V1 simple cell receptive fields that arise when animals are reared with normal and abnormal visual input. Receptive field structure exists prior to eye-opening, and there is ample evidence that important aspects of development are driven by intrinsic factors. However, that dramatic changes in receptive field properties occur that depend on the nature of the input suggests that substantial plasticity remains after eye opening. The account of these changes in our model provides evidence of a causal link between receptive field structure and optimal representations of visual input. This directly answers questions [99] that have been directed at models of purely normal development [26], [27], [30], [32], [37], [100] about the necessity or sufficiency of an impetus towards sparsity. We suggest that sparse coding provides a unifying framework for modelling receptive field changes under a wide variety of rearing conditions.
Understanding the mechanisms by which visual responsivity can be harmed and indeed potentially cured by aberrant or benificent input has important implications. Take, for instance, the case of partial monocular rearing [11], [16]–[19]. Monocularly deprived animals do not develop substantial V1 responses to the occluded eye [13], [14], [63], [64]. Recent experiments have demonstrated the recovery of near-normal visual acuity in animals allowed only a small fraction (1/7) of binocular experience. If this result depended on some intrinsic mechanism for detecting and reacting to this small amount of experience, then it might not generalize. Contrary to this, our findings suggest the enticing possibility that improvements may arise as a natural consequence of developmental optimisation of V1 coding.
Our results regarding the development of receptive fields in abnormal rearing conditions are in agreement with previous results which have examined other modalities and a more limited range of visual rearing conditions. Hsu and Dayan [42] showed that a monocular version of the products of experts model matched the over-representation observed in stripe rearing when trained on stripe filtered input, and a similar result has been found for other algorithms based on sparseness [35]. Saxe et al. [35] also demonstrated that unsupervised learning algorithms could match receptive field changes in abnormal rearing conditions across a range of different sensory modalities.
We used three unsupervised learning methods – independent component analysis [33], product of experts [101] and k-means clustering [102] to acquire sparse codes, thus ensuring that our results did not depend on the specifics of one particular algorithm. As in Saxe et al. [35], we found that all these algorithms learned qualitatively similar sparse codes. One advantage of k-means clustering and the product of experts is that they can readily learn over-complete codes. This is important as V1 is indeed significantly over-complete [23], [28], [42]. Due to the limitations of the experimental data, it is difficult to make any strong claims about which particular unsupervised learning approach provides a better fit. Although the mechanics of these algorithms are not biologically realistic, similar sparse learning algorithms have been implemented in neurobiologically realistic terms [103], [104].
We interpret our results, and those of others based on normal input, as placing the focus on the nature of the efficiency afforded by sparsity [21], [26], [105]–[107]. One notion is that sparse codes may represent a trade-off between the metabolic costs associated with neurons that are firing versus those that are quiescent (and maintaining their membrane potentials; [108]). Of more widespread note is the ability of sparse codes to capture the sort of latent statistical structure in input that can then underpin visual comprehension [105], [109], [110]. The idea is that sensory input arises from the superposition of causes that themselves occur only sparsely. These causes are what it is important to determine. Then, finding a sparse representation of the input in a computationally suitable context (formally, in the recognition component of a pair of recognition and generative models; [110]) can unearth those causes.
The agreement between the experimental results and the output of sparse coding models trained on visual input is perhaps surprising. V1 consists of much more than the feature detectors we have modelled it as here. Real V1 neurons incorporate temporal integration [3], bottom-up and top-down attentional modulation [20], [111]–[113], lateral connections between columns [114]–[116], feature maps [117], [118], and significant feedback from other cortical areas [119]. Additionally, our model does not include the influence of intrinsic activity which is known to be necessary for normal receptive field development [120], wiring constraints [121], and hemispheric asymmetries [16], [79]. These mechanisms may explain the presence of structured receptive fields in dark-reared animals, particularly as spontaneous retinal waves have statistical similarities to natural visual input [122]. However, the success of the model is evidence that, despite many mechanistic differences between the model and V1, sparse codes are good predictors of receptive field changes during the critical period.
Our models favor nurture at the expense of nature. One hint in our results that this favoritism may be too extreme comes from the critical importance of the 10% binocular input in the partial monocular rearing case (and thus the 10% normal input mixed in the stripe and orthogonal rearing conditions). From a technical viewpoint, the primary effect of the normal input is to avoid the collapse of the principal component step which results in the almost complete loss of oriented responses [42], figure 2b. The choice of 10% is somewhat flexible, and Hsu and Dayan [42] found that 75–95% stripe reared input resulted in strong over-representation without a collapse in oriented responses. It could be argued that the 10% is a stand-in for the ineluctable effects of the neural structures established prior to eye-opening, although further work would be required to establish this claim. In the models, the effect of normal input on stripe rearing appears to be more proportionate than in blind rearing; we are not aware of experimental tests of partial stripe rearing for comparison.
Trained with normal input, the models over-represented cardinal orientations. There is evidence that cardinal axes are indeed slightly over-represented in real animals [46], [123]–[127], presumably due to a prevalence of cardinal edges in natural scenes [125], [128], [129]. The degree of over-representation was in reasonably good agreement with a recent study of natural scenes [47]. However, the over-representation in the model may be accentuated because biological retinas are not arranged on a square lattice, unlike most digital representations of images. Such differences in representation are known to exert a small effect on the results of sparse coding models [130].
Modeling binocularity allowed us to observe an effect on coding due to the asymmetry of inter-ocular correlations in visual input that was predicted by Li and Atick [48]. They showed that, because inter-ocular correlation decays more rapidly with horizontal edges than vertical edges, a redundancy reducing code should have an increased number of monocular receptive fields for vertical edges and an increased number of binocular receptive fields for horizontal edges. In our models, despite higher order interactions not considered by Li and Atick, this effect is observed. Hoyer and Hyvärinen [39] have previously examined binocular encoding and reported that learned receptive fields had similar disparity preferences to experiment, but did not report on any relationships between orientation and ocularity. To our knowledge this asymmetry has not been observed in experiment. It is possible that supervised constraints on depth estimation [131] explain why this has not been observed in experiment.
We modelled the six conditions that have been the subject of the most intense investigation. Various other experimental manipulations of visual input have been performed in cats including rearing with random spots [132], exposure to constant speed and direction of motion [133]–[135], astigmatism [136] and opposite rotations of visual input in each eye [137]. We have not attempted to model these results here, either because the studies have not been replicated by other groups or because the effects involved temporal manipulation. In order to constrain the problem size, our models did not include the temporal dimension in receptive field responses, although previous work indicates that this is unlikely to change the results dramatically [37].
However, we did use the models to design a rearing regime that might provide a novel and strong test of the predictive power of sparse coding. According to our models, animals reared with exposure primarily to sparse white noise (similar to that used as a test of the sparse coding algorithm in Olshausen and Field [27]), should develop strongly localized receptive fields. This should be discernable with electrophysiological or optical imaging of receptive fields. Our models predict that the developed receptive fields will be small and mostly non-oriented. This effect is specific to noise with a sparsity near that of natural scenes, and Gaussian noise or distributions with a sparsity much greater than that of natural scenes (such as the spot stimuli examined in Ohshiro, Hussain and Weliky [97]) are predicted to have much less influence, particularly if the animal also receives some naturalistic input [98]. An experimental test of this prediction would provide evidence that sparse coding is a key driver of early receptive field development, or alternatively provide insights into the limits of plasticity during early visual development.
Our models consider only a single cortical area. It would be interesting to look systematically at the effects of the abnormal input statistics on the responses of neurons in higher cortical areas and, concomitantly, on the receptive fields and responses of units in multi-layer, hierarchical [138], [139] unsupervised learning models of those higher areas [35], [140]–[146].
Unsupervised learning can only take us so far in understanding brain function. At some point, brains have goals, seek rewards and avoid punishments. However, transforming high-dimensional input into representations that are more useful is an essential part of artificial forms of machine learning [147], and has offered a critical and realisable metaphor for understanding representations in the brain and the way that they are malleable to changes in the input.
We acquired a training set of naturalistic binocular image patches from eighteen stereo images of high-quality, binocular images of natural scenes (from http://home.comcast.net/~toeppen/, this image library was the same as used in Hoyer and Hyvärinen [39]). Each image was photographed using a binocular camera with lenses spaced approximately a pair of human eye's distance apart at varying focal distances. As in Hoyer and Hyvärinen [39], 5 focal points were chosen randomly in each image, the two stereo images were aligned at the focal point and then stereo image patches were acquired in a 300×300 pixel square around the focal point. This has the effect of approximating the view of an observer focussing at 5 different points in the scene. Each patch was 2×25×25 pixels. These patches naturally contain varying degrees of disparity. All images were first converted from color to greyscale. A total of 100,000 training patches were created for each rearing condition.
To model modified rearing conditions, the training input was filtered to match the visual experience of the animals. For stripe and orthogonal rearing the off-axis spatial frequencies were attenuated by using an oriented Gaussian filter [as in 42]. For monocular rearing and alternating monocular rearing the occluded eye's images were convolved with a square kernel with a length of 150 pixels, which removed all but extremely low spatial frequencies. To simulate strabismus, the focal points were chosen independently for each eye. In keeping with previous work, 10% of the training input was unfiltered in the stripe and orthogonal rearing conditions. Hsu and Dayan [42] found that retaining 10% normal input gave a better match with experiment because it reduced the “collapse” of receptive field structure that occurred in the absence of any normal input.
We used three different models for learning sparse codes: product of experts, k-means (which is also known as k-nearest neighbour) and independent component analysis. In all cases, the training data was whitened and dimension reduced using principal component analysis. We retained the first 150 principal components.
We used FastICA [148] to learn independent components and the built-in MATLAB ‘kmeans’ function to learn k-means clustering. Since these algorithms are well-known we do not describe them further here.
The product of experts model [44], [149] models the input distribution as a product of Student-t distributions. Representing each input patch as a column vector , and the ensemble average over all training examples as , the probability of input is modelled (with neurons) as:(1)(2)The parameters of this model (encapsulated in the term ) describe the receptive field and the sparseness of each neuron . The normalisation constant is dependent on . The model is trained by maximising the log-likelihood of the data with respect to the model parameters . When the number of neurons is equal to the dimensionality of the input there is a closed-form solution for and the model performs independent component analysis [44]. However, when over-completeness is introduced there is no general closed-form solution for the normalisation constant . Contrastive divergence [101], which performs gradient descent on a cost function that is within a small constant of the log-likelihood function, was used to fit the model with a learning rate of and a batch size of . We used an over-completeness factor of , as in [42].
Binocularity was quantified using the index described in Hoyer and Hyvärinen [39]:(3)where and refer to the portions of the receptive field corresponding to each eye and is the norm. For comparison with experiment we binned the values of into 7 bins with boundaries at [as in 150]. Bins 1 or 7 correspond to highly monocular responses while values in the middle correspond to binocular responses.
The orientation and spatial frequency preferences for each receptive field was calculated as in Hyvärinen, Hurri and Hoyer [33, ch. 6]. Each eye was treated separately. The response of the receptive field to a quadrature sinusoidal grating over a range of spatial frequencies and orientations was recorded. We examined spatial frequencies between and 0.5 cycles/pixels with spacing of 0.02 cycles/pixels and all orientations with a spacing of 1°. This provided an orientation response curve from which the circular variance [151], which is a measure of the orientation selectivity, could be determined. Circular variance is defined as where:(4) is the response of the neuron to stimuli of orientation . When plotting distributions of orientation preference we only included receptive fields which had a circular variance , as receptive fields with low circular variance cannot be reliably assigned an orientation.
|
10.1371/journal.pntd.0003192 | Physician Survey to Determine How Dengue Is Diagnosed, Treated and Reported in Puerto Rico | Dengue is a major cause of morbidity in Puerto Rico and is well-known to its physicians. Early case identification and timely initiation of treatment for patients with severe dengue can reduce medical complications and mortality. To determine clinical management and reporting practices, and assess knowledge of dengue and its management, a survey was sent to 2,512 physicians with a medical license in Puerto Rico. Of the 2,313 physicians who received the survey, 817 (35%) completed the questionnaire. Of the respondents, 708 were currently practicing medicine; 138 were board certified (Group 1), 282 were board eligible (Group 2), and 288 had not finished residency (Group 3). Although respondents clinically diagnosed, on average, 12 cases of dengue in the preceding three months, 31% did not report any suspected cases to public health officials while about half (56%) reported all cases. Overall, 29% of respondents correctly identified early signs of shock and 48% identified severe abdominal pain and persistent vomiting as warning signs for severe dengue with the proportion of correct respondents highest in Group 1. Reportedly about sixty percent (57%) appropriately never give corticosteroids or prophylactic platelet transfusions to dengue patients. One third (30%) of respondents correctly identified administration of intravenous colloid solution as the best treatment option for dengue patients with refractory shock and elevated hematocrit after an initial trial of intravenous crystalloids, and nearly one half (46%) correctly identified administration of a blood transfusion as the best option for dengue patients with refractory shock and decreased hematocrit after a trial of intravenous crystalloids. Even though dengue has been endemic in Puerto Rico for nearly 4 decades, knowledge of dengue management is still limited, compliance with WHO treatment guidelines is suboptimal, and underreporting is significant. These findings were used to design a post graduate training course to improve the clinical management of dengue.
| Dengue is a major cause of morbidity in Puerto Rico and is well-known to its physicians. Early case identification and timely initiation of treatment for patients with severe dengue can reduce medical complications and mortality. We conducted a survey among physicians who practice in Puerto Rico to determine clinical management and reporting practices and assess knowledge of dengue and its management. We found that although respondents clinically diagnosed, on average, 12 cases of dengue in the preceding three months, one third did not report any suspected cases to public health officials while about half reported all cases. We found that knowledge of dengue management was limited and compliance with WHO treatment guidelines was not optimal. As other dengue endemic countries have reported similar findings, a sustained continuing medical education training initiative may be necessary to improve case detection and clinical management even in countries where the disease is common. Our findings were used to design a postgraduate training course to improve the clinical management of dengue.
| Dengue is a mosquito-borne disease caused by any one of four dengue virus (DENV) types -1, -2, -3, and -4. Each DENV is capable of causing the full spectrum of disease from an asymptomatic infection to severe, life-threatening illness including dengue hemorrhagic fever (DHF) and dengue shock syndrome (DSS) [1]. Dengue is a major public health problem throughout the tropics and subtropics worldwide. There is currently no vaccine available to prevent dengue and vector control measures to prevent DENV transmission have not been sustainable or effective [2], [3]. Once a person has dengue, there is no licensed antiviral medication to treat or prevent severe manifestations of the disease. However, implementation of other secondary prevention measures including timely identification of dengue cases and initiation of intensive supportive treatment can reduce case fatality rates from 10% to less than 1% among severe cases [4]–[8].
The incidence and severity of dengue has been steadily increasing over the last three decades throughout much of South and Central America, Mexico, and the Caribbean including Puerto Rico [9]. Even though dengue has been endemic in Puerto Rico since the late 1960s [10], how physicians identify, diagnose, and report patients with suspected dengue is not well known. Similarly, even though the severity of dengue has increased with every subsequent dengue outbreak in Puerto Rico since 1994 [11], little is known about clinical management practices for dengue on the island. However evidence from fatal dengue case review suggests that treatment practices in Puerto Rico may differ from the World Health Organization (WHO) guidelines [12], [13] even though efforts to educate physicians concerning dengue management based on the most current guidelines had been performed from the late 1980s to the early 1990s and then intermittently with each subsequent outbreak [14]. To better understand diagnostic, treatment and reporting practices, we conducted a survey among physicians practicing in Puerto Rico in 2007–2008. Findings were used to develop a post graduate training course on the clinical management of dengue to minimize dengue morbidity and mortality, and to improve reporting of suspected dengue cases in Puerto Rico so that we can better understand the true burden of disease.
In 2007, there were 8,051 physicians residing in Puerto Rico who had a Drug Enforcement Administration (DEA) number and active license to practice medicine in Puerto Rico, according to the Puerto Rico Department of Health (PRDH) (Figure 1). Physicians who were unlikely to diagnose and treat patients with dengue were excluded from the list of 8,051 physicians, including surgeons, pathologists, radiologists, allergists, dermatologists, endocrinologists, geneticists, nephrologists, neurologists, ophthalmologists, otolaryngologists, oncologists, psychiatrists, rheumatologists, and sports medicine physicians. The remaining 5,997 physicians consisted of 5,635 generalists involved in primary and emergency care of dengue patients (e.g., general practitioners, family practitioners, pediatricians, emergency department physicians, obstetricians and gynecologists, and internal medicine physicians) and 362 specialists, most notably cardiologist and pulmonologists who are most likely to work in intensive care units in Puerto Rico, and intensive care physicians.
To determine physicians' knowledge of how to diagnose and treat dengue according to the 1997 WHO guidelines and to assess their treatment and reporting practices, we determined that 1,068 generalists were needed to be 95% confident of being within ±3% of the assumed population proportion of 50%, used because this value gives the most conservative sample size estimates. We expected, based on the literature, that there would be a 50% non-participation rate, and therefore a simple random sample of 2,150 generalists was selected from the 5635 primary care and emergency medicine physicians. In addition, all 362 specialists were invited to participate. This survey underwent institutional review at the Centers for Disease Control and Prevention (CDC) and was determined to be public health practice and not research; as such, Institutional Review Board approval was not required.
We sent a 37-item questionnaire, a personalized cover letter explaining the purpose of the survey, and a pre-addressed, prepaid return envelope to the 2,512 physicians in October of 2007 (Figure 1). One hundred and ninety-nine questionnaires were returned because of an inaccurate mailing address. Three weeks after the initial mailing, a reminder postcard was sent. A second copy of the questionnaire and a pre-addressed, prepaid return envelope was sent in mid-January 2008, and a second reminder postcard was sent in early February 2008. No incentives were given for participating.
The questionnaire, which was pilot tested among a small, diverse group of physicians practicing in Puerto Rico, asked for demographic information including age, sex, and training history (e.g., location of medical school and year of graduation, location of residency training, and whether or not they were board certified or finished a residency program). Further, physicians were asked whether or not they were currently practicing medicine, the location and type of healthcare facility of their current practice, and the average number of suspected dengue patients seen per week. Respondents who reported that they were no longer practicing medicine or did not see patients with dengue were excluded from the final analysis. Respondents who were currently practicing were asked questions about the clinical and laboratory diagnosis of dengue, and how frequently they report suspected dengue cases. Practicing respondents were also asked about their hospital referral criteria for patients suspected of having dengue, and to identify warning signs for severe dengue and early signs of shock. The questionnaire also asked about specific treatment practices including use of corticosteroids, prophylactic platelet transfusions, and intravenous immune globulin for patients suspected of having dengue, and knowledge of when to use intravenous colloid solutions and blood products in dengue patients with refractory shock after an initial trial of an intravenous crystalloid. A copy of the questionnaire is available upon request.
A pre-specified analysis of management practices and knowledge of dengue was conducted by respondent level of training; groups included board-certified physicians (Group 1), residency trained physicians without board-certification (Group 2), and physicians who did not complete residency training and were therefore not qualified to take the Board examination (Group 3). Throughout, point estimates, 95% confidence intervals (CI), and statistical tests were computed accounting for the sampling design (stratified, simple random sample) and incorporating a finite population correction [15]. Comparisons among groups were made using a chi-squared test for categorical data accounting for the survey design and using the Rao-Scott adjustment [16]. Simultaneous CIs for the differences among group proportions were adjusted using the Bonferroni adjustment. Group means were compared using the likelihood ratio test with the Rao-Scott adjustment and accounting for the sampling design [16].
In order to assess the internal and external validity of the survey's results, we evaluated the questionnaires for completeness and recorded those questionnaires that were discarded and the reasons for doing so. Response proportions were compared to available population proportions for sex and physician location (San Juan metro area or Ponce), and these were assessed using a chi-squared goodness-of-fit test accounting for the sampling design [16]; other demographic information was unavailable on a population level to provide reference for evaluation. All data analyses were conducted using STATA version 10 and the survey package in R [15], [17].
In the tables and text we report both the raw counts of the numbers of respondents in the stated categories and the relevant total numbers of respondents. We also report estimates of population proportions and means based on survey data analysis, a statistical method which incorporates weighting, so these estimates may not match crude proportions calculated from the values reported. All proportions and means are such survey-based population estimates. To ease presentation, we do not include CIs in the text below when these can be found in the tables.
Of the total 2,313 physicians who received the survey, 817 (35.3%) completed the questionnaire (Figure 1). Of the 817 respondents, 109 were excluded from the final analysis because they reported that they were no longer practicing medicine. The remaining 708 physicians were separated into three mutually exclusive groups: Group 1, board certified (n = 138); Group 2, residency only (n = 282); and Group 3, no residency (n = 288) (Table 1). The majority of respondents were male, more than fifty years old, and reported attending medical school outside of Puerto Rico, mostly notably in the Dominican Republic, Spain, or Mexico. Respondents reported practice locations throughout the island. The proportion of male respondents (62%) differed significantly, if not dramatically, from the reported physician population proportion of males (69%) (p = 0.01), while the rates for physician office location did not (San Juan metro area, p = 0.42; Ponce, p = 0.48).
Respondent characteristics varied by group (Table 1). Groups 1 and 2 had roughly similar age distributions, but a higher proportion of Group 1 respondents attended medical school in Puerto Rico and reportedly practiced medicine in the San Juan Metro Area when compared to Group 2 or 3 respondents. Group 3 respondents were more likely than those from Group 1 and 2 to be older, trained in Dominican Republic, and practicing outside of the San Juan Metro Area.
Respondents clinically diagnosed on average 12 cases of dengue in the three months before participating in the survey (Table 2). Slightly more than half (56%) of all respondents stated that they report all clinically suspected dengue cases to public health officials while about one third (31%) said that they do not report any suspected cases. During this same time period, respondents requested dengue diagnostic testing for only three cases on average. Few respondents were able to correctly identify laboratory assays used to diagnose acute DENV infections, however, this varied significantly by group with Group 1 respondents being most likely to respond correctly (Table 3).
Methods used to clinically diagnose patients with dengue did not differ by group (Table 3). The majority (92%) reported always using criteria consistent with the 1997 WHO case definition to identify suspected dengue cases while a similar proportion (96%) reportedly always use platelet count or white cell count (89%) to identify suspected cases. Less than one quarter (19%) of all respondents reported using the tourniquet test to identify suspected dengue cases.
Knowledge of warning signs for severe dengue and early signs of shock was low overall and knowledge varied by group (Table 3). One-third (29%) of respondents overall correctly identified tachycardia and delayed capillary refill as early signs of shock, and this proportion increased from Group 3 to Group 1. One half (48%) of all respondents were able to correctly identify severe abdominal pain and persistent vomiting as warning signs with a higher proportion of Group 1 respondents than Group 2 or 3 respondents being able to do so. Ability to identify all warning signs from a list was low (26%), with Group 1's ability significantly higher than Groups 2 and 3 (p = 0.04).
About one third (31%) of respondents reportedly use hospital referral criteria consistent with the 1997 guidelines (Table 3). Not all respondents reportedly refer suspected dengue patients who have a hemorrhagic manifestation; 68% of all respondents refer suspected dengue patient with minor bleeding (e.g., epistaxis, gum bleeding) in the absence of shock or hemoconcentration. About one third (32%) of all respondents use a platelet count of ≤100,000 cells/mm3 in the absence of bleeding, hemoconcentration, or shock as a criteria for hospital referral.
Knowledge of the WHO treatment guidelines varied among groups (Table 4). When given the scenario of a suspected dengue patient with persistent shock and an elevated hematocrit level after a trial of an intravenous crystalloid solution, one third (30%) of all respondents correctly responded that they would give the patient an intravenous colloid. This proportion varied from 39 to 23% among groups with Group I having the highest proportion of correct answers. However, a higher proportion of Groups 1 (48%) and 2 (47%) respondents said that they would give the patient a vasopressor given this scenario. Given a second scenario where a suspected dengue patient has persistent shock and a decreasing hematocrit level after a trial of intravenous crystalloids, about half (46%) of all respondents correctly identified blood transfusion as the treatment of choice, and the proportions among the groups were not statistically significantly different.
Many (57%) of the respondents appropriately never give corticosteroids to their patients with suspected dengue (Table 4). The majority (72%) of Group I respondents reportedly do not use corticosteroids while slightly more than half of Group 2 and 3 respondents reported not giving corticosteroids to their suspected dengue patients. Likewise, the same proportion (57%) of respondents correctly does not give prophylactic platelet transfusions, and this practice varied in a similar fashion by group. Among the 63 respondents who reportedly give prophylactic platelet transfusions, 41 (65%) individuals stated that their threshold for giving platelets is between 25,000 and 50,000 cells/mm3, while 22 (35%) gave ≤20,000 cells/mm3 as their threshold. The overwhelming majority (92%) of respondents appropriately do not give intravenous immunoglobulin to their patients with dengue.
This survey demonstrates that knowledge and management of dengue vary among physicians practicing in Puerto Rico, particularly between Board-certified physicians and non-Board-certified physicians, especially those who did not complete residency training. There were four important findings from this survey. First, while most reportedly use WHO case definition to clinically diagnose dengue cases, we found that case reporting to public health authorities is not optimal and knowledge of laboratory diagnosis of dengue was poor. Second, many respondents, regardless of their level of training, were unable to identify early signs of shock and warning signs for severe dengue, knowledge needed to effectively give anticipatory guidance and inform triage and referral decisions. Third, reported compliance with treatment guidelines of dengue patients in refractory shock was generally low. Fourth, corticosteroids and prophylactic platelet transfusions were reportedly used by about 40% of respondents; practices that are not recommended by current or past treatment guidelines [13], [18]–[20].
While respondents reportedly had clinically diagnosed 12 dengue cases on average in the preceding three months, about one-third of respondents did not report any cases to the Puerto Rico Department of Health and half reported all cases as required by law. Taken together, these findings suggest that dengue is underreported in Puerto Rico. This finding is consistent with past studies that estimated that for every case of suspected dengue reported to the passive dengue surveillance system (PDSS) in Puerto Rico, ten to 27 cases are not reported, and for every case of dengue hemorrhagic fever (DHF) reported, 15 DHF cases are not reported [21], [22]. Given these findings, it was not surprising to find that few (∼6%) respondents knew which laboratory tests are used to diagnose acute dengue as they do not routinely report cases to PDSS, a system that requires submission of a case investigation form and serum sample for case reporting and free diagnostic testing.
Treatment guidelines for the clinical management of dengue were first introduced by WHO in 1975 [18] and they were then updated in 1997 [13] and 2009 [19]. The 1997 WHO guidelines were translated into Spanish, widely distributed throughout the Caribbean, and in use when this survey was administered (October 2007 to February 2008) [20]. Identification of dengue patients with early signs of shock and warning signs for severe dengue and timely initiation of supportive care is the cornerstone of dengue clinical management. Survey respondents, regardless of their level of training, were largely unable to identify early signs of shock and warning signs for severe dengue. Moreover, because of their lack of knowledge, most respondent's reported hospital referral criteria deviated from WHO guidelines. Those guidelines recommended that dengue patients without bleeding or warning signs could be monitored at home by family members while clinicians monitor platelet count and hematocrit as an outpatient. Suspected dengue patients with any hemorrhagic manifestation and patients with a platelet count <100,000 cells per mm3 concurrent with an elevated hematocrit for age and sex were to be referred to a hospital for further evaluation. Dengue patients with signs of shock and/or warning signs were to be referred for inpatient hospitalization. Consistent with these findings fatal case review studies conducted in Puerto Rico have found missed opportunities for referral and hospital admission [12], [23].
Both the 1997 guidelines and the current 2009 WHO guidelines have comparable treatment algorithms for the use of intravenous crystalloids, colloids, and blood transfusions in dengue patients with refractory shock. Findings from our survey suggest that these treatment algorithms, especially the use of intravenous colloids for refractory shock due to severe plasma leakage, may not be as widely used as should be. In the same year as the survey was conducted, a medical record review from a case-series of laboratory-positive fatal dengue cases in Puerto Rico found that only one patient was given an intravenous colloid solution before the terminal event even though six of eight case-patients who died in the hospital had refractory shock [12]. This is noteworthy because application of the WHO treatment guidelines have been associated with a reduction in case fatality rates from 10 to less than 1% among patients with severe dengue [4]–[8].
Even though WHO guidelines [13], [18]–[20] and a 2006 Cochrane review [24] recommend against the use of corticosteroids in patients with dengue, 43% of respondents reported prescribing corticosteroids, a finding corroborated by the 2007 fatal dengue case-series that found that 55% of fatal laboratory-positive dengue cases were given a corticosteroid [12]. This practice also occurs in other dengue endemic countries [25], [26]. Reasons given by respondents for use of corticosteroids included use as an immune modulator given that severe manifestations of dengue are thought, in part, to be immune mediated (CDC, data not presented). However, a recent randomized clinical trial evaluating the early use of oral prednisolone in dengue patients found treatment to have little impact on the host immune response [27]. In addition, while the trial was not powered to assess efficacy, there was no evidence that treatment lead to a reduction in the severity of plasma leakage, or the development of shock or clinical bleeding. In short, with no evidence of therapeutic benefit and multiple potential side effects including hyperglycemia, immunosuppression, secondary infections, and gastrointestinal bleeding in critically ill patients, corticosteroids should not be used to treat patients with dengue [24], [27].
Despite a lack of evidence, many of our survey respondents reported giving prophylactic platelet transfusions to their patients with dengue; a practice that may be relatively common among physicians in dengue endemic countries [25], [28]. Several studies have found no correlation between platelet count and bleeding or bleeding severity in patients with dengue, and when given, prophylactic platelet transfusions do not expedite platelet recovery [29]–[32]. Moreover, the practice is costly and may contribute to fluid overload and the development of pulmonary edema resulting in increased hospital stays among dengue patients. A recent randomized control trial suggested that prophylactic platelet transfusions have no therapeutic benefit when given to patients with dengue and they may be associated with adverse outcomes including transfusion reactions [33]. A clinical trial is currently ongoing to further evaluate the use of prophylactic platelet transfusions among patients with dengue (ClinicalTrials.gov identifier NCT01030211).
Although these findings contribute to our understanding of the knowledge and management of dengue in Puerto Rico, our study has several limitations. First, non-response might have biased the results [34], [35]. Demographic and training characteristics of respondents and non-respondents were similar in most respects, though there was somewhat higher response by females, but we do not expect that this or other differences likely relate to their level of knowledge and practice. Second, our survey relied on self-reported practices and the accuracy of this information is not known. Previous studies suggest that physicians often over state their compliance with clinical guidelines when compared with chart review [36], [37]. An evaluation is ongoing to confirm actual practice patterns for hospitalized dengue patients in Puerto Rico. Lastly, differences in knowledge and practices we found among physician groups may be explained by non-Board certified physicians having less contact with severe dengue patients. However, there was no difference among the groups in the average number of clinical diagnoses made in the three months before participating in the survey.
In summary, our survey suggests that despite dengue being endemic in Puerto Rico for more than 40 years, physicians' diagnosis and clinical management of dengue in Puerto Rico are not optimal. As other dengue endemic countries have reported similar findings, a sustained continuing medical education training initiative may be necessary to improve case detection and clinical management even in countries where the disease is common [25], [28]. Findings from this survey were used to develop and implement a post graduate clinical management course attended by more than 8,000 physicians licensed to practice in Puerto Rico in 2010 and create an on-line version of the course that was released in March of 2014. Further study is needed to determine if focused training can improve clinical management by minimizing failed early recognition of severe dengue and delayed initiation of supportive care that can result in higher rates of medical complications, longer hospital stays, and increased hospital costs. An evaluation of the course and its impact on the clinical management of dengue in Puerto Rico is ongoing.
|
10.1371/journal.pntd.0001183 | The Use of a Mobile Laboratory Unit in Support of Patient Management and Epidemiological Surveillance during the 2005 Marburg Outbreak in Angola | Marburg virus (MARV), a zoonotic pathogen causing severe hemorrhagic fever in man, has emerged in Angola resulting in the largest outbreak of Marburg hemorrhagic fever (MHF) with the highest case fatality rate to date.
A mobile laboratory unit (MLU) was deployed as part of the World Health Organization outbreak response. Utilizing quantitative real-time PCR assays, this laboratory provided specific MARV diagnostics in Uige, the epicentre of the outbreak. The MLU operated over a period of 88 days and tested 620 specimens from 388 individuals. Specimens included mainly oral swabs and EDTA blood. Following establishing on site, the MLU operation allowed a diagnostic response in <4 hours from sample receiving. Most cases were found among females in the child-bearing age and in children less than five years of age. The outbreak had a high number of paediatric cases and breastfeeding may have been a factor in MARV transmission as indicated by the epidemiology and MARV positive breast milk specimens. Oral swabs were a useful alternative specimen source to whole blood/serum allowing testing of patients in circumstances of resistance to invasive procedures but limited diagnostic testing to molecular approaches. There was a high concordance in test results between the MLU and the reference laboratory in Luanda operated by the US Centers for Disease Control and Prevention.
The MLU was an important outbreak response asset providing support in patient management and epidemiological surveillance. Field laboratory capacity should be expanded and made an essential part of any future outbreak investigation.
| A mobile laboratory unit (MLU) was deployed to Uige, Angola as part of the World Health Organization response to an outbreak of viral hemorrhagic fever caused by Marburg virus (MARV). Utilizing mainly quantitative real-time PCR assays, this laboratory provided specific MARV diagnostics in the field. The MLU operated for 88 consecutive days allowing MARV-specific diagnostic response in <4 hours from sample receiving. Most cases were found among females in the child-bearing age and in children less than five years of age including a high number of paediatric cases implicating breastfeeding as potential transmission route. Oral swabs were identified as a useful alternative specimen source to the standard whole blood/serum specimens for patients refusing blood draw. There was a high concordance in test results between the MLU and the reference laboratory in Luanda operated by the US Centers for Disease Control and Prevention. The MLU was an important outbreak response asset providing valuable support in patient management and epidemiological surveillance. Field laboratory capacity should be expanded and made an essential part of any future outbreak investigation.
| Marburg virus (MARV) is classified as members of the family Filoviridae, genus Marburgvirus, type species Lake Victoria marburgvirus. A single species has been described which includes several virus strains [1]. Today, the geographic distribution of MARV seems to primarily involve areas in East Africa within 500 miles of Lake Victoria, Zimbabwe, but also western Africa [2], [3]. MARV is of zoonotic nature with an as yet unidentified reservoir in nature, but with strong cumulative evidence that bats are involved in the zoonotic cycle [4], [5] as this has also been implicated for Ebola virus [6].
MARV is the causative agent of Marburg hemorrhagic fever (MHF), a disease that was first described in 1967 among laboratory workers in Germany and former Yugoslavia [7]–[9]. Until 1998, only sporadic MHF cases have occurred in Zimbabwe/South Africa (1975) and in Kenya (1980 & 1987) [10]–[12]. The first community-based MHF outbreak was reported in 1998–2000 from the Watsa/Durba region in the Democratic Republic of the Congo (DRC) [13], [14]. In 2004/2005 MARV first appeared in western Africa, Angola, causing to date the largest MHF outbreak on record [15], [16]. The latest MHF episodes involved 4 reported cases from western Uganda associated with a single mine (2007) [5], and two imported cases into the US and the Netherlands, who independently visited the same cave in Uganda (2008) [17], [18] (Table 1). In addition, three laboratory exposures, one of them fatal, have been reported [9], [19], [20].
In March 2005, the National Microbiology Laboratory (NML) of the Public Health Agency of Canada (PHAC) offered assistance to the World Health Organization (WHO) as a partner of the ‘Global Outbreak Alert & Response Network’ (GOARN) (http://www.who.int/csr/outbreaknetwork/en/) for the MHF outbreak in Angola. Under GOARN, a Mobile Laboratory Unit (MLU) was deployed to Uige, the epicentre of the outbreak, to assist in clinical management and epidemiological surveillance with MARV-specific and limited differential diagnostic capacity. Here we discuss the usefulness of this latest response capacity for the management of viral hemorrhagic fever outbreaks.
Laboratory space was made available for the MLU in the Paediatric Ward of the Uige Provincial Hospital (Figure 1). Four rooms were used for the laboratory set up to ensure isolation of infectious work from other activities and to separate PCR assay steps to minimize contamination. Two rooms were located on one side of a central hallway; the smaller of the two rooms was accessible by a single door and had no windows or other opening and was utilized for infectious work (‘hot room’). The anteroom to this room was used for the preparation for entry to the infectious room and the subsequent disinfection of the worker following infectious work. Opposite these rooms were two additional rooms; one was used for RNA extraction and running the Q-RT-PCR and the other room was utilized as a ‘clean room’ for master mix preparation. Reagents and the laboratory team (2–3 members) were replaced every three weeks; in total NML deployed six teams to Angola to cover the period of April 1 to June 27, 2005.
Clinical samples were collected by personnel wearing personal protective equipment (PPE) including a surgical mask, cap, shield or goggles, gown, apron, gloves (two pairs) and boots. Swab samples (nasal and oral) were collected using cotton tipped applicators (AMG Medical, VWR, Mississauga, ON, Canada). Applicator tips were stored in 700 µl of Dulbecco's modified essential medium (DMEM) or phosphate buffered saline (PBS) supplemented with 5% bovine serum albumin (Invitrogen, Burlington, ON, Canada). Whole blood and serum samples were collected using EDTA and serum vacutainer tubes, respectively. For transport, tubes were sealed in plastic bags, surface disinfected with a 1% hypochlorite solution, sealed into a second bag or container and again surface disinfected. Collection of human specimens occurred on an outbreak response protocol and was approved by the local Scientific and Technical Coordination Committee in Uige, Angola.
Infectious specimens were manipulated in the field laboratory by personnel wearing Tyvek suits and HEPA filter-equipped powered air purifying respirators, in a room isolated and dedicated for this work (Figure 1). An aliquot (140 µl) was removed from each sample and inactivated by adding 560 µl of the guanidine thiocyanate lysis buffer AVL. The sample tubes were submerged in 1% hypochlorite solution for 10 minutes and released from the infectious area. All further work was performed with PPE as outlined above. For RNA isolation we used the QIAamp Viral RNA mini kit (Qiagen, Mississauga, ON, Canada). All waste material was treated with 1% hypochlorite solution and incinerated on the same day. Two separate sample aliquots were prepared for transportation to the reference laboratory in Luanda operated by the Special Pathogens Branch of the US Centers for Disease Control and Prevention (US-CDC). Remaining samples were forwarded to the National Institute for Communicable Diseases (NCID), Sandringham, South Africa, and finally shipped to the US-CDC (Atlanta) or NML (Winnipeg) for further testing. Transportation was carried out in compliance with International Air Transport Association (IATA) regulations after prior approval by the appropriate national authorities of the sending or receiving countries.
Initially, two quantitative real-time PCR (Q-RT-PCR) assays were used that targeted regions of the polymerase (L) [MARVLF-TTATTGCATCAGGCTTCTTGGCA, MARVLR–GGTATTAAAAAATGCATCCAA (AY358025; bp.13321–133517)] and the glycoprotein (GP) genes [MARVGPF–AAAGTTGCTGATTCCCCTTTGGA, MARVGPR–GCATGAGGGTTTTGACCTTGAAT (AY358025; bp.6131–6355)]. Later, an assay that targeted the nucleoprotein (NP) gene [MARVNPF–TGAATTTATCAGGGATTAAC, MARVNPR–GTTCATGTCGCCTTTGTAG (AY358025; bp.967–1146)] was used in place of the GP assay. The switch to an NP target was the result of testing that indicated this target was potentially more sensitive and provided a more distinct melting curve which simplified interpretation. MARV RNA was detected using the Lightcycler RNA Amplification SYBR Green I kit (Roche, Laval, PQ). Briefly, 5 µl of RNA was added to 20 µl of master mix containing 1X SYBR Green I mix, 5 mM MgCl2, 0.6 µM forward and reverse primers and 0.5 µl of the enzyme mix. Q-RT-PCR assays were run on Smartcycler thermocyclers (Cepheid, Sunnyvale, CA). A reverse transcriptase step at 50°C for 20 minutes and a 2 minute inactivation step at 94°C were followed by 40 cycles at 94°C for 15 seconds, 50°C for 30 seconds and 72°C for 30 seconds where a single acquisition point was taken. Melt curve analysis was performed to confirm the identity of amplification products. Samples were considered positive if they produced melting point confirmed amplification products in two assays. Amplification products were later confirmed by sequencing at NML (Winnipeg).
The algorithm for the laboratory testing and the rational for positive/negative test results are presented in Figure 2. Overall, the MLU tested 620 clinical specimens from 388 patients/individuals over an operation period of 88 days. The clinical specimens included mainly oral swabs and EDTA blood/serum samples; the remainder consisted of nasal and conjunctival swabs and breast milk. The sample source and test results of individuals tested are presented in Table 2.
The daily case load of the MLU fluctuated, with the number of individuals analyzed per day varying between 0 and 14 (Figure 3). This analysis often included multiple samples per individual on a single day and serial surveillance sampling of suspect and confirmed cases. The age and sex distribution of individuals tested were slightly shifted towards females (68%) and the younger age groups, in particular children under the age of 5 years (by far the largest single age group at 21%). The distribution of positive cases clearly demonstrated a larger proportion of females and children among the infected individuals (Figure 4).
A comparison of detection of MARV from oral swabs and EDTA blood was performed on 63 individuals from whom both specimen types were available from the same day. Both samples sources yielded identical test results in 98.5% of the individuals with roughly 33% positive and 66% negative for MARV. Cycle threshold (Ct) values for most paired samples did not differ markedly indicating similar viral loads in both specimen sources (Figure 5). Testing on some patients did provide disparate results for blood and swab samples but test results were identical even in these instances. Similarly, for 12 individuals, both oral and nasal swabs specimens were collected which resulted in identical test results and no significant differences in Ct values for the positives. Additionally, 3 breast milk specimens from laboratory-confirmed female MHF cases were analyzed and shown to be positive for MARV (data not shown).
We did not experience any evidence for PCR contamination during the entire operation. All controls produced the expected positive and negative results. Nevertheless, all samples tested in Uige were subsequently shipped to Luanda for confirmation at a US-CDC established biosafety level 3 (BSL3) laboratory using a real-time PCR hybridization assay targeting the matrix protein (VP40) gene, an antigen capture enzyme-linked immunorsorbent (ELISA) assay and antibody (IgM and IgG) detection ELISAs [16]. Overall, the reference laboratory confirmed test results of the MLU in 97.5% of all specimens analyzed and in all but one case. The high concordance between field and reference laboratory results supported the on-site report of the MLU results to the ward and the surveillance teams, allowing a turn-around time of <4 hours from sample receiving to laboratory diagnosis. After closing the MLU, further clinical specimens were shipped to Winnipeg for diagnosis via Luanda (US-CDC) and Sandringham (NCID). Eventually, all specimens were shipped to the BSL4 laboratories in Atlanta and/or Winnipeg for additional analysis. Sequence analysis of all amplified products and of several virus isolates obtained at the US-CDC [16] and NML (authors, unpublished data) demonstrated a high degree of conservation indicating a single or very few introductions into the community, with subsequent human-to-human transmission.
Differential diagnostic testing was only performed for malaria (Plasmodium spp.) using a real time PCR assay targeting the ssuRNA gene [21]. Test results for 19 individuals demonstrated two groups of patients, mild or asymptomatic (Ct values >20) and symptomatic individuals (Ct values <20), based on parasitemia levels (data not shown). The value of this diagnostic tool needs to be further evaluated.
Under current filoviral hemorrhagic fever outbreak operation protocols several activities are undertaken where accurate and rapid diagnostic testing can have significant impact:
To obtain diagnostic testing, specimens have normally been shipped to an international reference laboratory such as the Institut de recherche pour le développement (IRD), Franceville, Gabon; NCID in Sandringham, South Africa; or the US-CDC in Atlanta, United States resulting in a significant delay (days to weeks due to shipment issues) in laboratory diagnosis with limited or no benefit for acute case patient or outbreak management [22]–[24]. Therefore, such operation protocols require a fairly large infrastructure, longer hospitalization periods, and more staff and consequently increase resources and exposure risks. An MLU, providing testing results in a 4 hour turn around, can be an integral part of the outbreak response and simplify lessen many of the efforts needed to quickly contain and control the outbreak.
Laboratory testing of a symptomatic individuals during triage will allow the team to quickly assess if the person is a case or not. Confirmed cases can be appropriately isolated and supportive care initiated. Symptomatic individuals with negative test results can be maintained separate from confirmed cases either by releasing to another ward or kept in an observation ward for follow up testing or discharging. In Uige, and to a lesser extent also at previous outbreak locations, the isolation ward was largely unacceptable to the local population and significant resistance was present to have family members admitted [15], [25]. However a positive test result for MARV was normally sufficient to convince people of the necessity for admission to the ward. Isolating only those individuals who require it will reduce the infrastructure needed for isolation, minimize the hospitalization time for non-cases, reduce the number of staff and consequently reduce the risk of exposure for both staff and non-cases.
Cases that can be confirmed or excluded by laboratory testing can significantly contribute to one of the most important outbreak control measures, contact tracing. The current protocols call for the follow-up of contacts of suspected cases for 21 consecutive days. The presence of a field laboratory can help to arrive at a rapid confirmed final diagnosis for each suspected case, thereby decreasing the burden of field teams who may frequently be conducting contract tracing of cases with uncertain diagnosis.
Testing in this outbreak found that oral swabs from severely ill or deceased patients were a suitable sample for MARV testing by Q-RT-PCR. This allowed the MLU to safely test samples from corpses of unknown cause and when possible, to release MARV-negative bodies to the family members for traditional and religious burial procedures, a sensitive issue with almost all local communities in endemic areas. The value of swabs from corpses for diagnostic purposes needs to be further evaluated in future outbreaks and perhaps confirmed by other technologies such as immunohistochemistry [26]. Post mortem RNA degradation might render a test falsely negative even so infectious Ebola virus has been detected in blood samples more than a month after blood draw and storage at room temperature [27]. Any test results should take clinical presentation and epidemiology into account.
A growing concern is the return of negative and convalescent patients to the community, which may increase with the implementation of more advanced case patient care and the perspective of treatment options in the future [2], [24], [28]. These people are often shunned by their families and neighbours and a timely negative test result as provided through the MLU may aid in their re-acceptance and safe re-introduction into the community.
In Angola, field diagnostic support was used for the first time in response to a MHF outbreak. Also the first time, the combined operation of a field and reference laboratory allowed for a unique evaluation of field diagnostic capacity under difficult circumstances and proved it to be accurate, efficient and safe in operation. There have been previous attempts to provide field laboratory diagnostics for outbreaks of Ebola hemorrhagic fever. In 1976 during the Zaire ebolavirus outbreak an immunofluorescence assay was used for acute case identification but the results were considered poor [29]. In 2000 during the Ebola outbreak (Sudan ebolavirus) the US-CDC operated a laboratory within the Gulu district at St. Mary's Lacor Hospital, Uganda, and used antigen capture and reverse transcription nested PCR (RT-PCR) to successfully diagnose infection in suspected patients [30]. In 2003 during the Ebola outbreak (Zaire ebolavirus) in Mbomo, The Republic of the Congo, NML together with partners from the IRD, Franceville, Gabon, and the Bundeswehr Institute of Microbiology, Munich, Germany, operated a small field laboratory under the lead of WHO using antigen capture and Q-RT-PCR to diagnose acute cases [31], [32].
In general, the usefulness of on-site laboratory support during filovirus outbreaks is not really questioned [2], [24], and, in particular, the positive experience from this MHF outbreak demonstrate that rapid turn-around RT-PCR diagnostics can clearly aid in surveillance and case management [15], [25]. PCR-based techniques can be prone to contamination resulting in false positive results. Here we used a technique that did not require opening of tubes largely reducing the risk of contamination. Other concerns have been raised towards the reliability of RT-PCR assays during early disease stages and for survivors in the early convalescent stage, the consequences of false-positive and false-negative results of RT-PCR assays could be dire to outbreak management [30]. Indeed, PCR-based assays, like other diagnostic tests, have weaknesses and do not produce reliable results under all circumstances. Therefore, independent, methodologically different, confirmatory assay such as antigen capture to support RT-PCR should be mandatory. However, nowadays most laboratories depend on PCR detection as their first and most rapid diagnostic methods and there are good reasons to support that choice [33]. If a confirmatory assay is not available or unsuccessful, alternatives for RT-PCR confirmation include sample re-extraction, a second clinical specimen and/or an assay with independent targets (Figure 2). Nevertheless, any diagnostics should not replace general and common sense precautions in case patient management and on-site laboratory diagnostics should be in close proximity to the ward allowing for continuous interaction between physicians/nurses and laboratory personnel [15], [25]. Importantly, during this field laboratory deployment, Q-RT-PCR proved to be very sensitive and reliable even in this challenging environment. Patient samples were positive in our testing beginning on the day of onset of symptoms but we did see that detection in swab samples could be delayed by a few hours when compared to blood this early in the course of illness.
The collection of appropriate clinical specimens for diagnostic testing has become an increasing problem during filovirus outbreaks. The reasons for this can include the lack of properly trained personnel, fear of personnel to apply invasive procedures, cultural objections to bleeding and any other invasive pre- and post mortem sampling procedure, and insufficient infrastructure for sampling and transportation [22], [24], [34]. In that respect, the MHF outbreak in Angola was not different from previous outbreaks. In particular, resistance in the community to bleeding and post mortem invasive procedures, such as cardiac puncture or liver biopsy, and the increasing resistance of aid personnel to apply invasive procedures in the field (community) made oral swabs the predominant clinical specimen available for testing. As demonstrated here on paired blood/oral swab samples, in general there was no significant difference in viral load between oral swabs and EDTA blood taken at the same time (Figure 5). This supported oral swabs as an alternative diagnostic specimen to blood. The few incidences when oral swabs were less suitable than EDTA blood related to early disease stage and early convalescent stage samples. Lower viral loads in oral swabs compared to EDTA blood, at these stages, are likely to explain this discrepancy. Additionally, there are inherent sampling variables associated with oral swabs (the technique and efficiency of swabbing; moisture level of the oral cavity) that are not present in a blood draw, which may also have a role in these differences. However, despite the fact that oral swabs seemed to have been an appropriate specimen source for laboratory testing during this outbreak, and oral/nasal swabs are valuable alternatives in cases of resistance in the affected population to invasive procedures, EDTA blood should remain the priority choice for a clinical specimen due to the longer period of detectable viremia, the suitability to serological-based testing, and the value for monitoring potential point of care therapies in future.
While this study is not a detailed epidemiologic study, brief mention of some of the data is warranted as it has not been yet published elsewhere. This MHF outbreak was unique in regards to its location, case number and case fatalities, but also showed a large proportion of paediatric cases and cases among woman in the child bearing ages [2], [24]. Since MARV, as Ebola virus, are usually transmitted through close contact with blood, secretions or excretions from infected patients, family members and medical personnel caring for patients or preparing bodies for burials are considered high risk exposure groups [2], [34]. It has been proposed that because women provide the majority of in-home care that this was the reason for the preponderance of cases in women [35]. Certainly women provide the majority of care for the children and since, especially early in the outbreak, children less than 5 years of age represented the largest single age group affected may also be reflective of this fact. Furthermore, the detection of MARV in breast milk during this outbreak indicates that breastfeeding might have played a role in virus transmission. This is supported by epidemiological data indicating transmission from infected mothers to their nursing babies followed, after death of the mothers, by virus transmission from the infected babies to wet nurses who subsequently infected their own nursing child (authors, unpublished observation). Other factors may have come into play including the alleged lack of appropriate infection control within the paediatric ward prior to the identification of the outbreak [36]. It is very unlikely that the predilection of women and young children represents a biological predisposition, given that the demographics of the outbreak changed through the course of the outbreak (i.e. early in the outbreak a very high percentage were paediatric cases whereas later cases became more evenly distributed by age), and yet the virus changed very little [16]. Without more detailed epidemiologic data, it remains unclear which of these transmission routes constituted significant mechanisms for virus spread in the Uige outbreak.
Offering differential diagnosis significantly increases the value of on-site diagnostics. This is much harder to achieve in the field and requires variable clinical specimen (in particular blood or stool), more manpower and more extensive and continuous supplies. At a minimum, malaria diagnostics (e.g. commercially available rapid dipstick tests) and diagnosis for severe gastrointestinal infections should be available. Proper case patient management including intravenous fluid administration would also require blood chemistry and haematology analysis, another capacity that needs to be considered for expansion of a field laboratory response capacity.
Most of what constitutes the MLU can be sourced from equipment that most reference laboratories would have access to from their normal compliment of equipment and supplies, however a dedicated MLU would likely require the investment of approximately $100 000 and a weekly deployment cost of $2000 for reagents and supplies. Logistic needs and costs during a mission can be best managed through a close working relationship with other organizations including the WHO and Médecins Sans Frontières (MSF). The greatest challenge to the operation of the MLU was the lack of consistent electrical power and our reliance on portable generators. This necessitated the use of battery backup systems for thermocyclers and did not allow for storage of samples or reagents at freezing temperatures as freeze-thaw cycles could not be avoided. Fortunately, all reagents were relatively stable at 4°C over a three week rotation period before replacement teams replenished the reagents. We were able to efficiently operate the MLU using teams of two members as the workload and workflow rarely justified additional staff. We have since recommended that teams of three be deployed to allow for rest and health issues.
In conclusion, the combined operation of a field and reference laboratory in this outbreak allowed for a unique evaluation of field diagnostic capacity under difficult circumstances. Rapid MARV-specific Q-RT-PCR was useful for triage and assessing the need for isolation. The quick turn-around of laboratory diagnosis on the basis of Q-RT-PCR assays significantly improved outbreak response efforts. Therefore we propose: “On-site laboratory diagnosis should become a routine part of any future filovirus outbreak response as it provides all responders with valuable information to help minimize the extent and durations of these events”.
|
10.1371/journal.ppat.1003830 | Lysine Acetyltransferase GCN5b Interacts with AP2 Factors and Is Required for Toxoplasma gondii Proliferation | Histone acetylation has been linked to developmental changes in gene expression and is a validated drug target of apicomplexan parasites, but little is known about the roles of individual histone modifying enzymes and how they are recruited to target genes. The protozoan parasite Toxoplasma gondii (phylum Apicomplexa) is unusual among invertebrates in possessing two GCN5-family lysine acetyltransferases (KATs). While GCN5a is required for gene expression in response to alkaline stress, this KAT is dispensable for parasite proliferation in normal culture conditions. In contrast, GCN5b cannot be disrupted, suggesting it is essential for Toxoplasma viability. To further explore the function of GCN5b, we generated clonal parasites expressing an inducible HA-tagged dominant-negative form of GCN5b containing a point mutation that ablates enzymatic activity (E703G). Stabilization of this dominant-negative GCN5b was mediated through ligand-binding to a destabilization domain (dd) fused to the protein. Induced accumulation of the ddHAGCN5b(E703G) protein led to a rapid arrest in parasite replication. Growth arrest was accompanied by a decrease in histone H3 acetylation at specific lysine residues as well as reduced expression of GCN5b target genes in GCN5b(E703G) parasites, which were identified using chromatin immunoprecipitation coupled with microarray hybridization (ChIP-chip). Proteomics studies revealed that GCN5b interacts with AP2-domain proteins, apicomplexan plant-like transcription factors, as well as a “core complex” that includes the co-activator ADA2-A, TFIID subunits, LEO1 polymerase-associated factor (Paf1) subunit, and RRM proteins. The dominant-negative phenotype of ddHAGCN5b(E703G) parasites, considered with the proteomics and ChIP-chip data, indicate that GCN5b plays a central role in transcriptional and chromatin remodeling complexes. We conclude that GCN5b has a non-redundant and indispensable role in regulating gene expression required during the Toxoplasma lytic cycle.
| Toxoplasma gondii is a protozoan parasite that causes significant opportunistic infection in AIDS and other immunocompromised patients. Acute episodes of toxoplasmosis stem from tissue destruction caused by the rapidly growing form of the parasite, the tachyzoite. In this study, we identify a lysine acetyltransferase (KAT) enzyme called GCN5b that is an essential driver of tachyzoite proliferation. Our studies show that GCN5b is present at a wide variety of parasite genes and that expression of defective GCN5b compromises gene expression through its diminished ability to acetylate histone proteins. We also identified the likely mechanism by which GCN5b is recruited to target genes by co-purifying this KAT with plant-like AP2-domain proteins, a subset of which function as DNA-binding transcription factors in Apicomplexa. Our findings demonstrate that KATs play a critical role in parasite replication, which leads to tissue destruction and acute disease in the host. Parasite KAT enzyme complexes may therefore serve as attractive targets for future drug development.
| Lysine acetylation of histones is a well-characterized post-translational modification linked to the activation of gene expression. Initially identified in the free-living protozoan Tetrahymena thermophila, the first histone acetyltransferase (HAT) was homologous to a transcriptional adaptor protein in yeast known as GCN5 [1]. It has since been elucidated that GCN5 is a highly conserved catalytic component present in multiple protein complexes linked to the regulation of gene expression [2]. When GCN5 HATs were found to have non-histone substrates as well, they became referred to as lysine (K) acetyltransferases (KATs) [3]. The number of GCN5 KATs and their impact on cells or organisms depends on the species. GCN5 generally appears to be required for stress responses [4]–[6]. Consistent with this idea, the single GCN5 is dispensable in Saccharomyces cerevisiae, but required for growth on minimal media [7]. In contrast, mammals possess two GCN5 family members, one of which is required for mouse embryogenesis [8]. Null mutants of the other GCN5 (also known as PCAF, or p300/CBP-associating factor) have no discernible phenotype in mice [8], [9]. Together, these reports suggest that GCN5-mediated acetylation is an important facet of cellular biology, particularly during stress or adaptive responses.
The relevance of lysine acetylation in pathogenic protozoa is underscored by potent antiprotozoal activity of lysine deacetylation inhibitors like apicidin and FR235222 [10], [11]. Histone acetylation has also been linked to a number of processes that underlie pathogenesis of apicomplexan parasites, including antigenic variation in Plasmodium (malaria) and developmental transitions in Toxoplasma gondii [5], [12], [13]. An extensive repertoire of histone modification machinery is present in these parasites, suggesting that epigenetic-based regulation contributes to gene expression control [14]. A related oddity of the Apicomplexa is that these early-branching eukaryotes appear to use an expanded lineage of so-called Apetela-2 (AP2) proteins as transcription factors rather than the basic leucine zipper (bZIP) factors that are conserved throughout most of the eukaryotic kingdom [15], [16]. ApiAP2 proteins harbor a plant-like DNA-binding domain and emerging evidence supports that at least some function as bona fide transcriptional regulators [17], [18].
Toxoplasma has a number of unusual features with respect to its GCN5 KATs. First, there are two GCN5-family members in Toxoplasma (GCN5a and b) whereas other invertebrates, including Plasmodium, possess only one [19], [20]. Second, both TgGCN5s have long N-terminal extensions devoid of known protein domains. These N-terminal extensions are not homologous to those seen in higher eukaryotes, nor are they homologous to each other or other apicomplexan GCN5s [20]. One function of the TgGCN5 N-terminal extensions is to translocate the KAT into the parasite nucleus via a basic-rich nuclear localization signal [21], [22]. Yeast two-hybrid studies have suggested that the N-terminal extension of Plasmodium GCN5 plays a major role in mediating protein-protein interactions [23]. We previously generated a gene knockout of GCN5a, but similar methods have not produced viable GCN5b knockouts. GCN5a was found to be dispensable for parasite proliferation in vitro, but required for the parasite to respond properly to alkaline stress [5]. These findings are consistent with the well-documented role of GCN5 KATs in the cellular stress response.
The inability to knockout GCN5b suggested it is essential for parasite viability. To gain a better understanding of its function in parasite physiology, we expressed a dominant-negative form of GCN5b that lowered histone acetylation, altered gene expression, and arrested parasite proliferation. We also biochemically purified the multi-subunit GCN5b complex to identify interacting proteins and performed a genome-wide ChIP-chip analysis. Collectively, these findings establish that the KAT GCN5b interacts with AP2 factors to regulate the expression of a wide variety of genes that are essential for parasite replication.
To define the role of GCN5b in Toxoplasma, we attempted to generate a gene knockout. Repeated attempts to disrupt or replace GCN5b using homologous recombination in haploid type I RH strain tachyzoites have not been successful, in contrast to our ability to knockout GCN5a using the same approach [5]. More recent attempts to knockout the GCN5b locus in a Δku80 background also failed to generate viable parasites, further suggesting that GCN5b is essential in tachyzoites.
We then pursued an inducible dominant-negative strategy to ascertain the importance of GCN5b in Toxoplasma. As GCN5 KATs function in multi-subunit complexes, GCN5b is a good candidate for a dominant-negative strategy whereby an ectopically expressed enzymatically dead version would compete for essential interacting proteins from the endogenous protein. Consequently, the activities of the endogenous GCN5b complex would be attenuated. We generated clonal parasites in an RH background expressing a catalytically inactive form of GCN5b (mutated glutamic acid 703 to glycine, E703G [24]) fused to a destabilization domain and HA tag (ddHA) at the N-terminus. In vitro HAT assays using purified ddHAGCN5b proteins confirm the E703G mutation ablates enzymatic activity (Supplemental Figure S1). We also generated a clone expressing wild-type (WT) GCN5b in the same fashion to serve as a control in phenotypic analyses.
The dd domain directs its fusion partner to the proteasome for rapid degradation, but this can be averted by adding Shield ligand to the culture medium [25]. Fusion of ddHA to the N-terminus of GCN5b or GCN5b(E703G) allowed their ectopic expression to be regulated via Shield, as assessed in immunofluorescence assays (IFAs) and immunoblots using anti-HA (Fig. 1). Fusion of ddHA did not disrupt nuclear localization of WT or mutant GCN5b (Fig. 1). No difference in parasite replication was observed between parental wild-type parasites and those expressing ectopic ddHAGCN5b protein at any concentration of Shield (Fig. 2A and 2B). In contrast, parasites induced to express ddHAGCN5b(E703G) underwent rapid growth arrest in 48 hours with as little as 10 nM Shield (Fig. 2C). At 500 nM Shield, over 80% of the parasite vacuoles contained only 16 parasites, compared to the control in which most vacuoles contained 64 parasites. Similar results were obtained when we used a PCR-based assay for the B1 gene to measure parasite replication [5] (data not shown). The growth arrest observed for the Shield-treated ddHAGCN5b(E703G) parasites is reversible, as parasite plaques were evident in monolayers 48 hours after removal of Shield (Supplemental Figure S2). No plaques were present in monolayers infected with ddHAGCN5b(E703G) parasites that were maintained on Shield. As an additional control, ddHAGCN5b(E703G) parasites were treated with 1.0 µM pyrimethamine for 48 hours, which irreversibly kills the parasites as indicated by no plaque formation after removal of the drug (Supplemental Figure S2). These studies indicate that induction of the dominant-negative GCN5b attenuates the activity of the endogenous GCN5b complex, which results in stalled replication. However, as suggested by the genetic studies, complete ablation of GCN5b is not tolerated by tachyzoites.
We considered that the replication arrest observed in ddHAGCN5b(E703G) parasites may be due to a reduction in histone acetylation, which in turn would lead to dysregulation of the transcriptome. To assess this possibility, we analyzed the acetylation level of individual lysine residues in histone H3, the preferred substrate of GCN5 family KATs, purified from Shield- versus vehicle-treated ddHAGCN5b or ddHAGCN5b(E703G) expressing parasites. While total levels of H3 protein remain unaltered, acetylation of lysines K9 and K14 was specifically reduced in parasites expressing ddHAGCN5b(E703G) (Fig. 3). Interestingly, the acetylation status of H3K18 was not affected, which may be explained by the fact that GCN5a, which would not be attenuated by the expression of ddHAGCN5b(E703G), has an exquisite affinity for this particular lysine residue on H3 [20]. These data indicate that the expression of ddHAGCN5b(E703G) protein diminishes acetylation on histone H3.
We reasoned that if the growth arrest of Shield-treated ddHAGCN5b(E703G) parasites was due to hypoacetylation of histones, then inhibition of lysine deacetylases (KDACs) should help restore replication. We therefore incubated ddHAGCN5b(E703G) parasites with 500 nM Shield in combination with increasing sublethal concentrations of either apicidin or TSA, two independent, broad-spectrum KDAC inhibitors. Parasite plaque assays were performed one week post-infection. Consistent with results seen in Fig. 1, virtually no plaques were evident in infected monolayers that contained Shield with no KDAC inhibitor (Fig. 4). However, Shield-treated ddHAGCN5b(E703G) parasites were able to resume proliferation with the inclusion of either KDAC inhibitor in dose-dependent fashion (Fig. 4).
To identify genes regulated by GCN5b, we performed a genome-wide ChIP-chip analysis on HA-tagged GCN5b expressing parasites. Immunoprecipitated DNA associated with GCN5b was identified following hybridization to custom Nimblegen microarrays that tile the entire Toxoplasma genome. Three replicates were performed. GCN5b was detected at 195 tachyzoite genes in all 3 replicates and 1090 genes in two ChIP-chip replicates, when 0.05 was used as the FDR for significant peaks (Supplemental Dataset S1). Although we observe variability in the location of GCN5b peaks, we detect a statistically significant overlap between the three ChIP-chip replicates that is not present when we randomize the peak positions for statistically significant peaks (FDR<0.05; see methods for details).
It was expected that GCN5b would localize to gene promoters, but GCN5b was also detected in gene bodies, with no detectable preference for gene bodies or intergenic regions (Fig. 5) when significant peaks (FDR<0.05) were statistically compared. Analyses that compared “promoters” (H3K9ac; H3K4me3), “active genes” (H3K4me1), and “centromeres” (CenH3) using genome-wide epigenome mapping yielded similar results [26], [27]. The distance between each GCN5b-associating site (FDR<0.05) and the nearest transcription start site (TSS) was also calculated and plotted as a histogram (Supplemental Figure S3). Results show that enrichment of GCN5b occurs up- and down-stream, rather than at the TSSs. These data are consistent with recent studies showing a role for GCN5 in transcriptional elongation by promoting nucleosome eviction [28]–[30]. Overall, the ChIP-chip results establish that GCN5b is present within or near the loci of tachyzoite genes involved in a wide variety of cellular functions (see Supplemental dataset S1 for breakdown of KEGG and GO classifications of the 195 genes identified in all 3 replicates). Many of the genes are annotated as hypothetical genes of unknown function in the ToxoDB. The high confidence genes coinciding with GCN5b localization are associated with gene expression and RNA processing, as well as metabolic genes, rather than genes linked to virulence.
Although GCN5b did not show a detectable preference for functional regions of the genome, there was a significant association with genes containing introns versus those that do not. By hypergeometric test, each of the three ChIP-chip replicates has a statistically significant (p<0.05) preference for intron-containing genes (5,989) versus those that do not (2,083). While associated with intron-containing genes, GCN5b did not show a statistical association with introns versus exons.
We then performed a more targeted approach to verify if select genes detected by the ChIP-chip study were modulated by GCN5b activity. Primers were designed to amplify selected mRNAs in ddHAGCN5b or ddHAGCN5b(E703G) parasites in the presence or absence of Shield. Six primer pairs were designed to amplify mRNAs from GCN5b-associated genes and another 6 primer pairs were designed to mRNAs of genes that were not detected in the GCN5b ChIP-chip (Supplemental Table S1). Virtually no changes in mRNA levels were detected in ddHAGCN5b parasites regardless of whether GCN5b was detected at the gene locus (Table 1). However, expression levels of mRNAs in ddHAGCN5b(E703G) parasites were clearly altered, suggesting a role for GCN5b in gene activation. All 6 genes detected in the GCN5b ChIP-chip had lowered mRNA levels in the ddHAGCN5b(E703G) parasites following Shield treatment (Table 1). Genes that were not detected in the GCN5b ChIP-chip experiments generally exhibited no significant difference in mRNA levels in the ddHAGCN5b(E703G) parasites following Shield treatment, however one gene showed higher mRNA levels and another showed lower mRNA levels (Table 1). The altered expression of these two genes may be due to sensitivity of the ChIP-chip or indirect effects impacting their mRNA levels. Nevertheless, the results show a general trend that is consistent with the decreased acetylation observed in ddHAGCN5b(E703G) parasites leading to decreased transcription of GCN5b-associated genes. This independent qRT-PCR analysis not only highlights the fidelity of the ChIP-chip dataset, but supports the idea that the dysregulation of gene expression induced by the accumulation of ddHAGCN5b(E703G) protein contributes to the arrest in parasite replication.
As GCN5 KATs lack DNA-binding domains, the complex must be recruited to target genes by virtue of a DNA-bound transcription factor, e.g. GCN4 in Saccharomyces cerevisiae or its human counterpart, ATF4 [31]. However, this well-conserved class of transcription factor is not present in Apicomplexa. We therefore performed biochemical purifications of the GCN5b complex from nuclear fractions of intracellular tachyzoites to define the KAT's interactome. Two independent co-immunoprecipitations were performed using RH parasites stably expressing HA-tagged GCN5b. Supplemental dataset S2 is a complete list of proteins identified in each pull down experiment.
As expected, GCN5b itself as well as the known interacting co-activator protein, ADA2-A [20], were identified in each purification. Supporting the idea that DNA-binding proteins would recruit GCN5b to specific gene sites, four AP2 factors were identified in association with GCN5b (AP2IX-7, AP2X-8, AP2XI-2, and AP2XII-4). Another GCN5b-interacting protein with a probable DNA-binding domain is the AT-hook protein (TGME49_109250). Consistent with its frequent localization on gene bodies, GCN5b was associated with RTF1, LEO1, and CTR9, three components of the PAF (polymerase associated factor) complex associated with mRNA elongation [32]. GCN5 activities are typically coordinated with those of SWI/SNF complexes [33], and we detected two distinct SWI/SNF ATPases (TGME49_120300 and TGME49_078440) associated with GCN5b.
The finding that plant-like AP2 factors may partner with KAT complexes to alter transcription is of particular relevance. Associations between GCN5 and AP2 proteins have yet to be demonstrated for any species, including plants. To validate that GCN5b and these AP2 factors reside in the same complex, we endogenously tagged AP2IX-7 and AP2X-8 with a C-terminal 3×HA tag (Fig. 6A). Reciprocal co-immunoprecipitation of each AP2 factor pulled down GCN5b and many of the other proteins seen in the previous GCN5b IPs (Supplemental dataset S2). We refer to proteins that were pulled down consistently in all three purifications as the GCN5b/AP2 “core complex” (Table 2). To further confirm the association of GCN5b with these AP2 factors, we performed Western blots for GCN5b in AP2IX-7HA and AP2X-8HA immunoprecipitates. We also endogenously tagged AP2X-5, an AP2 factor that was not seen in the GCN5b IPs, to serve as a control. As shown in Fig. 6B, GCN5b was detected in a Western blot of HA-immunoprecipitated AP2IX-7HA and AP2X-8HA, but not AP2X-5HA. Collectively, these results demonstrate specific interactions between GCN5b, AP2IX-7, and AP2X-8. Whether these three proteins interact directly, or through contact with other proteins in the complex, remains to be elucidated.
The GCN5b complex associates with components of the TFIID transcription complex (TAF1/TAF250 and TAF5), affirming its role in facilitating transcriptional activation. Intriguingly, a surprising number of proteins associated with pre-mRNA splicing were also identified in the GCN5b complex (Table 2). In keeping with prior observations that Toxoplasma lacks homologues of most proteins found in the GCN5 complexes in other species [20], the majority of proteins in the GCN5b complex are novel interacting partners, with four being hypothetical proteins with unknown function (Table 2).
Twelve of the 20 subunits compromising the GCN5b core complex were detected as acetylated in a previous study (shaded gray in Table 2). Up to four proteins in the GCN5b complex, including GCN5b itself as well as TAF1/TAF250 [34], contain bromodomains that recognize acetylated lysine residues [35] (Supplemental dataset S2). The high degree of acetylated subunits and the presence of multiple bromodomain modules in the complex supports the idea that they participate in intracomplex protein-protein interactions through the binding of acetylated lysines [36].
The objective of this study was to gain a better understanding of the role played by the second GCN5 KAT in Toxoplasma parasites through biochemical purification of associating proteins, ChIP-chip analyses, and the generation of mutants. Our findings reveal that GCN5b interacts with a large number of novel proteins and is enriched at genes involved in transcription, translation, and metabolism. Consistent with previous failures to knockout GCN5b, inducible expression of a catalytically dead version acted like a dominant-negative mutant and displayed replicative arrest, supporting that GCN5b is essential in tachyzoites.
The GCN5b interactome is unique and includes components possessing plant-like AP2 DNA-binding domains, thereby providing a probable mechanism by which the complex can be recruited to target promoters. A high-throughput yeast two-hybrid approach previously identified the single Plasmodium falciparum GCN5 to be the most interconnected protein in the parasite integrating chromatin modification, transcriptional regulation, mRNA stability, and ubiquitination [23]. The discovery of AP2 factors in the GCN5b interactome prompted us to re-examine the PfGCN5 data, since this analysis was done prior to the identification of AP2 domains. PfGCN5 interacts with two predicted AP2 factors, PF3D7_1007700 and PF3D7_0802100, but they do not have similar DNA-binding domains or other conserved domains that might suggest orthology to the TgAP2s interacting with GCN5b. Interestingly, no other PfGCN5-interacting proteins cross-reference to the GCN5b interacting proteins. This may be due to the different techniques that were used (yeast two-hybrid using only the N-terminal extension of PfGCN5 as bait versus biochemical purification of full-length GCN5b), or could indicate that GCN5 complexes between apicomplexan species have significantly diverged. In support of this, the lengthy N-terminal extensions between PfGCN5 and GCN5b share no obvious sequence homology. Another possibility is that the PfGCN5 complex may be more analogous to that of GCN5a, whose interactome has yet to be resolved.
It is well-established that histone acetylation complexes generally aggregate at gene promoters, but the considerable proportion of GCN5b located within gene bodies is not without precedent. Govind et al. reported that yeast GCN5 plays a role in transcriptional elongation by promoting histone eviction [28]. GCN5 was also found to be predominantly localized to coding regions of highly transcribed genes in fission yeast, where it interplays with an HDAC to modulate H3K14-Ac levels and transcriptional elongation [30]. Interestingly, an Spt6 homologue was purified with the GCN5b complex, a protein that has been implicated in transcription elongation through binding of RNA polymerase II [37]. In other species, GCN5 has also been shown to play a role in co-transcriptional splicing [38]. The disproportionate number of pre-mRNA splicing components that we identified in the GCN5b complex might be suggestive of additional roles for GCN5b in splicing. Although it was not preferentially associated with introns versus exons, GCN5b preferentially associated with genes containing introns in our ChIP-chip analysis, providing further evidence that GCN5b macromolecular complexes may be involved in the modulating splicing. Recently, a novel Toxoplasma G1 cell cycle mutant was found to map to an RRM protein that interacts with the splicesome [39]. Further studies are required to delineate the interaction of GCN5b with the splicesome and whether the dominant negative effects of ddHAGCN5b(E703G) affect Toxoplasma splicing.
It is probable that GCN5b forms multiple complexes and contributes to an assortment of cell biological functions, as seen in other species [40]. The GCN5b-interacting proteins described in this report were isolated from intracellular, replicating tachyzoites. It is possible that GCN5b partners with different components in extracellular tachyzoites or when parasites are subjected to different stress conditions. While our data clearly shows that histone acetylation is decreased in the dominant-negative clone, we cannot conclude that the arrest in parasite replication is solely due to dysregulation of gene expression. Complicating matters is the recent observation that lysine acetylation is widespread on hundreds of non-histone proteins, many of which reside in the parasite nucleus [36]. It is conceivable that GCN5b has non-histone substrates and decreased efficiency in acetylation of those substrates contributes to the replication arrest in the dominant-negative parasites.
Toxoplasma is unique as a lower eukaryote to possess a pair of GCN5 KATs. Studies to date suggest that these two GCN5 KATs have non-redundant functions in tachyzoites. GCN5b was not sufficient to compensate for a lack of GCN5a, which is required for adequate responses to alkaline stress [5]. Similarly, GCN5a is not able to compensate when the function of GCN5b is attenuated through expression of a dominant-negative version. It has been previously reported that inhibition of parasite histone modifying enzymes is deleterious to protozoan pathogens [41]. Our findings suggest that pharmacological inhibition of GCN5b or disruption of the GCN5b complex may be novel avenues for therapy against toxoplasmosis.
All Toxoplasma lines (RH strain) were propagated in monolayers of human foreskin fibroblasts (HFFs) in Dulbecco modified Eagle's medium (DMEM) supplemented with 1% heat-inactivated fetal bovine serum (Gibco/Invitrogen). Cultures were maintained in a humidified, 37°C incubator with 5% CO2. To isolate parasites for experiments, intracellular tachyzoites were harvested through syringe passage of infected host cells followed by filtration through a 3 micron filter [42]. Where designated, Shield-1 (CheminPharma), dissolved in ethanol, was added to culture medium. For some experiments, KDAC inhibitors were added to the culture: TSA (Sigma #T8552) or apicidin (Calbiochem #178276). Plasmids were introduced into Toxoplasma via electroporation, subjected to drug selection (20 µM chloramphenicol or 1 µM pyrimethamine), and cloned by limiting dilution as previously described [42]. Parasite replication assays were performed as previously described [5], [43].
Immunofluorescence assays (IFA) were performed as previously described [21]. Briefly, HFF monolayers grown on coverslips were inoculated with the designated parasite line, sometimes containing Shield-1 or EtOH vehicle. After removal of culture medium, infected HFFs were fixed in 3% paraformaldehyde for 10 min and then were permeabilized with 0.3% Triton X-100 for 10 min. For visualization of HA-fusion proteins, rat monoclonal anti-HA primary antibody (Roche #11867423001) was applied at 1∶2,000 followed by goat anti-rat Alexa Fluor 488 secondary antibody at 1∶2,000 (Invitrogen #A-11006). Nuclei were co-stained with 4′,6-diamidino-2-phenylindole (DAPI). Samples were visualized using a Leica DMLB fluorescent microscope.
Western blots to monitor Shield-based protein stabilization were performed by resolving 50 µg parasite lysate on a 4–12% Tris-acetate polyacrylamide gradient gel (Invitrogen) and probing with 1∶2,000 rat anti-HA monoclonal antibody as the primary antibody (Roche #11867423001). Analysis of histones and tubulin used the following primary antibodies: rabbit polyclonal anti-H3 antibody (Abcam #ab1791, 1∶2,000), rabbit polyclonal antibodies against acetyl H3K9 (Millipore #06-942, 1∶2,000), acetyl H3K14 (Millipore #06-911, 1∶2,000), acetyl H3K18 (Abcam #ab1191, 1∶2,000), and rabbit polyclonal antibody against Toxoplasma β-tubulin (kindly provided by Dr. David Sibley, 1∶5,000). Anti-rat or anti-rabbit antibodies conjugated with horseradish peroxidase (GE Healthcare) were used as secondary antibodies at 1∶5,000. The blots were visualized using Chemiluminescence Western Blot Substrate (Pierce).
The GCN5b complex was purified from RHΔhxgprt parasites stably transfected to ectopically express an Nt HAmyc-tagged form of full-length GCN5b driven by the Toxoplasma tubulin (TUB1) promoter (described in [21]). Two experiments were performed following an initial co-IP done by Dr. Ali Hakimi. AP2 factor complexes were independently purified from parasites expressing AP2X-8 and AP2IX-7 endogenously tagged with HA (see above). Parental RHΔhxgprt (for GCN5b line) and RHΔku80 (for AP2 lines) were run in parallel as negative controls. Large-scale tachyzoite cultures were grown in monolayers of HFF cells at 37°C for 42 hours post-infection. Prior to egress, culture medium was removed and the cell monolayers were washed once with PBS, scraped into cold PBS and then collected by centrifugation at 4°C for 10 min at 700×g. The cell pellets were resuspended in 25 ml cold PBS, and sequentially passed through 20/23/25-gauge needles in a 30 ml syringe to release intracellular parasites from the host cells. To prepare parasite nuclear extracts, 3×109 parasites were incubated 5 min on ice in lysis buffer A (0.1% [v/v] NP-40, 10 mM HEPES pH 7.4, 10 mM KCl, 10% [v/v] glycerol, 20 mM sodium butyrate, plus protease inhibitors), and the nuclei were pelleted by centrifugation 6,000×g for 8 min at 4°C. The parasite nuclei were then incubated 30 min at 4°C in lysis buffer B (0.1% [v/v] NP-40, 10 mM HEPES pH 7.4, 400 mM KCl, 10% [v/v] glycerol, 20 mM sodium butyrate, plus protease inhibitors) with rotation, and subjected to five freeze-thaw cycles followed by vortexing for 1 min at 4°C before freezing. The nuclear extracts were clarified by centrifugation at 12,000×g for 30 min at 4°C. The mixture of the clarified nuclear extracts (1 part) with lysis buffer A (2 parts) was used for co-immunoprecipitation. Nuclear extracts were incubated with mouse monoclonal anti-HA-tag magnetic beads (μMACS Anti-HA Microbeads; Miltenyi Biotec) overnight at 4°C with rotation. After the beads were washed 4 times with cold wash buffer 1 (150 mM NaCl, 1% NP-40, 0.5% sodium deoxycholate, 0.1% SDS, 50 mM Tris-HCl pH 8.0) and once with wash buffer 2 (20 mM Tris-HCl pH 7.5) by using μ Column (pre-washed with buffer containing 0.1% [v/v] NP-40, 10 mM HEPES pH 7.4, 150 mM KCl, plus protease inhibitors) in the magnetic field of the separator, the bound proteins were eluted from the magnetic beads by applying 50 µl of Laemmli's sample buffer (pre-heated to 95°C) to the column. Eluted proteins were separated by SDS-PAGE (Any kD™ precast polyacrylamide gel; Bio-Rad), and stained with Coomassie blue (GelCode Blue Stain Reagent; Pierce). The entire length of each sample lane was systematically cut into 24 slices and the gel slices were maintained in MilliQ water until trypsin digestion.
Proteins from a Coomassie-stained gel were reduced then alkylated with TCEP and iodoacetamide prior to digestion with trypsin. Trypsin (Sequencing grade, Promega) digestion was carried out for 1 hour at 50°C using 10 ng/ul solution in 25 mM ammonium bicarbonate/0.1%ProteaseMax (Promega). The resulting digest was then diluted with 2%Acetonitrile/2%TFA prior to LC-MS/MS analysis. Nanospray LC-MS/MS was performed on a LTQ linear ion trap mass spectrometer (LTQ, Thermo, San Jose, CA) interfaced with a Rapid Separation LC3000 system (Dionex Corporation, Sunnyvale, CA). Thirty-five µL of the sample was loaded on an Acclaim PepMap C18 Nanotrap column (5 µm, 100 Å,/100 µm i.d. ×2 cm) from the autosampler with a 50 µl sample loop with the loading buffer (2% Acetonitrile/water+0.1% trifluoroacetic acid) at a flow rate of 8 µl/min. After 15 minutes, the trap column was switched in line with the Acclaim PepMap RSLC C18 column (2 µm, 100 Å, 75 µm i.d. ×25 cm) (Dionex Corp). The peptides are eluted with gradient separation using mobile phase A (2% Acetonitrile/water +0.1% formic acid) and mobile phase B (80% acetonitrile/water+0.1% formic acid). Solvent B was increased from 2 to 35% over 40 min, increased to 90% over a 5-min period and held at 90% for 10 min at a flow rate of 300 nL/min. The 10 most intense ions with charge state from +2 to +4 determined from an initial survey scan from 300–1600 m/z, were selected for fragmentation (MS/MS). MS/MS was performed using an isolation width of 2 m/z; normalized collision energy of 35%; activation time of 30 ms and a minimum signal intensity of 10,000 counts. The dynamic exclusion option is enabled. Once a certain ion is selected once for MS/MS in 7 sec, this ion is excluded from being selected again for a period of 15 sec.
Mgf files were created from the raw LTQ mass spectrometer LC-MS/MS data using Proteome Discoverer 1.2 (ThermoScientific). The created mgf files were used to search the Toxo_Human Combined database [45] using the in-house Mascot Protein Search engine (Matrix Science) with the following parameters: trypsin 2 missed cleavages; fixed modification of carbamidomethylation (Cys); variable modifications of deamidation (Asn and Gln), pyro-glu (Glu and Gln) and oxidation (Met); monoisotopic masses; peptide mass tolerance of 2 Da; product ion mass tolerance of 0.6 Da. The final list of identified proteins was generated by Scaffold 3.5.1 (Proteome Software) with following filters: 99% minimum protein probability, minimum number peptides of 2 and 95% peptide probability. All searches were performed against a decoy database and yielded no hits (i.e. a FDR of 0%). For final presentation of data, proteins appearing in the negative control parental lines and human proteins were treated as non-specific contaminants.
Parasites from AP2IX-7HA, AP2X-5HA, AP2X-8HA, and parental RHΔku80 lines were harvested in lysis buffer (150 mM NaCl, 50 mM TrisCl pH 7.4, 0.1% NP-40) with 1× protein inhibitor cocktail (Sigma) and 1 mM PMSF. The lysates were then sonicated and centrifuged to remove the insoluble fraction. Immunoprecipitations were performed using anti-HA high affinity matrix (Roche) and 300 µg total parasite protein. After overnight incubation at 4°C, the beads were washed 3× in lysis buffer and treated at 95°C for 10 minutes to elute proteins. Eluted proteins were resolved by SDS-PAGE and analyzed by Western blot with HA (1∶2,000), GCN5b (1∶500), or β-tubulin (1∶1,000) antibodies.
ChIP was performed as described [26] with some modifications. Briefly, intracellular tachyzoites grown in HFF cell monolayers for 42 hours were cross-linked for 10 min with 1% formaldehyde in PBS and quenched with 125 mM glycine for 5 min at room temperature. The cell monolayers were washed with PBS, scraped into PBS and then collected by centrifugation at 4°C for 10 min at 700×g. The cell pellets were resuspended in cold PBS, and sequentially passed through 20/23/25-gauge needles in a syringe to release intracellular parasites from the host cells. The parasites were then centrifuged at 4°C for 15 min at 700×g, resuspended in lysis buffer (50 mM HEPES, pH 7.5, 150 mM NaCl, 1% NP-40, 0.1% SDS, 0.1% sodium deoxycholate, 1 mM EDTA, plus protease inhibitors), and the chromatin was sheared by sonication yielding DNA fragments of 500–1,000 bp. The chromatin was clarified by centrifugation at 12,000×g for 10 min at 4°C; 10% of the clarified chromatin was saved as the input sample and the remaining 90% was used for immunoprecipitation.
Immunoprecipitations were performed with mouse monoclonal anti-HA-tag magnetic beads (μMACS Anti-HA Microbeads; Miltenyi Biotec) overnight at 4°C with rotation, washed extensively, and the GCN5b chromatin was eluted with 1% SDS in TE buffer. Both input and GCN5b chromatin were reverse cross-linked by incubation overnight at 65°C and purified using the Qiagen MinElute PCR purification kit. Purified DNA was amplified by using GenomePlex Complete Whole Genome Amplification kit (WGA2; Sigma) and amplified DNA was further purified using the Qiagen MinElute PCR purification kit.
The genome-wide chip was designed using the Nimblegen isothermal protocol design based upon Release 4.1 of the Toxoplasma ME49 genome with a total of 732,672 toxo-specific probes with average spacing of probes every 86 bp using an estimated genome size of 63 Mb. Chip design is available under GEO identifier GPL15563 and GPL15564 and the data is supplied in supplemental dataset S1. H3K4me3, H3K9ac, H3K4me1, H3K9me2, and CenH3 data are also accessible at www.toxodb.org. Hybridization to arrays was performed using standard Nimblegen protocols at Nimblegen or in the Albert Einstein College of Medicine Epigenomics Facility as described in [26] with an input DNA and experimental DNA hybridization performed simultaneously for each experiment on the same chip. ChIP-chip analysis was performed three times. Microarrays were scanned once with excitation at 635 nm for the immunoprecipitated DNA and 532 nm for total genomic DNA. The ratio of probe intensities is calculated and then log-base 2 transformed. After this, the bi-weight mean is calculated for all of the probe ratios. This mean is then subtracted from each of the probe ratios in order to scale the data.
GCN5b localization within the genome was determined using NimbleScan's peak calling algorithm. The algorithm employs a sliding window to look for consecutive probes with high ratios. The threshold for high ratio is percentage of the theoretical maximum ratio value (mean plus six standard deviations). The threshold steps down from 90% to 15% to calculate the significance of the peak. The false discovery rate (FDR) was estimated by randomly permuting probe ratio values. Positions of peaks with a false-discovery rate (FDR)<0.05 within each chromosome were compared between biological replicates using custom Perl scripts.
Reproducibility of GCN5b peaks was judged using the “makeVennDiagram” function from the ChIPpeakAnno package of the Bioconductor project [46]. We first selected those peaks that have an FDR below 0.05. We then identified peaks from different replicates that overlap by 50 nucleotides or more. Finally, we used the hypergeometric distribution to calculate the significance of the overlap. The probability of finding the number of overlapping peaks by chance was calculated as: 1 vs 2 7.33 E-105; 2 vs 3 2.56 E-19; 1 vs 3 0.0005. Association between GCN5b and specific genes was established by using the “findOverlappingPeaks” function from the ChIPpeakAnno package of the Bioconductor project. We used the gene annotations from ToxoDB, release 6.1, ME49 strain. Gene associations for each replicate were determined independently.
We used two different randomization strategies to test the reproducibility of our GCN5b ChIP. First, we reassigned the starting position of each of our significant peaks (FDR< = 0.05) to a random probe from the microarray while keeping the peak widths the same. We used the hypergeometric distribution to calculate the significance of the overlap with the randomly selected peaks in each replicate. The probability of chance overlap in each of these cases was close to one. In a second test, we randomly selected an equal number of peaks from each replicate, regardless of FDR. These peaks were reanalyzed using the hypergeometric distribution to calculate the significance of the overlap with the peaks in each replicate. These randomizations also resulted in a probability of close to one that the overlap was due to chance. Both randomization tests were performed 1,000 times.
For examination of GCN5b peaks with introns or genes with introns, gene models and annotations from www.toxodb.org V6 were used. Significance of overlaps was determined by hypergeometric test for intron-containing genes (5,989) versus intronless genes (2,083). To further test this, we selected random peaks from each experiment rather than only the significant peaks (FDR<0.5%). We repeated the same test with the randomized peaks. Using the same methods, we also tested for an association between GCN5b binding with introns or exons. We also tested for an association with active promoters (as defined by dual marking with H3K9ac and H3K4me3 [26]), active coding regions (defined by the H3K4me1 mark [26]), or centromeres [27] and measured the distribution of GCN5b peaks from the inferred transcription start site using data from Yamagishi et al. [47].
1.0 µg of total RNA purified from intracellular parasites was transcribed into cDNA using Omniscript reverse transcriptase with oligo-dT primers according to the manufacturer's protocol (Qiagen). qRT-PCR was performed in 25 µl volume reactions containing SYBR Green PCR Master Mix (Applied Biosystems), 0.5 mM of each forward and reverse primer (Supplemental Table S1), and 1.0 µl of a 1∶10 dilution of cDNA. Target genes were amplified using the 7500 Real-time PCR system and analyzed with relative quantification software (7500 software v2.0.1, Applied Biosystems). The ratio of mRNA levels in Shield-treated parasites versus EtOH-treated parasites was calculated using Toxoplasma β-tubulin as an internal control for normalization (GCN5b was not detected at β–tubulin in any of the three ChIP-chip experiments). Reactions were performed in triplicate and Student's t-test was applied to RT-PCR data.
|
10.1371/journal.pgen.1007874 | Allele-specific RNA imaging shows that allelic imbalances can arise in tissues through transcriptional bursting | Extensive cell-to-cell variation exists even among putatively identical cells, and there is great interest in understanding how the properties of transcription relate to this heterogeneity. Differential expression from the two gene copies in diploid cells could potentially contribute, yet our ability to measure from which gene copy individual RNAs originated remains limited, particularly in the context of tissues. Here, we demonstrate quantitative, single molecule allele-specific RNA FISH adapted for use on tissue sections, allowing us to determine the chromosome of origin of individual RNA molecules in formaldehyde-fixed tissues. We used this method to visualize the allele-specific expression of Xist and multiple autosomal genes in mouse kidney. By combining these data with mathematical modeling, we evaluated models for allele-specific heterogeneity, in particular demonstrating that apparent expression from only one of the alleles in single cells can arise as a consequence of low-level mRNA abundance and transcriptional bursting.
| In mammals, most cells of the body contain two genetic datasets: one from the mother and one from the father, and—in theory—these two sets of information could contribute equally to produce the molecules in a given cell. In practice, however, this is not always the case, which can have dramatic implications for many traits, including visible features (such as fur color) and even disease outcomes. However, it remains difficult to measure the parental origin of individual molecules in a given cell and thus to assess what processes lead to an imbalance of the maternal and paternal contribution. We adapted a microscopy technique—called allele-specific single molecule RNA FISH—that uses a combination of fluorescent tags to specifically label one type of molecule, RNA, depending on its origin, for use in mouse kidney sections. Focusing on RNAs that were previously reported to show imbalance, we performed measurements and combined these with mathematical modeling to quantify imbalance in tissues and explain how these can arise. We found that we could recapitulate the observed imbalances using models that only take into account the random processes that produce RNA, without the need to invoke special regulatory mechanisms to create unequal contributions.
| Gene expression in genetically identical individual cells often deviates from that of the cell population average [1], which in mammals can impact cell fate and development [2–5], response to environmental stimuli [6–9] and disease [10–13]. Over the past few years, it has emerged that at least some of this variability arises due to random fluctuations in the biochemical processes that underlie transcription and translation. In the case of transcription, a primary source of fluctuations is so-called transcriptional bursting, where a gene alternates between an active state, during which RNA is produced, and an inactive state, where no RNA is transcribed. Because both the time of onset of these bursts and the amount of RNA produced in a single burst are random, this process can lead to cell-to-cell variability [14–16].
An additional nuance to the effects of bursting on cellular variability is that diploid mammalian cells carry two sets of chromosomes (one from each parent), which means that they also have two copies of each individual gene. It is typically assumed that for most genes both copies, called alleles, are capable of being expressed, thus providing protection through redundancy if one of them is mutated [17,18]. Recent studies, however, which made use of crosses between distantly related mouse strains and high-throughput sequencing, uncovered that there can be extensive differences in the relative expression levels of the two alleles [19–23]. Additional work then showed that even if transcripts from both alleles are detected at the population level, there may be substantial variation in the degree of allelic imbalance in single cells. For example, while for most genes individual cells express RNA from both alleles, for other genes the population can be a mixture of cells expressing RNA from only one or the other allele. This latter expression pattern has been termed random monoallelic expression, and certainly, genes with such an expression profile exist: most X-linked genes are expressed only from one X chromosome due to random X-chromosome inactivation [24–26], and a similar pattern has also been shown for some autosomal genes, such as olfactory receptors or antigen receptors [27–29]. Understanding random monoallelic expression is of particular interest given that quantitative cell-to-cell differences or spatial heterogeneity in allele-specific gene expression have the potential to modify phenotypic outcome if the two alleles harbor different functional variants, as has been described for both X-linked (eg. [30–34]) and autosomal traits [35,36].
Beyond these prototypic cases, it has been proposed that many more autosomal genes may be subject to random monoallelic expression [37–43], but some key properties of this extended class sets them apart from the more established examples. Similarly to the initial group, these genes were classified as displaying random monoallelic expression because cells with only transcripts from one or the other gene copy were observed, with both types of cells present in the same experiment. In addition, some of these genes maintain their monoallelic expression status over multiple passages in clonal cell lines [40,43], so a specific, heritable mechanism could limit transcription to only one allele, as is the case for X chromosome inactivation. However, thus far such a mechanism remains elusive [40,41]. Moreover, unlike previously established cases, many genes in this extended set are not expressed exclusively monoallelically, and typically a subset of cells or clones with transcripts from both alleles can also be detected [37–41]. This suggests that if a specific mechanism does exist to regulate monoallelic expression, it is limited to only a subset of cells.
To resolve this question of mechanism, Reinius et al proposed an elegant scenario (so-called dynamic random monoallelic expression), whereby the many genes with random monoallelic expression may arise not by differential cell- and allele-specific regulation, but instead monoallelic expression may arise by chance [43,44]. In this scenario infrequent transcriptional bursting would lead to cells that contain only RNA from one of the two gene copies. This observed monoallelic expression of mRNA would be temporary and the allelic state of a cell could change over time. The authors confirmed this model in clonal cell lines [43], but whether the same is true in tissues is still an open question. Different groups have deployed single-cell transcriptomics to determine the degree of cell-to-cell allelic imbalance [42,43,45,46], but technical limitations inherent to low abundance RNA quantification, as well as parameter choice can impact the interpretation of allele-specific sequencing data [47,48]. Thus, it has been hypothesised that the level of monoallelic expression, especially at single-cell level, could be an overestimate [45,49]. This absence of precise quantitative data has made it difficult to definitively answer if random monoallelic expression observed in vivo requires a dedicated mechanism or if it could arise as a consequence of transcriptional bursting.
In this study, we adapted a previously described single-molecule RNA fluorescent in situ hybridization technique that is sensitive to single-nucleotide differences between RNAs for the analysis of transcripts in snap-frozen, cryosectioned tissues from different mouse strains and their hybrids [50–52]. This allowed us to determine the allelic origin of individual RNAs in single cells, while preserving both their spatial context and their in vivo expression levels. We used this method to measure allele-specific expression of multiple autosomal genes and of Xist, a gene for which it is well-documented that individual cells randomly and exclusively transcribe either the maternal or the paternal copy. Quantitative analysis of the data enabled us to answer whether the autosomal genes we investigated were expressed from one or both gene copies in single cells in tissue. While we observed monoallelic expression in some cells, mathematical modeling showed that this pattern was compatible with random transcriptional events (including transcriptional bursting) from the two alleles producing low levels of RNA, rather than an explicit mechanism governing random monoallelic expression.
Our goal in tissues was to quantitatively measure the amount of cell-to-cell variability in transcript abundance from either the maternal or paternal allele of a gene to determine the degree of imbalance between transcripts arising from the two alleles. To make these measurements, we modified a protocol previously developed by our group [50] that enables the detection of single nucleotide variants (SNVs) on individual RNA molecules in situ in cultured cells (Methods, Fig 1A). We applied this method to mouse kidney tissues from C57BL/6J (BL6) and JF1/Ms (JF1) mice and their F1 progeny, since these two strains harbor a large number of genomic differences, allowing us to target most genes with multiple SNV-specific probes.
To demonstrate our ability to detect expression from the BL6 vs. JF1 allele using our technique in tissue, we first examined the expression of Xist, a prototypical example of random monoallelic expression [25,53]. In female cells Xist is transcribed exclusively from the (randomly chosen) inactivated X chromosome, is expressed at high levels, and contains a large number of SNVs between the two substrains, making it an ideal test case for our method. We collected kidney tissues from mice at day 4 of postnatal development, snap-froze them in liquid nitrogen or on aluminium blocks in dry ice and cryosectioned them at 5 μm thickness. We then applied our in situ hybridization protocol using allele-specific probes. As expected, we observed the appropriate sex-specific expression pattern: no fluorescent signal in male tissues (S1 Fig), whereas in female tissues, our guide probes clearly labelled nuclear Xist foci, which colocalized with signal from the strain-specific probes. For SNV-targeting probes, controls in homozygous tissues confirmed that only the appropriate probes bound, with little binding from the probes targeting the other strain (S1 Fig). Whereas in heterozygous samples the two strain-specific probes each labelled a subset of the nuclei in the anticipated mutually exclusive pattern (Fig 1B). To further analyse the data, we developed a pipeline to computationally identify Xist RNA foci, and automatically classify them as BL6 or JF1 within an entire scanned kidney section (Methods, Fig 1B). This algorithm identified tens of thousands of Xist foci per section (mean 22k, min 8.2k, max 30k), and—in agreement with our manual inspection of the data—predominantly identified Xist foci of the correct identity in the homozygous samples (S1 Table), while the overall population ratio of BL6:JF1 foci ranged from 45:55 to 65:35 in heterozygous samples (Fig 1C, S2A Fig).
Having verified that we could correctly measure the allelic origin of clusters of Xist RNA accumulated on the X chromosome in mouse tissues, we next ascertained whether the method would also work for monodisperse spots corresponding to single RNA molecules, which is how most mRNAs appear in the cell. We considered this challenging, because single, punctate RNA spots would be both smaller and considerably dimmer compared to Xist RNA, which accumulates multiple copies on the inactivated X chromosome. Thus, to test our protocol for use on single RNA molecules, we designed probes for 8 autosomal genes that contained at least one polymorphism in BL6 versus either JF1 or the C7 strain (which carries both copies of chromosome 7 from the M. musculus castaneus strain in a BL6 background). These genes were selected to represent genes with (Aebp1, Churc1, Lyplal1) and without (Aqp11, Mpp5, Podxl, Prcp, Stard5) putative random monoallelic expression [38–41] and with different expression levels, patterns and chromosomal locations. Some of the chosen genes were also linked to specific kidney disease phenotypes (Aqp11 [54], Mpp5 [55,56], Prcp [57]) (Methods, S4 Table). As expected, these genes were typically expressed at much lower levels than Xist and punctate individual mRNA spots could not be as readily observed in low magnification scans. We therefore combined whole tissue scans in a single plane at low magnification with random sampling of the tissue at higher magnification, where we imaged the entirety of the section. This approach allowed us to identify individual mRNA spots within the context of the whole tissue section in the 60x scan, while the additional data collected from the 100x z-stacks facilitated our ability to precisely determine colocalization between the guide probes and the strain-specific probes (Fig 1D). Colocalization between the strain-specific signal and the mRNA probes allowed us to determine from which allele a given mRNA originated. Accordingly, this could be used as a key readout for SNV-specific single molecule RNA FISH, and we characterized the quality of the experiments using colocalization rate (i.e. what percentage of the guide spots colocalized with allele-specific spots). When we assayed the overall colocalization rates for these autosomal genes in kidney sections, we found that 4 out of 8 genes had mean colocalization rates >50% (S4 Fig, S4 Table), which is comparable to colocalization rates previously observed in cultured cells [50], showing that we were able to perform quantitative SNV-specific single molecule RNA FISH directly in tissue. We also observed an apparent trend between colocalization rate and the number of SNV probes, where genes with fewer SNV probes had lower colocalization rates than those with more SNV probes. We tested a series of parameters that could affect probe binding (base composition, GC content, probe secondary structure and folding energy), but found no parameter that differentiated between the probes with high and low colocalization rates (S4 Fig). However, it should be noted that SNV probes were only tested as full sets (i.e. all SNV probes for a given gene were tested together), and we therefore do not know the binding behaviour of individual probes.
Collectively, these results showed that we could directly visualise and assign strain-specific identities to both focally localised RNA (Xist) and single molecule RNA spots in the context of whole tissue sections. This motivated us to ask if we could directly quantify cell-to-cell heterogeneity in the allelic origin of RNAs in tissue.
To determine how the chromosomal origin of RNAs contributes to cell-to-cell heterogeneity in tissue, we focused on two different questions. For Xist, we investigated the spatial clustering of cells based on their X inactivation choice, i.e. to what extent cells expressing Xist from either the JF1 vs BL6 X chromosome intermix. For the autosomal genes, we quantified their allelic imbalance in single cells, i.e. whether individual cells expressed RNAs from the two chromosomes at different ratios.
In the case of Xist, we measured spatial clustering of allele-specific expression because each cell is randomly and fully committed to expressing RNA from only the BL6 or the JF1 chromosome. Thus, allelic imbalance in tissue is not due to quantitative expression differences between the two alleles in single cells. Instead, it can arise either as a consequence of overall skewing of X chromosome inactivation rates or due to uneven spatial distribution of cells with a given inactivated X chromosome. Such spatial partitioning can arise through the local expansion of cells in which X chromosome inactivation “choice” has already been fixed, resulting in extended patches of cells carrying the same inactive X chromosome [58–61]. Because we had observed fairly balanced expression from the BL6 and JF1 chromosome in our initial analysis of heterozygous samples (see previous section, Fig 1C), we next determined whether cells expressing Xist from either the BL6 or JF1 allele segregated spatially. Although visual inspection of the sections revealed no extended regions where cells expressed Xist foci with the same strain-specific identity, computational modelling could potentially reveal a more precise view of spatial patterning. We therefore developed a metric that characterized the distribution of cells expressing BL6 Xist RNA (Methods and S2 Fig) and then compared this to either completely randomized BL6 and JF1 assignments or randomizations where we introduced different sized clusters of cells. We found that in all tissues BL6 cells were less evenly distributed than the random assignments (S2C Fig), suggesting that cells cluster together more than expected in a completely random scenario. Our subsequent comparison with the clustered assignments further supported and refined this interpretation: it showed that our data was most similar to simulations with smaller cluster sizes. The closest matching seed size was different, depending on what scale we assessed spatial partitioning, but centered around a cluster size of 2–4 cells (mean cluster size: 3.2, standard deviation 0.7) (S2 Table, Fig 2A and S2D Fig). Thus, cells expressing Xist from the same chromosome clustered together in small patches in tissue sections, resulting in a spatially fairly mixed population of cells and showed no evidence of extended patches with the same allele-specific expression.
For the autosomal genes we wanted to know how much the allelic imbalance differed between cells. Visual inspection of our data indicated a range of allelic ratios, including BL6 monoallelic, JF1 monoallelic, as well as biallelic mRNA expression. To quantify this, we focused on three autosomal genes (Aebp1, Lyplal1, Mpp5) where >50% of guide spots colocalized with signal from SNV probes (we excluded Podxl, because the high expression levels of Podxl mRNA in podocytes precluded separating individual RNA spots (S7 Fig)). We selected this cutoff based on testing in cultured mouse fibroblasts, where we had found that in cells with >50% colocalization rate, our method consistently identified the correct allele (Methods, S5 Fig). First, we considered that the observed chromosomal origin (BL6 vs. JF1) could either be due to true biological variability or to technical error (as seen when we detected RNA from the “incorrect” strain in homozygous tissues). To distinguish these, we determined the BL6 and JF1 signal for these genes in kidney tissue from both BL6 and JF1 homozygous mice, as well as in tissue from reciprocal heterozygous crosses. For all three genes, we counted only a few mRNAs in the majority of cells (mean number of RNA spot counts per cell: 3.5 for Aebp1, 3.2 for Lyplal1 and 2.6 for Mpp5). These RNA counts measured by RNA FISH were higher than those observed via single-cell RNA sequencing of the kidney [62]. The mean RNA count by single-cell sequencing ranged between 1.0–1.78 UMI/cell for Aebp1, 1.0–1.2 UMI/cell for Lyplal and 1.0–1.63 UMI/cell for Mpp5, depending on the cell type, considering only those cells where transcripts for these genes were detected (S3B Fig). Given the higher detection rate of our method, it may be particularly useful for assessing allele-specific expression for these lowly expressed genes.
We predominantly detected the correct allele in the homozygous kidney samples, both in bulk and at the single cell level (Fig 2B, 2D and 2F). In heterozygous samples we observed a more balanced presence of both BL6 and JF1 mRNAs, with the reciprocal crosses showing similar results (heterozygous data in Fig 2B, 2D, 2F and 2G, and S5 Table). These results indicated that our technical error (false positive rate) was less than the biological variability and that we could use our method to measure quantitative single-cell differences. Moreover, when we compared the BL6 allelic ratios in homozygous and heterozygous cells with >2 mRNA, we found some heterozygous cells with allelic ratios similar to those of the homozygous samples, but also a subset of cells with an allelic ratio that was intermediate to that of homozygous cells (Fig 2G). We obtained similar results when we switched the fluorescent dyes conjugated to the SNV detection probes used to label the BL6 and JF1 mRNA (Methods, S6 Fig). The observation that single cells had transcripts from both the BL6 and JF1 allele were particularly intriguing for Aebp1 and Lyplal1, because these two genes had been previously identified as genes with putative random monoallelic expression in other tissues [38–41]. Still, given that we did in fact observe cells with either BL6 or JF1 monoallelic mRNA expression in heterozygous tissue, akin to the random monoallelic expression pattern, we wanted to know which transcriptional models our quantitative single-cell allelic imbalance data was compatible with, in order to explain how this expression pattern could arise.
Our results showing that individual cells could have mRNA from either one or both alleles motivated us to assess whether existing models of transcription were sufficient to explain the observed cell-to-cell variability in allelic imbalance. To evaluate these models, we first used the correlation coefficient between the BL6 and JF1 mRNA expression in our population as a simple metric that captures the joint behaviour of the two alleles: a correlation coefficient of zero represents no coordinated expression between the two alleles, positive values indicate coordinated expression, while negative correlation could be indicative of anti-correlation or repulsion between the two alleles. The correlation between BL6 and JF1 mRNA counts were 0.41 for Aebp1, 0.29 for Lyplal1 and 0.35 for Mpp5 suggesting somewhat coordinated expression. To better understand how these values compare to the expectations from different models of transcription, we initially considered two extreme cases: an “all-or-none” scenario in which every cell has transcripts exclusively from one or the other gene copy, and a “coin flip” scenario in which the allelic origin of every individual transcript is essentially indistinguishable from random coin flipping (Fig 3A). The former scenario could suggest the existence of regulatory mechanisms that limit transcription to only one gene copy per cell (an extreme form of random monoallelic expression, as is the case with Xist), whereas the latter corresponds to a null model with no distinct allele-specific transcriptional regulation. We used computational modeling to simulate these two scenarios and to discriminate between them.
We first checked if our data were similar to those expected in the “all-or-none” scenario, in which the transcripts in each individual cell were either solely from one or the other gene copy. Looking in heterozygous cells, it seemed qualitatively apparent that (as noted) many cells have mRNAs from both gene copies, which was seemingly incompatible with this scenario. However, it was still formally possible that cells in reality only had transcripts exclusively from one of the gene copies and that the apparent transcripts from the other copy were technical artifacts due to false detection events. We used homozygous tissue to measure the rate at which these false detection events occur, and thereby estimated the expected false detection rate in heterozygous cells. We then computationally simulated hypothetical “all-or-nothing” heterozygous cell populations taking into account these false detection rates, and found that the RNA counts and distributions observed in such a simulated all-or-none population were still inconsistent with our data (compare Fig 3B with heterozygous data in Fig 2B, 2D and 2F). Next, we calculated the correlation of the BL6/JF1 counts for the “all-or-nothing” simulations, which showed much lower correlation values than we had observed in the real data (Fig 3C). Thus, the two alleles in our simulated data were uncorrelated or anti-correlated, as would be expected in the case of random monoallelic expression, and were unlike the positively correlated BL6 and JF1 mRNA we had observed our measurements. In addition, the probability of observing the strain-specific mRNA counts we measured versus simulated cell populations was also different (S8 Fig). Collectively, these results indicate that in the kidney, none of the three genes we interrogated displayed “all-or-none” expression.
In our alternative scenario, it is possible that the two copies of the gene transcribe RNA independently and each random transcription event produces just a single RNA, thus leading to the “coin-flipping” model in which most cells would have RNAs from both alleles in them, but with some statistical noise about this population average (Fig 3A). We modeled the outcome of such a scenario and found that the real versus simulated single-cell RNA distributions looked very similar for all three genes (compare Fig 3D with heterozygous data in Fig 2B, 2D and 2F). We also saw that the correlation between the BL6 and JF1 allele in our modeled data was similar to our measured correlation in the case of Lyplal1 and Mpp5 (Fig 3E). Statistical analysis showed that when we treated the strain-specific RNA counts per cell as a series of independent coin-flips, the probability of the observed distributions for both Lyplal1 and Mpp5 fell within the distribution of likelihoods from our simulated model (S8 Fig). Together, these results demonstrated that the allelic imbalance observed for Lyplal1 and Mpp5 was compatible with a simple coin-flipping null model of transcription from the two alleles.
In the case of Mpp5, where we had found false detection rates to differ most depending on which fluorescent dyes were coupled to the allele-specific probes, we collected data from heterozygous tissues swapping the dyes used on the BL6 and the JF1 probes and repeated our analysis and simulations. We obtained similar results, showing that a model for all-or-none expression did not recapitulate the allelic counts observed in our data, whereas the coin-flip model did (S6B and S6C Fig). We also tested whether the detection efficiency of ~50% influenced our conclusions about the “all-or-none” vs. “coin flip” scenarios by repeating our simulations assuming 100% detection efficiency and then randomly downsampling to the measured detection rate. The results were essentially indistinguishable from those obtained without downsampling, further verifying that our results were not due to the technical noise introduced by the low detection efficiency (S5C and S5D Fig).
All analyses of our simulations therefore supported the conclusion that allelic imbalances for Lyplal1 and Mpp5 were compatible with a simple coin-flipping null model of transcription. Aebp1, however, showed higher levels of imbalance per cell than could be explained by this model. This was not clearly visible on a scatterplot of the simulated RNA counts per cell (Fig 3D), but the measured BL6 and JF1 counts were less correlated than those in the simulated dataset (Fig 3E). Yet this gene did not exhibit the all-or-none behavior either. We therefore considered a third, intermediate scenario motivated by the phenomenon of transcriptional bursting [14–16]. Transcriptional bursts refer to the fact that most mammalian genes are transcribed in short pulses during which multiple transcripts are synthesized, interspersed between periods during which the gene remains inactive. When there are two copies of a gene, each bursting independently [63], the expected result would be that some cells may have more transcripts from one of the copies than expected by the coin flipping model above; bursting would be akin to getting several heads or tails in a row every time one flipped a coin.
To test whether transcriptional bursting could explain the observed data, we first wanted to confirm that Aebp1 was indeed transcribed in a burst-like fashion. To verify this, we measured Aebp1 transcriptional activity directly in kidney cells by using intronic probes (Fig 3F), which, owing to the extremely short half-life of introns, detect almost exclusively nascent transcripts at the site of transcription [63]. This showed that 19% of cells with Aebp1 mRNA were also actively transcribing Aebp1, and that these transcription sites contained more than 1 RNA based on their fluorescence intensity relative to cytoplasmic RNA spots (average 1.6x higher fluorescence intensity in 3 independent experiments, S9 Fig). This data also showed that the majority of actively transcribing cells had only one Aebp1 transcription site, although a small subset (6 out of 55 (11%) cells with Aebp1 transcription sites) showed simultaneous expression from both alleles. This corroborated that cells indeed produced Aebp1 in transcriptional bursts and that it was possible for individual cells to transcribe RNA from both alleles simultaneously.
Next, we turned to simulations to assess the RNA distributions that we would expect in a scenario where the two alleles transcribed RNA independently from each other and in bursts. Initially, we estimated the expected burst size (average number of RNAs that were transcribed together in a single burst) and burst frequency for the two alleles based on the observed RNA counts for each allele independently and used these parameters as inputs for our model (see methods for details). BL6 and JF1 RNA counts simulated this way closely matched our measurements, which was also reflected by the likelihood of the real data falling within that of the simulated data for the two alleles separately (Fig 3G). We then wanted to see whether the degree to which there was allelic imbalance in single cells could be explained by the two alleles bursting independently; i.e., whether for per-cell RNA counts from the two alleles was more or less correlated than one would expect by chance. To simulate the null hypothesis of no interaction between alleles we randomly paired up the modeled BL6 and JF1 counts, mimicking cells that contain RNA from both alleles. When we compared this simulation to the real data we consistently observed that the modeled per cell BL6 and JF1 counts were less correlated than the real pairwise measurements (Fig 3G and S10 Fig). Thus, while in our measurements cells with high BL6 expression typically also expressed JF1 at higher levels (compare heterozygous data in Figs 2B and 3G top panel), in the modeled data BL6 and JF1 counts showed little correlation. This was also true when we incorporated false detection events in our model to account for possible incorrect allelic assignment, as we had done in the “all-or-none” model (S11 Fig). Together, our transcription site measurements and simulations showed that the observed allele-specific single-cell RNA counts for Aebp1 were compatible with transcriptional bursting of the two gene copies individually and that expression from the two alleles was correlated.
Thus, for Aebp1, Lyplal1 and Mpp5 the observed monoallelic expression in some single cells can likely be explained by low levels of transcription occurring randomly from the two gene copies without having to invoke a special mechanism that limits expression to one of the alleles.
There has been great interest in recent years to precisely measure expression from the two alleles of a gene in diploid cells, ideally directly in tissue and at the single-cell level. The RNA fluorescence in situ hybridization method described here is a step in this direction: by visualizing endogenous SNVs it enables the assignment of single RNA molecules to their allele of origin in single cells and in the context of whole tissue sections. We now provide quantitative information about cell-to-cell heterogeneity with single transcript resolution, which is an extension of our previous work, where we used this method for a more qualitative assessments (i.e. presence-absence) of parental origin of mRNA in tissue [52]. When we applied our method to autosomal genes we observed that individual cells in heterozygous tissue spanned the entire range from all RNAs originating from the BL6 chromosome through various more mixed populations to all RNAs originating from the JF1 chromosome. These observed allelic imbalances were not due to the parental origin of the gene copies, because reciprocal crosses (BL6xJF1 vs JF1xBL6) showed similar results. We therefore asked what model could explain the observed single-cell allelic imbalance pattern and combined our data with computational simulations to address this question. We found that the observed allelic distributions could be recapitulated by a model where transcription occurred randomly from the two alleles, perhaps with moderate transcriptional bursting (e.g. in the case of Aebp1). Thus, we did not have to invoke a special mechanism that restricted expression to only one allele to explain the presence of cells with either BL6 or JF1 monoallelic expression status.
Our results suggest that cells with RNA expression from only one of the alleles occur due to the low levels of expression and thus the limited number of random sampling events from the two gene copies. Because transcription is a dynamic process this state is likely transient so that while a cell may have mRNA from only one allele at a particular time, it may gain mRNA from the other allele a short time later if another transcriptional event occurs. For Mpp5, this conclusion is in line with the fact that the reported haploinsufficient phenotype is thought to be caused by overall reduced dosage [55,56] rather than by a special subpopulation of cells with monoallelic expression. More generally, this scenario links the observation of transcriptional bursting with that of random monoallelic expression, as put forward by Sandberg et al [43,44], and explains why we observed the co-occurrence of cells with one and two transcription sites in the same population. In the case of Aebp1 and Lyplal1, which had previously been identified to display random monoallelic expression, our data suggest that no additional mechanism is needed to explain the presence of cells with monoallelic expression in kidney tissue. However, we cannot extrapolate these results to other tissues and mouse strains, given that random monoallelic expression is tissue specific. Thus, we cannot conclude whether mechanisms for allele-specific regulation or for maintenance of monoallelic status are present in the mouse strains and cell types used in the studies that originally identified genes with random monoallelic expression [37–41]. Interestingly, though, those studies also observed that many genes with monoallelic expression in one clonal cell line are expressed biallelically in others and that the expression of a given gene is often lower in cell lines with monoallelic expression than in those with biallelic expression [38–40], which is consistent with our results.
In addition to the data on Aebp1, Mpp5 and Lyplal1, we also demonstrated our ability to distinguish Xist expression from the BL6 vs JF1 chromosome and to assess the spatial relationship of cells expressing different parental alleles. We detected a spatially fairly mixed population in the kidney, where cells expressing Xist from the same chromosome clustered together in small patches in transverse sections. This is contrary to other tissues, such as intestinal crypts or the skin [58,59,61,64], and suggests either a larger number of kidney precursor cells or extensive cell migration during development. Because our method only provides a snapshot in time we cannot easily distinguish between these scenarios (lineage vs. migration), especially given that the kidney is a complex organ, composed of cells originating from different embryonic lineages that undergo extensive migration during development even after birth [65,66]. Regardless of the developmental mechanism, however, our data indicate that there is likely no major spatial segregation due to X chromosome inactivation in this tissue, and in the case of mutations, any phenotypic effects would be fairly evenly distributed.
Through the examples detailed above we have shown how our method can be used to directly quantify cell-to-cell differences that arise due to differential expression from the two alleles in diploid cells. Our approach overcomes multiple limitations imposed by previous methods: First, because this method enables sensitive SNV-specific detection of even single mRNA molecules it provides more information than RNA FISH measurements that rely solely on quantifying the number of transcription sites in individual cells [36,37,40,41,67]. Second, although our approach tests only single genes in a given experiment and thus has much lower throughput than single-cell sequencing-based methods, it relies on direct detection of transcripts and is therefore not subject to subsampling and dropout, which complicate the interpretation of sequencing-based cell-to-cell variability results [44,46,68–70]. Finally, by making use of pre-existing endogenous SNVs it eliminates the need for genetic manipulation, for example to label the gene of interest with fluorescent tags, as has been done to measure X chromosome inactivation choice [61,71] or to monitor the transcription of autosomal genes from the two chromosomes [9,72]. Moreover, because the breeding history for classical inbred mice has lead to extended regions of shared ancestry (and shared SNVs) between different strains [73,74], a probeset developed for one strain can often easily be adapted for another strain. For example, while we measured Xist expression from the BL6 and JF1 allele, the probes were designed so that they should distinguish equally well between the 129-strains and CAST/EiJ, which are also commonly used to study strain-specific expression.
In addition, while our quantitative analysis focused solely on genes with a relatively high number of SNVs and high mean colocalization rates (>50%), it should be noted that we did not systematically explore the relationship between SNVs and colocalization rates, and also that lower colocalization can be sufficient to address specific questions, as was demonstrated in a recent single cell in situ analysis of A-to-I RNA editing [75]. It is therefore likely that depending on the scientific question, less stringent cutoffs can be applied to colocalization rates and/or the number of SNVs required. In addition to the number of SNVs/allele-specific probes, the applicability of our method also depends on the expression level of the gene of interest. Based on the genes tested here, our method is applicable for genes with low-to-moderate expression levels, corresponding to approximately 2–25 FPKM in bulk sequencing data (see Methods). The lower boundary for this estimate is set by Churc1, a gene that is consistently expressed at low levels in all cell types of the kidney (0.1–1 UMI/cell, ~2 FPKM). It is technically possible to determine colocalization events at this mRNA density, though a large number of cells needs to be screened for robust quantification. The upper boundary is based on the observation that a high density of transcripts makes it impossible to precisely assign colocalization. In our experiments only Podxl in podocytes showed such high expression levels. Based on the relatively low proportion of podocytes in the kidney we determined that this matched an expression level of >800FPKM in these cells (based on bulk sequencing data) and an average count of 6.9 UMI/cell in single-cell sequencing data. Importantly, given the higher detection efficiency of RNA per cell, allele-specific single-molecule RNA FISH provides a valuable complementary tool to single-cell transcriptomics for low-to-moderate expression level genes, and will likely be useful in confirming findings made by sequencing.
In conclusion, we demonstrated how quantitative measurement of allele-specific expression in tissue could be used to directly determine the level of allelic imbalance in single cells. By combining these measurements with modeling, we showed that random monoallelic expression could arise in vivo by chance alone. Beyond this application, our methods could have a number of additional uses. Similar analyses could be performed in other tissues and, for example, could enable the evaluation of genetic variants directly in the tissue believed to be affected if there are genic SNVs in linkage with those variants or to study mutations thought to lead to haploinsufficiency. Furthermore, with single cell resolution, our method allows for the interrogation of particular cellular subtypes within a tissue. In concert with recent genome-wide association studies in single cells [76,77], this technique provides a useful tool for quantitative assessment of allele-specific genetic effects.
C57BL/6J and JF1/Ms founder mice were obtained from Jackson Laboratories. All mouse work was conducted in accordance with the University of Pennsylvania Institutional Animal Care and Use Committee. For tissue collection we used either homozygous C57BL/6J or JF1/Ms pups, or F1 heterozygotes from both C57BL/6J x JF1/Ms or JF1/Ms x C57BL/6J crosses. We dissected pups at postnatal day 4 using standard techniques, and mounted tissues in Tissue-Plus O.C.T. compound (Fisher Healthcare), flash-froze them in liquid nitrogen or on an aluminum block in dry ice, and then stored tissues at −80°C. We determined sex of the animals by visual inspection and verified this by SRY-specific PCR on DNA extracted from a tail sample, collected during dissection. Tissues were cryosectioned at 5 μm using a Leica CM1950 cryostat. We adhered tissue samples to positively charged Colorfrost plus slides (Fisher Scientific), washed slides in PBS, fixed them in 4% formaldehyde for 10 min at room temperature, then washed again in PBS two times. Fixed slides were stored in 70% ethanol at 4°C.
We shortlisted genes that had previously been identified in studies of random monoallelic expression, genes with known function and/or phenotypes in kidney and genes with documented expression in different cell types in the kidney. We then filtered shortlisted genes for expression levels of >20 TPM in publicly available bulk kidney sequencing data [78], accessed via Expression Atlas (https://www.ebi.ac.uk/gxa/). We later established that this corresponded to expression levels of >10FPKM in bulk and 0.1–1 UMI/cell (in the tissue(s) with highest expression; S3 Fig and S9 Table) using additional bulk [79] and single-cell [62] transcriptome data, respectively. The only gene with lower expression was Churc1 (TPM <10, FPKM ~2.2), for which we were nevertheless able to robustly detect transcripts using the guide probes (in agreement with a count of 0.1-1UMI/cell in all cell types of the kidney), and identify colocalization events above random (pixel-shifted) controls. The only gene with higher expression was Podxl with a UMI of 6.9 per cell in podocytes.
To identify exonic SNVs between the C57BL/6J and JF1/Ms strains we used the NIG Mouse Genome Database (http://molossinus.lab.nig.ac.jp/msmdb/index.jsp) [80]. For Aebp1 and Lyplal1 we confirmed these SNVs through PCR amplification and sequencing of exonic sequences of genomic JF1/Ms DNA. All guide probes and the Aebp1 intron probe set were designed using the Stellaris probe designer (Biosearch Technologies), SNV-specific probes were designed as specified in Levesque et al. [50] and mask oligonucleotides were selected to leave a 7-11bp overhang (toehold) sequence (all probe sequences available in S6 Table). Guide probes were purchased labeled with Cal fluor 610 (Biosearch Technologies), while SNV-specific probes and intron probes were ordered with an amine group on the 3′ end. For these latter probes we pooled the oligonucleotides for each probe set and coupled them to either NHS-Cy3 or NHS-Cy5 (GE Healthcare) for the allele-specific probesets, or NHS-Atto488 or NHS-Atto700 (Atto-Tec) for the intronic probes. We purified dye-coupled probes by high-performance liquid chromatography. Mask oligonucleotides were used unlabelled.
DNA was extracted from tail biopsies using a quick-lyse protocol: 100μl of Solution A (25mM NaOH and 0.2mM EDTA) were added to the tissue and kept at 95°C for 30 min, before adding an equal volume of Solution B (40mM Tris, pH = 8). Samples were then spun at 6000 rpm for 10 min and 100μl of the top layer was transferred to a fresh tube. 1μl of this solution was used as template for PCR. To verify the presence of reported SNVs in Aebp1 and Lyplal1, we designed primers for the exonic segments of these genes (primer sequences available in S7 Table), and PCR-amplified genomic DNA using AmpliTaq Gold (ThermoFisher) with buffer II and 0.25mM MgCl2 according to the manufacturer's instructions. PCR amplicons were purified with ExoSAP-IT (ThermoFisher) and submitted for sequencing to the University of Pennsylvania DNA sequencing facility. For sex-specific genotyping of pups we used Sry-specific primers (S7 Table), since this gene is located on the Y chromosome and thus amplicons can only be detected in male tissues. PCR was performed as for sequencing, and the presence-absence of a product was revealed on a gel.
Allele-specific RNA fluorescence in situ hybridization works by first detecting the mRNA of interest (regardless of the allele of origin) using conventional single molecule RNA FISH probes labelled in one color (guide probes), and then colocalizing this signal with probes that discriminate specific single-nucleotide differences based on a “toehold probe” strategy and which are labelled in colors unique to the two different alleles. In this way, mRNAs are essentially “tagged” as being either from one or the other parental chromosome. In cultured cells, this approach can successfully distinguish RNA variants that contain just one single nucleotide variant (SNV) and thus can only be targeted with a single variant-specific probe [50,51], but the decreased signal-to-noise ratio makes reliable detection of single probes more difficult in tissue [81]. We therefore opted to work with C57BL/6J (BL6) and JF1/Ms (JF1) mice, which belong to two different Mus musculus subspecies (domesticus and molossinus, respectively) [80,82]. Due to their distant relationship, JF1 mice harbor a large number (>50,000) of SNVs in genic regions compared to BL6 [80], allowing us to target most genes with multiple SNV-specific probes.
For each gene of interest we first prepared a probe mix, containing a guide probe set (labelled with Cal fluor 610), the two allele-specific probe sets (labelled in Cy3 and Cy5, respectively) and a set of mask oligos (unlabelled, in 1.5x excess of the allele-specific probes) in hybridization buffer (10% dextran sulfate, 2× SSC, 10% formamide). For detection of nascent Aebp1 RNA we also included intronic probes labelled either with Atto488 or Atto700, and to verify the integrity of RNA in male tissues stained for Xist we also included Gapdh probes (labelled with Atto488). To stain the samples, we first washed the slides with tissues sections in 2x SSC, then incubated them in 8% SDS for 2 minat room temperature, washed again in 2x SSC and finally added the hybridization buffer with probes. Slides were covered with coverslips and left to hybridize overnight in a humidified chamber (ibidi) at 37°C. The next morning we performed two 30 min washes in wash buffer (2× SSC, 10% formamide), the second one including DAPI to stain nuclei. To label cell membranes (to clearly identify single cells) the first wash was sometimes substituted with a 15 min incubation in wash buffer containing wheat germ agglutinin coupled with Alexa488 (LifeTech) and a 15 min regular wash. After the final wash, slides were rinsed twice with 2x SSC and once with antifade buffer (10 mM Tris (pH 8.0), 2× SSC, 1% w/v glucose). Finally, slides were mounted for imaging in antifade buffer with catalase and glucose oxidase [83] to prevent photobleaching.
We imaged all samples on a Nikon Ti-E inverted fluorescence microscope using either a 60x or a 100× Plan-Apo objective and a cooled CCD camera (Andor iKon 934). For whole-tissue scans we imaged at 60x and used Metamorph imaging software (Scan Slide application) to acquire a tiled grid of images. We used the Nikon Perfect Focus System to ensure that the images remained in focus over the imaging area. For 100× imaging, we acquired z-stacks (0.3 μm spacing between stacks) of stained cells in six different fluorescence channels using filter sets for DAPI, Atto 488, Cy3, Calfluor 610, Cy5, and Atto 700. The filter sets we used were 31000v2 (Chroma), 41028 (Chroma), SP102v1 (Chroma), 17 SP104v2 (Chroma), and SP105 (Chroma) for DAPI, Atto 488, Cy3, Cy5, and Atto 700, respectively. A custom filter set was used for CalFluor610 (Omega).
For Xist image analysis we worked with whole tissue scans, where we had collected data for Cal fluor 610 (Xist guide probes), Cy3 and Cy5 (BL6 and JF1 probes, respectively) and DAPI (nuclei). To visualize scans, we used the “Grid/Collection stitching” feature available in Fiji [84] to assemble tiles. To identify Xist RNA and assign them an allelic identity we developed a custom pipeline in MATLAB. First, we reconstructed the scan taking into account the tile order provided in a supplementary file. Then, we used the data from the guide channel to detect Xist foci, regardless of allelic identity: we performed background subtraction, removed small objects and smoothened boundaries by border clearing and morphological opening, and then used LoG filtering to sharpen objects, binarized the observed signals and created connected components. Visual inspection of these connected components showed that they largely corresponded to Xist foci, but some areas with high background signal were also being detected as connected components. We therefore applied a number of filters (minimum fluorescence intensities for all RNA FISH channels, minimum cutoff for solidity, maximum area for connected components) to yield the final segmentation. Each obtained spot was then parametrized as the ratio of the signal intensity (background subtracted and normalized to the mean intensity of the scan) of the two SNV probe channels and we applied k-means clustering (2 means) to yield a critical angle above which we assigned spots JF1 identity, and below which we assigned BL6 identity. To verify the quality of these assignments, we designed a graphic user interface to manually annotate Xist foci and their allelic identity. We typically annotated 10 or more randomly selected tiles and the results of this quality control step are shown in S3 Table. On average ~90% (mean 90.9%, standard deviation 5%) of Xist foci were correctly detected, while the remaining 10% of identified spots were areas of high background intensity that had been miscategorized as Xist foci. When Xist spots were correctly identified, typically more than 90% were assigned the correct allelic identity (mean 94.4%, standard deviation 5%).
To assess spatial patterns of Xist allelic choice we then used the positional and identity information from our automatic assignments, and developed a metric for spatial heterogeneity. First, we tiled images into regular rectangles of equal size (i.e. 16 tiles all 1/16 of the full scan size). For each rectangle, we calculated the fraction of cells expressing Xist from the BL6 allele. Next, we obtained the variance of these BL6 cell fraction values across all rectangles of a given size. This protocol was repeated for different sizes of rectangles ranging from 16 to 256 rectangles spanning the entire tissue section. We also calculated a baseline for spatial heterogeneity of random allelic choice by repeating this analysis on 1000 random permutations of the data for each sample generated by MATLAB’s randperm. We performed a similar analysis to determine the minimal cluster size of Xist foci with identical allelic identity, but instead of random permutations we generated simulations, where kidney sections were randomly seeded with clusters of a fixed size (ranging from 1 to 10) while keeping the allelic ratio the same as for the measured data. For each seed size we generated 500 simulations. To obtain a likely minimal cluster size for cells with identical X chromosome inactivation we selected the seed size whose variance deviated least from the variance observed for the real data. We repeated this process for each subdivision size and determined the mean across all subdivision sizes.
For analysis of single molecule RNA spots we used a combination of 60x whole tissue scans in DAPI and Cal fluor 610 to determine the overall structure of the tissue and collecting z-stacks at 100x resolution of 5–10 individual positions within that tissue to identify individual mRNA molecules and characterize their allelic identity. To determine allelic identity we first segmented and thresholded images using a custom MATLAB software suite (downloadable at https://bitbucket.org/arjunrajlaboratory/rajlabimagetools/wiki/Home, changeset: d278b7d0012282ecb318fde3bebbe3beaba62032). To quantify colocalization rates we first determined the ideal colocalization radius for each gene. To do so, we segmented extended areas of the tissue (typically containing 10–50 cells). To ascertain subpixel-resolution spot locations the software then fitted each spot to a two-dimensional Gaussian profile specifically on the z plane on which the spot occured. Next, colocalization between guide spots and allele-specific spots was determined in two stages. In the first stage, we searched for the nearest-neighbor allele-specific probe for each guide spots within a 2.5-pixel (360-nm) window and ascertain the median displacement vector field, which was subsequently used to correct for chromatic aberrations. After this correction, we tested a range of different radii (r = 0.1 to 2.5 pixel) for each gene to calculate colocalization rates for the real data, as well for pixel-shifted data, where we took our images and shifted the guide channel by adding 2*r pixels to the X and Y coordinates. This pixelshifted data was used to test random colocalization due to spurious allele-specific spots. For each gene we then visually inspected colocalization rates for real and pixel-shifted data at the different radii and determined a radius where both the colocalization rate for the real data and the difference between the real and the pixel-shifted data was maximal. The selected colocalization radii for each gene are included in S4 Table. To obtain allele-specific data for single cells we then repeated the colocalization analysis, but segmentation of cells was done by drawing a boundary around nonoverlapping individual cells using brightfield or wheat germ agglutinin signal, and colocalization was determined using only the previously determined ideal colocalization rate.
Transcription site analysis was performed using a custom MATLAB software suite (downloadable at https://bitbucket.org/arjunrajlaboratory/rajlabimagetools/wiki/Home). For this, we segmented cells, thresholded RNA FISH signal and identified transcription sites for Aebp1 by co-localization of spots in the intron and exon channel. Relative fluorescence intensities of transcription sites vs cytoplasmic RNAs were determined based on the fluorescence intensity of the guide probes using custom scripts written with R packages dplyr and ggplot2.
To determine whether any biophysical properties could differentiate between allele-specific probes that had high vs low colocalization rates, we compiled a table containing the the following parameters (S8 Table): probe name, probe sequence, colocalization rate (the colocalization rate determined for an entire probeset was applied to each individual probe), number of predicted secondary structures and folding energies. The latter two parameters were extracted by running sequences on the mfold web server [85] for DNA probes, with Na concentration set to 0.3M. Frequency of individual nucleotides, dAT, dGC, purines and pyrimidines was determined through analysis of the probe sequences.
Since different fluorescent dyes have different detection sensitivity, we tested how this impacted our findings by performing “dye-swap” experiments in which we use the same probes but swap the fluorophore moieties used to label the probes for the BL6 and JF1 alleles. This enables us to measure the extent to which the labels on the probes affected probe binding efficiency and what the consequences were for our results and interpretations. In these experiments we probed Aebp1, Lyplal1 and Mpp5 expression in BL6 and JF1 homozygous tissues with allele-specific probes labelled either with Cy3 for BL6 and Cy5 for JF1, or the reverse (S6A Fig). The unique detection rate (i.e. the rate at which we could assign spots either a BL6 or a JF1 identity) differed by 7% or less between the dye combinations for all three genes, with the exception of Mpp5 in BL6 homozygous tissue. Similarly, the false detection rate, which includes both the error from incorrect probe binding and the differential detection sensitivity, deviated by less than 4% between the dye combinations for all three genes, with the exception of Mpp5 in BL6 homozygous tissue. For Mpp5 in BL6 homozygous tissue we observed that the BL6 allele was detectable at higher efficiency when labelled with Cy5 (66%) than when labelled with Cy3 (50%). We therefore also collected dye-swapped data for Mpp5 in heterozygous tissues (S6B Fig) and performed the “all-or-none” and “coin flip” simulations on both dye combinations (S6C Fig and Fig 3B–3D). In addition, we used the empirically determined false detection rate for each dye combination and gene for all of our analysis and modelling of allelic imbalance in heterozygous tissue to avoid artefacts due to variable sensitivity of detection for different dyes.
Given that for all three genes that we studied in more detail we could determine BL6 or JF1 identity for only approximately 50% of RNAs, we measured how colocalization rate impacted our ability to measure single-cell allelic imbalance. This was difficult to do in tissue, because the low number of RNAs per cell provided a limited dynamic range to probe this effect (i.e. if 1 out of 2 RNAs can not be assigned an identity, this results in 50% detection efficiency). We therefore collected extensive data for Aebp1 in primary fibroblasts, where Aebp1 is expressed at higher levels (10s to 100s of RNAs per cell) to determine the general effect of the colocalization rate on allelic ratios. In homozygous fibroblasts we found that there was an inverse relationship between colocalization rate and false detection rate, so that higher colocalization rate corresponded with a lower false detection rate and a higher rate of calling the correct allele (S5A Fig). Most likely this was caused by low colocalization rates being indicative of other technical issues that complicate determining colocalization rate (such as high background fluorescence). We observed this effect regardless of the dye combination used, but also noted that above a colocalization rate of ~0.5 (the overall colocalization rate in tissue) our ability to detect the correct allele was consistently high (above 75%). Next, we used 10 heterozygous fibroblast cells with the highest colocalization rates. The cells all had colocalization rates >70% and randomly downsampling RNAs to 65, 60, 55, 50 or 45% colocalization rate (1000 random sampling per condition) showed the the mean allelic ratio for each downsampling was similar to the original measured ratio, but the variability around that mean increased with lower colocalization rates. The biggest change in the allelic ratio that we observed was ~0.2 (S5B Fig). Next, we also tested how detection efficiency influenced our conclusions about the “all-or-none” or “coin flip” scenario in tissue. For this, we repeated both simulations using the total RNA count for each cell (i.e. including unclassified RNAs and RNAs that colocalized with both allele-specific probes). We then randomly downsampled the RNA in each cell to the number of RNAs that had been assigned a unique BL6 or JF1 identity in the original measurement. Each simulation was performed 5000 times. We found that the log likelihoods and correlations from these downsampled simulations were essentially indistinguishable from the original simulations performed without downsampling (S5C Fig).
To quantify cell-to-cell variability of allelic state in single tissue cells, we extracted colocalization data from our image analysis pipeline, and used this for further analysis. Using this data, we first compiled a quality control table for each experiment (S5 Table) and excluded those where colocalization rates were <40% (4 out of 35 experiments). For all remaining data we combined replicates from the same genotype, and in the case of heterozygous data, combined results from C57BL/6J x JF1/Ms and JF1/Ms x C57BL/6J tissues. We then processed and visualised single-cell results using custom scripts written with R packages dplyr and ggplot2.
To determine how the observed data compared to random monoallelic expression (all-or-nothing scenario) or binomially distributed (coin-flip scenario) allelic calls we simulated those scenarios through modelling. For the binomial distribution we considered a null model wherein all heterozygous cells share the same allelic ratio, which was determined to be the overall allelic ratio observed at the population level. Then, for an experiment with overall estimated C57BL/6J allelic ratio equal to pBL6 (above), we let nBL6j be the number of transcripts with C57BL/6J identity detected in cell j and nJF1j be the number the number transcripts with JF1/Ms identity detected in cell j. Under the null model, nBL6j was drawn from a binomial with (nBL6j + nJF1j) draws and probability pBL6. We simulated single-cell label counts for cells by drawing from these conditional null distributions for each cell 10,000 times. We then compared the negative log-likelihood of the observed data with the distribution of negative log-likelihoods of each simulation iteration.
To simulate random monoallelic expression each cell was assigned either a BL6 or a JF1 identity, based either on the majority of RNAs in a cell, or based on random assignment if both alleles had the same count. We then designated all RNAs in a given cell to the same allelic identity (eg a cell that originally contained 4 BL6 RNAs and 2 JF1 RNAs would be assigned a BL6 identity with 6 BL6 RNAs). Next, we randomly added “technical noise” 10,000 times, by changing some RNAs in the population to the opposite identity, based on the false positive rates measured in the original BL6 and JF1 homozygous populations. These steps were performed while keeping the final overall RNA assignments in the population the same as the original heterozygous population. For the 10,000 simulations we then calculated negative log likelihoods similarly as we did for the binomial distributions, assuming two separate null models for the BL6 and JF1 populations, whose parameters were determined by the original homozygous population.
To assess if the measured mRNA distributions were compatible with a transcriptional bursting scenario we used the negative binomial distribution to simulate expected mRNA counts [86]. First, we determined the burst size and frequency of the BL6 and JF1 alleles separately, by using the moments method to determine r and p parameters of the negative binomial distribution based on the mean and variance of our measurements (where p = mean/variance and r = mean^2/(variance-mean)), from which we obtained the burst size and frequency using: burst_size = (1-p)/p and burst_frequency = r. We then generated 10,000 RNA counts for the two alleles separately by drawing from a distribution with the r and p parameters we had calculated. We visualized the obtained mRNA counts for both alleles individually using a randomly selected simulation, and also calculated the negative log-likelihood distribution of the 10,000 simulated datasets. Next, we randomly paired the data for the two alleles to generate “cells” with RNA counts from both alleles and calculated the correlation between the BL6 and JF1 counts in each of these modeled cells. In addition to using the negative binomial parameters that we had calculated from our data, we also tested a series of additional burst sizes (from 0.5 to 5 RNAs per burst) and repeated the entire analysis, which showed that our findings were consistent across a range of burst values (see S10 Fig). Finally, to generate a model which included BL6-JF1 correlations that arise due to false assignments, we used the randomly paired data and changed some RNAs in the population to the opposite identity based on the false positive rates measured in the original BL6 and JF1 homozygous populations (one round of reassignments for each simulation). Following reassignment we again calculated the correlation between the BL6 and JF1 counts.
Raw and processed data, as well as scripts for all analyses presented in this paper, including all data extraction, processing, and graphing steps are freely accessible from the Open Science Framework (URL https://osf.io/sbjcw/?view_only=09393298e00e4b8c9d0dd1b24b0318d9).
Our image analysis software (changeset: d278b7d0012282ecb318fde3bebbe3beaba62032) is available here:
https://bitbucket.org/arjunrajlaboratory/rajlabimagetools/wiki/Home
All animal studies were approved by the Institutional Animal Care and Use Committee at the University of Pennsylvania (IACUC protocol 805433). The animals used in this study were treated humanely and with regard for alleviation of suffering.
|
10.1371/journal.pgen.1000178 | Stability and Dynamics of Polycomb Target Sites in Drosophila Development | Polycomb-group (PcG) and Trithorax-group proteins together form a maintenance machinery that is responsible for stable heritable states of gene activity. While the best-studied target genes are the Hox genes of the Antennapedia and Bithorax complexes, a large number of key developmental genes are also Polycomb (Pc) targets, indicating a widespread role for this maintenance machinery in cell fate determination. We have studied the linkage between the binding of PcG proteins and the developmental regulation of gene expression using whole-genome mapping to identify sites bound by the PcG proteins, Pc and Pleiohomeotic (Pho), in the Drosophila embryo and in a more restricted tissue, the imaginal discs of the third thoracic segment. Our data provide support for the idea that Pho is a general component of the maintenance machinery, since the majority of Pc targets are also associated with Pho binding. We find, in general, considerable developmental stability of Pc and Pho binding at target genes and observe that Pc/Pho binding can be associated with both expressed and inactive genes. In particular, at the Hox complexes, both active and inactive genes have significant Pc and Pho binding. However, in comparison to inactive genes, the active Hox genes show reduced and altered binding profiles. During development, Pc target genes are not simply constantly associated with Pc/Pho binding, and we identify sets of genes with clear differential binding between embryo and imaginal disc. Using existing datasets, we show that for specific fate-determining genes of the haemocyte lineage, the active state is characterised by lack of Pc binding. Overall, our analysis suggests a dynamic relationship between Pc/Pho binding and gene transcription. Pc/Pho binding does not preclude transcription, but levels of Pc/Pho binding change during development, and loss of Pc/Pho binding can be associated with both stable gene activity and inactivity.
| Cells make fate decisions as they progressively differentiate into specific cell types during development. The stability of these decisions is important and is achieved, in part, by changes to the chromatin that packages DNA in the nucleus. A key set of protein complexes that together constitute the Polycomb-group/Trithorax-group (PcG/TrxG) machinery is involved in chromatin modification and is known to operate at a large number of genes involved in developmental decisions. The PcG proteins establish stable gene repression, whereas the TrxG counteract the PcG to enable gene activation. How this PcG/TrxG balance works is not understood. By mapping PcG protein binding to chromatin in vivo, we show, in general, a relatively constant association of PcG protein at target genes during development. However, we also find changes in binding at specific genes. While some of these changes are consistent with a loss of PcG proteins associated with gene expression, we also find examples where PcG proteins are present at active genes and not present at inactive genes. Our analysis supports the idea that simply the presence of PcG proteins at a target gene does not necessarily result in gene repression and suggests a more dynamic balance between PcG protein binding and gene expression.
| As the cells of the embryo progress along developmental pathways they make fate decisions, becoming committed to particular lineages and ultimately to a specific differentiated cell state. Although cell fate decisions may be triggered by transient signals, the resultant cell states are generally stable and are maintained through time and cell division. A long-standing paradigm for understanding the mechanisms underlying the stability of cell fate decisions has been the maintenance of Hox gene expression through gene silencing by Polycomb-group (PcG) genes in Drosophila (reviewed in [1]). Hox gene expression domains, initiated in the early embryo through active transcriptional regulation by the transiently-expressed products of the segmentation genes, are thereafter maintained throughout the rest of development and adult life by the maintenance machinery of the PcG and Trithorax-group (TrxG) genes. The products of the PcG genes build the Polycomb Repressive Complexes (PRC1 and PRC2) that are required for gene silencing, whereas the TrxG genes are required for the maintenance of gene activation (reviewed in [2]). In this paradigm, the balance between gene repression and activation is set once and thereafter stably remembered.
A more dynamic view of the role of PcG silencing has recently been emerging, largely from work with embryonic stem cells, where several PcG genes have been shown to be required for both embryonic and adult stem cell maintenance (reviewed in [3]). Genome-wide analysis of the targets of PRC1 and PRC2 complex components reveals that a large number of genes with roles in cell fate decisions and cell differentiation are bound by PcG gene products in stem cells [4],[5]. Many of these genes are repressed by PcG proteins since loss or down-regulation of PcG genes results in their derepression. Upon stem cell differentiation many repressed genes become activated and some concomitantly lose binding of PcG complexes. In stem cells many developmental genes exhibit a “poised” bivalent chromatin organisation, carrying both repressive and active chromatin modifications [6]–[8]. The repressive H3K27me3 histone modification, dependent on the PRC2 complex, is lost from many genes on differentiation. Thus PcG silencing appears to maintain the stem cell state via repression of cytodifferentiation genes; this repression is not permanent and can be lifted upon receipt of differentiation signals.
When the human embryonic teratocarcinoma cell line NT2/D1 is induced to undergo neural differentiation by exposure to retinoic acid, two different scenarios are observed for PcG regulation of target genes [9]. For PcG target genes activated during neuronal differentiation (e.g. the neuronal transcription factor ZIC1 and the neurofilament light chain gene, NEFL), PcG proteins are associated with these genes prior to activation but are lost upon differentiation. In contrast, for PcG target genes repressed during differentiation (e.g. the pro-neural transcription factors OLIG2 and NEUROG2), PcG proteins are already associated with these genes in undifferentiated cells, even though the genes are expressed, and the Polycomb complexes remain after differentiation when expression is switched off. Thus it appears that, at some genes, the association of Polycomb complexes with target genes can change dramatically upon differentiation, but the presence of Polycomb complexes does not always accord with transcriptional repression.
PcG target genes have been identified in Drosophila by genome-wide mapping of PcG protein binding in tissue culture cells [10],[11] and by more limited genomic mapping (across 10 Mb of Drosophila euchromatin) with different developmental stages in vivo [12]. In this latter study, examples of target genes with clear developmental changes in PcG protein association were identified, suggesting that the chromosome association profile of Polycomb complexes in Drosophila may be more dynamic than previously thought.
Here we extend these studies, presenting a genome-wide analysis of PcG proteins in Drosophila embryos and in imaginal discs from the third thoracic segment. We examine the binding profiles of Polycomb (Pc), the canonical member of the PRC1 complex, and of Pleiohomeotic (Pho) a DNA-binding protein proposed to recruit the PRC2 complex [13]. Analysis of tissue derived specifically from the third thoracic (T3) segment allows us to examine Pc and Pho association with Hox genes that are known to be either active or inactive in this segment. Comparing binding profiles between the embryo and third larval instar imaginal discs also enables us to examine the dynamics of PcG binding during development of specific tissues. Finally, we compare our in vivo developmental analysis with a previously described genome-wide analysis of Pc binding in Drosophila tissue culture cells [10] identifying further examples of differential Pc binding.
We performed genome-wide mapping of binding sites for Pc and Pho in chromatin from two in vivo sources; Drosophila embryos and imaginal discs. We studied Pc as a representative of the four core PRC1 components, Pc, Polyhomeotic, Posterior sex combs and dRing [14]. We investigated Pho since this is a sequence-specific DNA-binding protein known to be associated with several PcG Response Elements (PREs). Pho binding sites have been shown to be required for PcG-mediated silencing at these PREs [15]–[19] and although Pho is not a component of purified PRC complexes, it interacts biochemically with both PRC1 and PRC2 [13],[20],[21]. Pho co-localises with PRC1 proteins at many sites on polytene chromosomes [16] and, by ChIP (Chromatin Immunoprecipitation) analysis, it is associated with PRC1 binding sites in Hox genes [22],[23].
The 0–16 hr embryonic chromatin provides a base-line for our analysis identifying a set of in vivo targets in a mixture of developmental time-points and tissues. In contrast, the imaginal disc chromatin provides a more focussed sampling of targets within a single tissue (epidermal imaginal), at a particular developmental time (wandering third larval instar) and at a specific position along the body axis (T3 segment). At a gross level, comparison of the binding profiles, e.g. across chromosome 3R as illustrated in Figure 1, reveals considerable similarity, suggesting that Pc and Pho are generally bound at the same locations and that their binding sites appear relatively constant with little change between embryo and imaginal disc.
To analyse these data in more detail, we defined upper and lower binding thresholds for each profile, allowing us to categorise binding over each Drosophila gene as positive, intermediate or negative (see Table S1). As well as counting binding events directly over transcription units, intergenic events were separately ascribed to the nearest transcript. Validation of the ChIP-array data by ChIP followed by specific PCR confirmed that thresholds were appropriate (Figure S1).
Using conservative thresholds, we find 386 genes with Pc binding over the transcription unit. 179 (46%) of these are also associated with Pho binding, which rises to 229 (59%) when we include intergenic Pho binding, supporting the idea that Pho has a general role as a DNA-binding protein targeting the assembly of Polycomb complexes.
Interestingly, we find a substantial number of genes (212) that exhibit Pho binding in the absence of Pc (Figure 2). The majority (85%) of these Pho-only binding sites are specific to embryo chromatin. Examination of the GO classifications that are enriched in this subset reveals a markedly different profile from the set of genes that bind both Pc and Pho (Figure 2C). In particular, we find significant overrepresentation (p<0.05) of genes involved in oogenesis, mitosis and mRNA splicing. In addition, we also note that several genes whose products are involved in chromatin regulation (e.g. brahma, Polycomblike, Cp190 and su(Hw)) are associated with Pho but not Pc. These observations indicate a role for Pho in the embryo that is independent from its association with Pc.
As illustrated in Figure 2D the local profiles of Pc and Pho binding are very different. Pc is often associated with a broad binding domain extending over tens of kilobases whereas Pho binding is characterised by much narrower isolated peaks. Since the sharp Pho peaks identify relatively short regions associated with Pho binding we searched for enriched sequence motifs underlying these peaks. In addition to Pho, several other DNA-binding proteins, including GAGA factor and Zeste, are associated with PREs [24]–[29]. We were interested to see if Pho-bound sequences exhibited the canonical Pho binding motif, as well as putative binding sites for these other factors.
We learned motif dictionaries from the central areas of 150 strong Pc-associated Pho peaks using various search parameters. The searches identified more than 70 partially redundant sequence motifs and we selected 18 of these for further analysis, based on their length, information content and/or similarity to known binding sites of Pho, GAGA and Zeste (Figure 3, see also Table S2 and Dataset S1). Binding sites of additional factors known to be involved in Pc recruitment (e.g. Grainy head) could not be identified. We found Pho-type motifs comprising a core GCCAT sequence with a more or less pronounced tail of four Ts. In addition, we found a novel TGGCC motif that may be related to the Pho-type as it has a GCC core (and GCCA on the reverse strand) but which lacks the tail of Ts. We also found a frequently occurring GTT repeat and a CGCACT sequence motif. The GAGA- and Zeste-type motifs differ in that Zeste-like motifs have a pronounced “GAG”, however we recognise that this classification is somewhat arbitrary.
The selected motifs were significantly over-represented when we compared their occurrence in all Pho peaks to random sets of sequences that were not occupied by Pho (Figure 3B). While longer motifs with positions of low information content are mostly over-represented when allowing for mismatches, short sequence motifs are only over-represented when considering perfect matches or small sub-optimal bit scores. This over-representation approach also enabled us to derive informed cutoff values for further analyses; for each motif we identified the sub-optimal bit score for which the motif shows the strongest evidence for over-representation. Using these cutoff values, we determined the presence of the different motif types in all 628 embryonic Pho peaks (Figure 3C). The short motifs (clustered to the right of the diagram) are well represented in the Pho peak sequences; 85% of the sequences contain Pho_6, 63% contain GAGA_6 and 64% contain Zeste_7. Longer Pho and TGGCC motifs occur in 51% of the peaks and have an interesting antagonistic clustering to the longer GAGA and Zeste-type motifs; i.e. peaks containing longer Pho/TGGCC-type motifs (region A in Figure 3C) cluster separately from peaks containing longer GAGA/Zeste-type (region B). Overall, the general association of Pho, GAGA and Zeste binding motifs with Pho binding is consistent with the clustering of these motifs previously used to predict PREs in the Drosophila genome [29] and we add novel enriched sequences that may improve such approaches. However, we emphasise that there is considerable variability in the motif occurrence at Pho peaks as we illustrate for representative peaks in Figure S2.
We were interested to determine if there was any qualitative difference in motif composition between Pc-associated and Pho-only peaks. We compared the number of peaks containing particular motifs for both these peak classes and tested for significance using chi-square statistics. Interestingly, we find significant differences with TGGCC_7 (34.7% vs 53.6%, p<3×10−6) and Zeste_7 (31% vs 46.8%, p<8×10−5) under-represented in the Pho-peaks that are not associated with Pc. In contrast, the Pho_12a (43.9% vs 26.2%, p<3×10−6) and Pho_14b (35.2% vs 20.4%, p<3×10−5) motifs are over-represented in the Pho-only group. This observation highlights the potential importance of a long Pho motif at the Pho-only sites. There is no compositional bias for the longer GAGA- and Zeste-type motifs.
What differentiates a silenced from an active Hox gene? Although the PcG and TrxG genes have antagonistic effects on gene expression they can nevertheless be present at the same gene. PcG proteins and TrxG proteins were found to co-localise at targets on salivary gland polytene chromosomes [30],[31] and at PREs in the Bithorax Complex (BX-C) [32]. In addition, Pc-binding does not correlate with gene expression in S2 cells [33]. Recently, Papp and Mueller have analysed the binding of PcG and TrxG at the Ubx gene in active and repressed states in vivo [22]. Sampling the occupancy of these complexes at 17 sites across 115 kb of the Ubx region, encompassing the transcription unit and regulatory sequences, they found that PcG proteins of both the PRC1 and PRC2 complexes as well as Trx protein are bound to Ubx PREs in both the ON and the OFF states. Similarly, Trx and PcG were found to co-localise at binding sites in both active and inactive Hox genes in tissue-culture cells [23].
Our ChIP-array data allows a more extensive assessment of Pc and Pho occupancy across the Drosophila Hox complexes in vivo. Chromatin derived from embryos represents a mix of active and inactive states for the different Hox genes, however, for the T3 imaginal discs we can compare silenced and active Hox genes. Focusing on the BX-C: Ubx is active in the T3 discs, where it is required in the haltere disc to specify haltere in contrast to wing development, and in the T3 leg disc to specify T3 characteristics. In contrast, both abdominal-A (abd-A) and Abdominal-B (Abd-B) are silenced in T3 discs.
The Pc binding profile shows an extensive domain of binding that covers the approximately 300 kb BX-C region (Figure 4A). Characterised PREs tend to be represented as regions of relatively higher binding within the overall domain. As with other regions of the genome, the Pho binding profile is very different with much sharper, more isolated peaks several of which coincide with characterised PREs. In chromatin from the T3 imaginal discs both Pc and Pho are associated, as expected, with the silenced genes, abd-A and Abd-B. However, we also find significant association with the active Ubx gene. The T3 disc Pc profile over Ubx is similar to the embryo chromatin profile with an extensive domain and significant binding at both the bx and bxd PREs as well as a peak close to the start of Ubx transcription. The Pho profile in T3 discs also shows clear peaks at these PREs and binding close to the Ubx 5′ end. These data show clear evidence of Pc and Pho association with an active gene and confirm and extend the results of Papp and Mueller [22].
There are, however, differences between the embryo and T3 disc profiles. For example, several strong Pho binding peaks in embryo chromatin are only weakly represented in the T3 disc chromatin. In addition, there is a generally lower level of Pc and Pho binding across the active Ubx gene in comparison to the inactive abd-A and Abd-B genes. The average enrichments (log2 binding ratios) across the three transcription units in disc chromatin for Pc are: Ubx 0.33, abd-A 1.03 and Abd-B 1.01 and in the case of Pho: Ubx 0.07, abd-A 0.31 and Abd-B 0.34.
Significant Pc and Pho binding associated with an active Hox gene is also found in the Antennapedia Hox cluster (ANT-C). This cluster contains the Hox genes lab, pb, Dfd, Scr and Antp, and all these genes are associated with widespread Pc binding and distinct Pho peaks in the embryo (Figure 4B). It is striking that peaks in the Pc distribution as well as strong Pho peaks are found close to the 5′-ends of lab, pb, Dfd and Scr. The longer Antp gene is covered by a domain of Pc binding and is associated with several Pho peaks. As with Ubx, Antp is active in T3 discs since it is expressed from the labial segment posteriorly. Indeed, Antp may be a better gene than Ubx for the analysis of the active state in T3 because it is expressed in both the ectoderm and mesoderm of the T3 segment [34],[35], whereas Ubx is only active in the ectodermal imaginal cells of the T3 disc and may be silenced in the small population of mesodermal adepithelial cells [36]. Despite this difference we find a similar situation with Antp as we observe with Ubx. Although it is active, Antp is nevertheless associated with a significant domain of Pc binding, which encompasses the Antp transcription unit, as well as strong peaks of Pho binding close to the 5′ and 3′ ends of the gene. Other Pho peaks over Antp that are prominent in embryo chromatin are less apparent in T3 discs.
As shown in Figure 4 we find many more Pho binding peaks across the Hox complexes than there are characterised PREs. As all these Pho sites may not be functionally equivalent we examined the motif composition in the 36 Pho peaks in the BX-C. We find a high variability of motif counts but characterised PREs do not appear as a distinct motif-rich group (Figure S3).
While our analysis of the Hox clusters demonstrates significant Pc and Pho binding associated with both active and inactive genes, a genome-wide comparison of the embryo and T3 imaginal disc profiles shows that PcG proteins are not constitutively associated with target genes. The binding profiles of embryo and T3 disc chromatin are similar, with 65% (252/386) of the genes significantly associated with Pc in the embryo also above threshold in the disc chromatin. This rises to 81% (314/386) if we include the genes with intermediate binding in discs. When we include the data from the genome-wide S2 tissue culture study [10] we also find considerable overlap. For genes with direct binding over transcription units there is a 42% (161/386) overlap across all three data sets rising to 58% (224/386) if we include intermediate binding in discs and S2 cells. A gene-by-gene comparison is provided in Table S1.
Comparison with other previously published datasets also reveals considerable overlap. For the analysis of Pc targets by the DamID method in Kc tissue culture cells [11] we find that our set of Pc targets in the embryo (386 targets) contains 136 targets from the Kc cell data. As the genome coverage of the Kc cell analysis is approximately 60%, this extrapolates to an estimated 60% overlap. This comparison is detailed in Table S1. There is also good correspondence with the in vivo data from Negre et al. [12] where, for example, 5 out of the 7 targets they identify in the 3 Mb Adh region in embryo chromatin are also present in our set of Pc targets in the embryo. A detailed comparison is presented in Figure S4.
For examining differential Pc binding, we focus on the most comparable datasets, the two chromatin samples in our dataset and the Schwartz et al. S2 data [10], as these three datasets are genome-wide and use the same Affymetrix array platform. Despite the overall similarity comparing the embryo, T3 disc and S2 cell chromatin samples, there are clear differences in Pc occupancy for some genes indicating potential sites of differential Pc activity. To reduce the level of artificially selected differential Pc targets resulting from automatic selection of peaks in the high-throughput analysis, we visually screened the binding profiles of all differentially bound regions and restricted our further analysis to differential gene sets that show significant enrichment of GO categories. We identify three robust sets of differentially occupied genes. We find 49 genes that are bound by Pc in the embryo but are not Pc-associated in imaginal discs, 107 genes that are bound by Pc in the embryo but not in S2 cells and 119 genes that are bound by Pc in imaginal disc but not in S2 cells.
By examining the genes in these differentially-bound sets we anticipated that we might identify cell fate-determining genes for cell fate decisions in particular developmental pathways. For example, genes bound by Pc in the embryo but not in imaginal discs might represent genes released from Pc silencing on the pathway of disc development and hence might represent key cell-fate determining genes for that pathway. However, the GO analysis of these gene sets (summarised in Figure 5 (also see Table S3)) presents a striking observation. The Pc target genes that are unoccupied by Pc in a particular tissue appear to have little to do with cell fate decisions that are relevant for that tissue. For example, the genes occupied by Pc in the embryo but not in the imaginal discs are enriched in genes involved in fate decisions in neuroblasts of the central nervous system. Similarly the genes that are occupied in embryos but not in S2 cells, which are mesodermally–derived cells of the haemocyte lineage, are genes required for ectodermal and neural fate decisions. The genes that are occupied in imaginal discs but not in S2 cells are relevant for fate decisions occurring in discs but not in S2 cells (e.g. sensory organ development, eye development, ectoderm development).
We further investigated the group of 49 genes which are associated with Pc in the embryo but not in T3 discs. Figure 6A lists these genes and shows their pattern of binding of Pc and Pho across the five data sets. Representative binding profiles are shown in Figure 6B.
As the plot in Figure 6A demonstrates, approximately half of the Pc targets also show Pho binding in the embryo (49%) and, as with Pc, Pho binding is absent in the T3 discs. For this gene set, target occupancy in the S2 cells is similar to that observed in imaginal discs with only a few targets (14%) showing Pc binding. The set of 49 genes specifically bound in the embryo contains several well-characterised genes; notably run, hb and tll that are involved in early embryonic patterning and in neuroblast specification as well as two genes, ind and vnd involved in the specification of the nervous system and in neuroblast fate. As mentioned above, this class of genes has a strong GO signature and is highly enriched in transcription factors (Figure 5). Some individual classes of transcription factor are particularly strongly enriched. For example, 3 out of the 19 forkhead domain proteins present in the Drosophila genome are represented (hypergeometric probability 3.5E-05) and 3 putative hormone-receptor C4-zinc finger genes of the 21 in the genome (4.7E-05).
The GO analysis and the individual gene functions suggest that this set of genes may be involved in early embryonic fate decisions but not in fate decisions that are relevant for the imaginal disc cells, where these targets are unoccupied. To explore this we asked whether these genes are actually expressed in imaginal discs. All of the six genes tested for expression by RT-PCR are expressed the embryo but show little or no expression in imaginal discs (Figure 6C). Thus, as with the Hox genes, we find Pc occupancy is not linked to expression state in a simple fashion.
Drosophila S2 cells are an embryo-derived cell line that appear to be related to haemocytes since they are phagocytic and express haemocyte markers [37]. We were interested to relate the Pc binding profile in these cells to the genes involved in cell fate decisions in the haemocyte lineage (reviewed in [38]). The embryonic haemocytes are derived from a head mesoderm primordium defined by the GATA transcription factor, Serpent (Srp), and differentiate into crystal cells or plasmatocytes. Lozenge (Lz), a Runx family transcription factor, is required for crystal cell development whereas U-shaped (Ush) antagonises crystal cell development and Glial cells missing (Gcm) promotes plasmatocyte development. The closely related Gcm2 acts redundantly with Gcm in plasmatocyte differentiation. Full maturation of plasmatocytes requires the PDGF/VEGF Receptor (Pvr). S2 cells express srp together with the plasmatocyte markers ush and pvr and do not express the crystal cell marker lz (FLIGHT database: http://flight.licr.org/, [37]). The expression status of gcm and gcm2 is less clear; they are not scored as expressed in S2 cells in the FLIGHT database but are reported to be detectable by RT-PCR [37].
The key cell fate-determining genes in the haemocyte lineage, srp, lz, gcm, gcm2 and ush, are all Pc targets. Figure 7 compares the Pc and Pho occupancy at srp, ush and lz in S2 cells with the occupancy in embryos and imaginal discs. Strikingly, the cell fate genes associated with the plasmatocyte fate, srp and ush, show strongly reduced Pc occupancy in S2 cells compared to embryos and imaginal discs whereas the crystal cell determining gene, lz, shows clear Pc binding. The comparative binding at srp is dramatic as there is a highly specific reduction in Pc binding in a specific domain over the srp gene in S2 cells, whereas the neighbouring gene GATAe is strongly associated with Pc binding.
Overall, this analysis of Pc binding at cell fate-determining genes in the haematocyte lineage shows clear differential binding in S2 cells that correlates with gene expression and the requirement for gene activity in the plasmatocyte pathway.
The Pc maintenance machinery functions to stably propagate states of gene activity through cell division and, for the Hox genes, stable expression patterns are preserved throughout development. We were interested to examine if this is also true for other Pc target genes. If Pc targets in general are stably expressed once activated, then differentiated cells may express the set of Pc target genes that have been activated along the developmental pathway they have followed. We used the FlyAtlas data set (http://www.flyatlas.org/) of gene expression profiles from selected adult and larval tissues to examine the pattern of deployment of Pc target genes in specific tissues [39]. Out of the 386 Pc target genes we identified with embryo chromatin, we obtained tissue-specific expression data for 373 genes from FlyAtlas. The cluster analysis of the expression data is presented in Figure 8 (and Figure S5). We find that the data for this small sample of genes out of the 18,770 transcripts in the data set nevertheless clusters according to tissue type. For example, the two neural samples, brain and thoracic/abdominal ganglia cluster together, as do the crop and hindgut samples representing ectodermal-derived gut ensheathed in visceral mesoderm. Thus the Pc target gene expression profile provides a tissue-type signature. The cluster diagram, in addition to the block of genes that are present in all samples, also provides several blocks of genes whose expression is related to particular tissues. For example, the block illustrated in Figure 8B includes genes expressed in foregut (crop) and in hindgut and contains the two key genes bin and bap that are required for the specification of the visceral mesoderm in mid-embryogenesis. This analysis indicates that not only the Hox genes but also other Pc target genes remain stably expressed through to adulthood, providing a basis for the stable specification of cell type.
The PcG target genes identified by several genome-wide binding studies represent an assembly of key regulators that generate cellular diversity and patterning in the developing organism [4],[5],[10]. Coupled with the ability of the PcG/TrxG maintenance machinery to stably transmit states of gene expression through cell division, this provides a system for the stable inheritance of cell fate decisions and for the stability of differentiated cell states. In this paper we have examined the linkage between the PcG machinery and cell fate decisions by comparing the binding sites of PcG machinery components in different tissues. We have generated genomic binding profiles for Pc and Pho from whole Drosophila embryos and from the specific tissue of the imaginal discs of the third thoracic segment. We combined our data from in vivo sources with the data from Drosophila S2 tissue culture cells [10] allowing comparison of PcG occupancy in chromatin from different tissues. In general we find considerable overlap of target sites in the three data sets. However Pc and Pho binding to target sites is not simply constitutive and we find clear examples of alterations in Pho and Pc binding at specific target sites in particular tissues.
We find a substantial number of genes associated with Pho binding but not Pc, suggesting a function for Pho separate from its role in PcG-mediated gene silencing. Pho has been found to be associated with two distinct protein complexes, Pho-dINO80 and PhoRC [40]. PhoRC is implicated in PcG-mediated gene silencing and it contains dSfmbt, a PcG protein required for Hox gene repression. PhoRC, but not the Pho-dINO80 complex is bound at PREs. The role of the Pho-dINO80 complex is unknown but it is a candidate for mediating Pho function at the Pc-independent Pho targets. Null pho mutants lacking any maternal contribution exhibit severe pleiotropic phenotypes and one pho allele shows a specific female sterility phenotype [15]. In this respect it is interesting to note that the Pho target genes we identify are overrepresented for oogenesis, mitosis and splicing functions. Of the 212 Pho-only targets, 60% are enriched for ovary expression and 66% are absent in testis according to FlyAtlas (http://www.flyatlas.org/). This suggests that pho may regulate a set of specific functions during oogenesis and we suggest that Pho continues to be associated with these targets in the embryo. A more general role for Pho, separate from PhoRC function, is also suggested by clonal analysis of pho and dSfmbt mutations in imaginal discs [40]. Mutant clones lacking pho (together with pleiohomeotic like which functions largely redundantly with pho) show not only loss of Hox gene silencing but also growth defects that result in the elimination of the clones from the disc epithelium. Clones lacking dSfmbt lose Hox silencing but do not show growth defects. Recently, Pho was found to be bound at active genes and is strongly recruited at sites with high transcription (chromosome puffs) on salivary gland polytene chromosomes. Based on the kinetics of Pho binding at heat-shock loci a role for Pho in the repression of previously active genes was proposed [23].
Examination of Pc and Pho binding in the BX-C in T3 imaginal discs provides a clear test case for the linkage between occupancy and gene expression since Ubx is expressed but abd-A and Abd-B are silenced. We find significant Pc and Pho binding associated with both the expressed and the silenced genes. This provides a clear demonstration that silencing is not simply established by the presence of PcG proteins at a target site and supports previous observations of a lack of correlation between PcG binding and gene silencing [22],[23],[33]. Although the Ubx gene is associated with significant Pc and Pho binding, there is overall less binding over Ubx in comparison with the two silenced genes. Also, the Pho binding profile in the embryo, representing a mixture of gene activity states, is markedly different from the T3 disc profile; the T3 profile shows prominent peaks at the bx and bxd PREs but the other peaks seen with embryo chromatin are less prominent. Similar effects are also seen at the Antp gene, which is also active in T3 discs. Reduction and rearrangement, rather than absence, of PcG protein at active genes suggests a dynamic interaction between silencing and gene transcription. Indeed, Pc complexes have been shown to be highly dynamic with rapid exchange of PcG proteins on chromatin [41]. Alteration of the Pho binding profile at Hox genes in cell lines with differential Hox expression has also been reported [23] with the striking observation of spreading of the Pho binding at active loci. Such dramatic Pho spreading is, however, not apparent in our data.
We identified a set of 49 Pc target genes that were bound by Pc in the embryo but not in the T3 imaginal discs. We had expected that such a class might contain genes required for fate decisions on the pathway to T3 imaginal disc cell differentiation and were surprised to find that this gene set was enriched in genes required for early cell fate decisions in the nervous system. We examined the expression of several of these genes and found little or no expression in the T3 discs, reinforcing the idea that Pc binding does not simply correlate with silencing of gene expression. In the case of Ubx and Antp we find that expressed genes have significant Pc binding and, with the set of genes that show Pc association in embryos but not in imaginal discs, we find inactive genes that lack Pc. This is reminiscent of an observation with the Pc target hedgehog (hh) which has an identified PRE and is silenced by PcG in the embryo and imaginal discs [42],[43]; in salivary gland polytene chromosomes, Chanas and Maschat found no PcG binding at the hh gene despite the fact that hh is not expressed in this tissue [43]. A similar observation was made for CycA [44]. In this case we note that although hh is a clear Pc target in S2 cells and in our in vivo analysis, CycA is not [10]. In all of these cases of differential Pc binding it is possible that the particular genes are inactive due to the absence of appropriate transcription factors to drive expression in particular tissues. This contrasts with the situation in the Hox genes where Pc is continuously required to maintain silencing against a background presence of gene activators [45]–[47]. If Pc complexes are only recruited to genes where they are required to counteract gene activation, this would provide an economical way to deploy the silencing machinery. It would also imply the existence of a mechanism that enables the PcG-machinery to identify genes that are capable of being activated. A possible basis for this mechanism could be the targeting of Polycomb complexes by non-coding RNAs; PcG proteins are recruited by non-coding RNAs in mammalian X chromosome inactivation [48] and recent studies implicate non-coding RNA in Hox gene repression [49],[50]. Alternatively, the lack of Pc associated with non-expressed genes may indicate that these genes are repressed through non-PcG dependent mechanisms. In the case of hh in the salivary glands, there is some support for this since attempts to activate hh by supplying activators were unsuccessful [43]. In addition, a study on histone modifications and cell lineage provided evidence for a class of genes that lose both the PcG-dependent H3K27me3 mark associated with silencing and the TrxG–dependent H3K4me3 mark associated with activation on lineage progression [8]. Loss of both marks was found to be associated with gene inactivity.
In the case of the haemocyte lineage cell fate-determining genes required for plasmatocyte development, these genes are expressed in S2 cells and are found to be selectively unoccupied by Pc. This is what would be expected for a non-silenced active gene and fits with the idea that Pc is lost from PRE/TREs following switching to the active state. However it raises the question of why Ubx or Antp, as genes expressed in T3 imaginal disc cells, are still associated with significant Pc binding whereas a Pc target gene such as srp appears to lack Pc binding in haemocyte lineage cells. S2 cells are tissue culture cells whose gene regulatory systems may have deviated considerably from the endogenous state and it is therefore possible that the observed Pc status represents a tissue culture artefact. However, another possibility is that it relates to plasticity. Imaginal disc cells are relatively undifferentiated precursor cells that only differentiate fully during metamorphosis. S2 cells, on the other hand, may represent a more committed cell state. In this respect it is interesting to compare the Pc profiles observed at the BX-C in S2 cells and T3 imaginal discs (Figure 4A). S2 cells express high levels of Abd-B but very low levels of Ubx and abd-A. Pc binding reflects this gene expression and in particular shows no binding over an Abd-B domain that includes the four active Abd-B promoters [10]. This contrasts with the situation in T3 imaginal discs where the active Ubx gene is associated with significant Pc binding. However, it should be noted that the state of Hox gene expression in S2 cells is rather curious since these cells are thought to derive from the head mesoderm, an area of the embryo that does not express any of the genes of the BX-C. Despite this caveat, the differences in the Pc binding at active genes in S2 versus imaginal disc cells may reflect the plasticity of the undifferentiated imaginal disc cells compared to the loss of plasticity in the S2 cells. In general, we have identified two situations where Pc target genes are not bound by Pc proteins, a set of inactive genes in imaginal discs and a set of active genes in S2 cells and the common feature may be that these both represent terminal stable gene states. In these situations, loss of Pc binding may be associated with loss of plasticity and may indicate final cell commitment.
Our analysis of T3 imaginal discs enabled us to investigate the Pc occupancy of genes in this specific tissue but does not immediately reveal the developmental history of these cells in terms of which cell fate switches had been turned on and which had been turned off along the developmental pathway leading to T3 imaginal epidermal specification. The Pc target genes which exhibited no Pc binding in the T3 imaginal discs did not obviously suggest a set of fate-determining genes for T3 disc specification. In the relatively undifferentiated imaginal disc cells it is apparent that Pc occupancy by itself does not differentiate a silenced from an active state and so to map the fate switching history of a cell we will either need to find markers that provide a clearer readout of the state of gene activation or else we will have to look at more differentiated cells where the PcG system has stabilised. Our analysis of a limited set of adult tissues, where gene expression data is available, provides support for stable activation of cell fate decision genes, suggesting that examining the expression of Pc target genes in differentiated cell types can provide information on the key developmental genes that are activated on a specific developmental pathway. Although the T3 imaginal discs represent a tissue sample of limited cell fate diversity they are nevertheless a complex mixture of cells with different states of gene activity. Many of the known key genes in imaginal disc development e.g. vg, Dll, hh and tsh are active in only a subset of disc cells and therefore the Pc and Pho binding profiles we observe may represent a mixture of active and silenced states. Further analysis examining more restricted tissues will be required to investigate to what extent the Pc target genes provide a stable “genetic address” [51] specifying cell differentiation.
The wild type strain used was OregonR. The Pc-GFP transgenic fly line was generated by Dietzel et al. [52]. The antibodies used were affinity purified rabbit anti-GFP [53], rabbit anti-Pho [18] and affinity-purified anti-Pc [54].
Chromatin from embryos aged between 0 to 16 h after egg laying was purified as described previously [55]. For the preparation of chromatin from T3 imaginal discs (haltere and third leg) late 3rd instar larvae were dissected in ice-cold Schneider's Medium. Dissected discs were washed with PBS, fixed in PBS/1.5% formaldehyde for 20 min and washed with PBS. Batches of material were snap-frozen in liquid N2 and stored at −80°C. Chromatin was prepared from a minimum of 100 discs. For Pc target analysis the specific reaction used chromatin from the Pc-GFP fly line immunopurified using anti-GFP, and the control reaction used wild type chromatin immunopurified using anti-GFP. The Pc-GFP protein binds to the same polytene chromosome loci as wild-type Pc [41] and we validated a selection of targets by ChIP using anti-Pc antiserum (Figure S1). For Pho analysis wild type chromatin was used with anti-Pho for the specific reaction and pre-immune antiserum for the control. For validation reactions anti-Pc and anti-Pho were used for the ChIP and enrichment was assayed using PCR with specific primers as described previously [55]. The primer sequences are given in Table S4.
Three biological replicates were used for each condition and enrichment profiles were generated by comparison of specific and control ChIP DNA samples. In order to identify regions bound by Pho or Pc, 10–20 ng of ChIP and control DNA samples were amplified using a random-primed PCR method according to Affymetrix recommendations (Affymetrix Chromatin Immunoprecipitation Assay Protocol; http://www.affymetrix.com/support/technical/manuals.affx). Purified DNAs were then fragmented, TdT labeled, and hybridized to the Affymetrix Drosophila genome Tiling Array 1.0 (reverse part no. 520,054) as described previously [56]. ChIP–array data have been deposited in the GEO database under accession code GSE11006.
Affymetrix CEL files were converted into chromosomal enrichment profiles using the TiMAT2 package (http://bdtnp.lbl.gov/TiMAT/TiMAT2/). Probe mapping information (“bpmap”) to D. melanogaster genome release 4 was obtained from David Nix. Each CEL file was visualised for manual inspection and artefacts were masked using CelMasker. Normalisation was subsequently performed with CelProcessor using default parameters. Enrichment profiles were generated using ScanChip, outputting windowed enrichment signals and Wilcoxon Rank Sum scores. The .sgr files are provided in Datasets S2–5).
We classified enrichment events into positive, intermediate and negative based on visual inspection in the Integrated Genome Browser (http://www.affymetrix.com/support/ developer/tools/download_igb.affx). We found that our manual classification could be automated using basic descriptive statistics. Positive bound regions (“peaks”) were characterised by enrichment values greater than an experiment-specific cutoff as well as a Wilcoxon Rank Sum score greater than 55. Intermediate regions were score-independent but showed an enrichment value greater than 50% of the experiment-specific cutoff. Negative regions accounted for all regions that did not fulfil these criteria. The experiment-specific cutoffs were empirically determined as the signal average plus three standard deviations for Pc (log2 enrichment ratio of 0.39 for the embryo and 0.77 for the T3 disc material) or five standard deviations for Pho (log2 enrichment ratio of 0.62 for the embryo and 0.80 for the T3 disc material). For the S2 Pc data of [10], we followed a similar classification with positive regions having enrichments greater than the signal average plus three standard deviations and negative regions showing ratios of less than 50% of this value.
We assigned each binding event to a target gene, based on complete or partial overlap with a gene model. Binding events that did not overlap with a gene model were assigned to the nearest gene. In most cases for Pc, this concerned bound regions that represented extensions of larger domains overlapping with the gene.
We selected the 150 strongest Pho peaks that overlapped with Pc binding and generated two different search sets comprising 200 nt or 700 nt of sequence around the centre of the peaks. We used NestedMICA [57] to search for statistically overrepresented sequence motifs in the search set. A first round of searches was performed with NestedMICA v0.72, specifying expected motif length and usage frequency. A second search was performed with NestedMICA v0.8 using default usage frequency and dynamic motif length. Both searches aimed to identify 10–15 overrepresented motifs. Candidate motifs were visually inspected in MotifExplorer and a set of promising candidates resembling Pho-, GAGA- or Zeste-like motifs as well as some with high information content were chosen for further analysis.
Statistical overrepresentation of motifs was determined by comparing the set of all Pho peak sequences to 1,000 sets of random sequences of the same length, representing regions of the Drosophila genome that are not bound by Pc. A Z-score was derived, incorporating the number of occurrences in real peaks and the numbers observed for the 1,000 random sets.
All downstream analyses were performed with custom-made Perl scripts. Clustering and visualisation was performed with Genesis v1.6 [58]. Binding sites in the sequence context were visualised in BioSAVE [59].
Enrichment of Gene Ontology terms was determined with the GeneMerge 1.2 software tool, comparing enrichment in specific lists with all Drosophila genes. Gene Association files used were March 2007 release of the Gene Ontology. The enrichments quoted in the text are corrected for multiple testing by applying a modified Bonferroni method within the Gene Merge algorithm. Enrichments with e-scores better than 0.05 are called significant. Tissue expression analysis used the data from FlyAtlas [39] with clustering and visualisation using Cluster [60] and Java Treeview [61].
OregonR embryos (0–20 hr) were dechorionated with bleach, then divided into aliquots, placed directly in Trizol and stored at –80°C. Homogenisation and RNA extraction were carried out according to the following protocol: http://www.flychip.org.uk/protocols/gene_expression/standard_extraction.php. T3 leg and haltere discs were dissected from wandering 3rd instar OregonR larvae in PBS. Each pair was transferred in a small drop of PBS directly into 100 µl Trizol and frozen immediately. For RNA extraction, 7 disc pairs were randomly pooled for each of 3 samples and RNA extracted as above. RNA samples were treated with RQ1 DNase to remove any genomic DNA. cDNA synthesis was performed by combining 10 µg DNase treated RNA with 500 ng anchored oligo(dT)23 primer (Sigma; Cat. No. 04387), 1 µl of 10 mM dNTPs, DEPC treated H2O to 13 µl. The reaction was heated to 65°C for 5 min then chilled on ice for 1 min. 4 µl of 5x First Strand Buffer (Invitrogen), 1 µl 0.1 M DTT (Invitrogen), 1 µl RNAsin (Promega; Cat. No. 18064-014) and 1 µl Superscript III Reverse Transcriptase (Invitrogen; Cat. No. 18080-044) were added. Reactions were incubated at 50°C for 60 min and inactivated at 70°C for 15 min.
0.5 µl of the resulting cDNA was used in PCR reactions with the following primers: hb-F ggcctcttcgttcacatgg, hb-R agcggcttaattggcttatg, ind-F aacgattatgccgattccag, ind-R gattgaaggtgggactttcg, vnd-F cgacgagatgtcctcgtacc, vnd-R ctcttgtaatcgccggaaag, fd59A-F ttcagtcaccgcacaagaag, fd59A-R gtccagaagttgccctttcc, run-F agtccttcacgctgaccatc, run-R gtagtccgcatagccgtagg, tll-F tacaacagcgtgcgtctttc, tll-R ttgtccaccacacagagtcc, Rp49_F catacaggcccaagatcg, Rp49_R tgggcatcagatactgtcc.
The Flybase (http://flybase.bio.indiana.edu) accession numbers of the genes and gene products discussed in this paper are: abdominal-A (abd-A), FBgn0000014; Abdominal-B (Abd-B), FBgn0000015; Antennapedia (Antp), FBgn0000095; bagpipe (bap), FBgn0004862; biniou (bin), FBgn0045759; brahma (brm), FBgn0000212; Centrosomal protein 190kD (Cp190), FBgn0000283; Cyclin A (CycA), FBgn0000404; Deformed (Dfd), FBgn0000439; Distal-less (Dll), FBgn0000157; dRing (Sce), FBgn0003330; GAGA factor (Trl), FBgn0013263; GATAe, FBgn0038391; gcm2, FBgn0019809; glial cells missing (gcm), FBgn0014179; grainy head (grh), FBgn0004586; hedgehog (hh), FBgn0004644; hunchback (hb), FBgn0001180; intermediate neuroblasts defective (ind), FBgn0025776; labial (lab), FBgn0002522; lozenge (lz), FBgn0002576; PDGF- and VEGF-receptor related (Pvr), FBgn0032006; pleiohomeotic (pho), FBgn0002521; pleiohomeotic like (phol), FBgn0035997; Polycomb (Pc), FBgn0003042; polyhomeotic distal (ph-d), FBgn0004860; polyhomeotic proximal (ph-p), FBgn0004861; Posterior sex combs (Psc), FBgn0005624; proboscipedia (pb), FBgn0051481; runt (run), FBgn0003300; Scm-related gene containing four mbt domains (Sfmbt), FBgn0032475; serpent (srp), FBgn0003507; Sex combs reduced (Scr), FBgn0003339; suppressor of Hairy wing (su(Hw)), FBgn0003567; tailless (tll), FBgn0003720; teashirt (tsh), FBgn0003866; trithorax (trx), FBgn0003862; Ultrabithorax (Ubx), FBgn0003944; u-shaped (ush), FBgn0003963; ventral nervous system defective (vnd), FBgn0003986; vestigial (vg), FBgn0003975; and zeste (z), FBgn0004050.
|
10.1371/journal.pgen.1004128 | Extreme Population Differences in the Human Zinc Transporter ZIP4 (SLC39A4) Are Explained by Positive Selection in Sub-Saharan Africa | Extreme differences in allele frequency between West Africans and Eurasians were observed for a leucine-to-valine substitution (Leu372Val) in the human intestinal zinc uptake transporter, ZIP4, yet no further evidence was found for a selective sweep around the ZIP4 gene (SLC39A4). By interrogating allele frequencies in more than 100 diverse human populations and resequencing Neanderthal DNA, we confirmed the ancestral state of this locus and found a strong geographical gradient for the derived allele (Val372), with near fixation in West Africa. In extensive coalescent simulations, we show that the extreme differences in allele frequency, yet absence of a classical sweep signature, can be explained by the effect of a local recombination hotspot, together with directional selection favoring the Val372 allele in Sub-Saharan Africans. The possible functional effect of the Leu372Val substitution, together with two pathological mutations at the same codon (Leu372Pro and Leu372Arg) that cause acrodermatitis enteropathica (a disease phenotype characterized by extreme zinc deficiency), was investigated by transient overexpression of human ZIP4 protein in HeLa cells. Both acrodermatitis mutations cause absence of the ZIP4 transporter cell surface expression and nearly absent zinc uptake, while the Val372 variant displayed significantly reduced surface protein expression, reduced basal levels of intracellular zinc, and reduced zinc uptake in comparison with the Leu372 variant. We speculate that reduced zinc uptake by the ZIP4-derived Val372 isoform may act by starving certain pathogens of zinc, and hence may have been advantageous in Sub-Saharan Africa. Moreover, these functional results may indicate differences in zinc homeostasis among modern human populations with possible relevance for disease risk.
| Zinc is an essential trace element with many biological functions in the body, whose concentrations are tightly regulated by different membrane transporters. Here we report an unusual case of positive natural selection for an amino acid replacement in the human intestinal zinc uptake transporter ZIP4. This substitution is recognized as one of the most strongly differentiated genome-wide polymorphisms among human populations. However, since the extreme population differentiation of this non-synonymous site was not accompanied by additional signatures of natural selection, it was unclear whether it was the result of genetic adaptation. Using computer simulations we demonstrate that such an unusual pattern can be explained by the effect of local recombination, together with positive selection in Sub-Saharan Africa. Moreover, we provide evidence to suggest functional differences between the two ZIP4 isoforms in terms of the transporter cell surface expression and zinc uptake. This result is the first genetic indication that zinc regulation may differ among modern human populations, a finding that may have implications for health research. Further, we speculate that reduced zinc uptake mediated by the derived variant may have been advantageous in Sub-Saharan Africa, possibly by reducing access of a geographically restricted pathogen to this micronutrient.
| Zinc homeostasis is critically important for human health. Similarly to iron, zinc has manifold functions in the body, such as in the immune system [1], aging [2], DNA repair [3], signaling [4] and in diseases such as diabetes [5] and cancer [6]. On the molecular level, zinc acts as a co-factor in hundreds of metallo-enzymes as well as in hundreds of DNA-binding proteins (e.g. zinc finger proteins). Zinc homeostasis is tightly regulated by 10 zinc efflux transporters and 14 zinc influx transporters (encoded by the SLC30A and SLC39A gene families, respectively). ZIP4 (SLC39A4) is the most important intestinal zinc uptake transporter and is expressed at the apical membrane of enterocytes [7], [8]. Loss-of-function mutations in ZIP4 cause acrodermatitis enteropathica [9], [10] [MIM 201100], a congenital disease characterized by extreme zinc deficiency if left untreated without supplemental zinc [11], [12]. Fittingly, it was recently reported that the loss of expression of this gene in a ZIP4 intestine-specific knockout mouse caused systemic zinc deficiency, leading to disruption of the intestine stem cell niche and loss of intestine integrity [13].
The single nucleotide polymorphism (SNP) c.1114C>G (rs1871534) in the ZIP4 gene (SLC39A4; NM_130849.2) results in the substitution of leucine for valine at amino acid 372 (Leu372Val) in the human ZIP4 transporter. This non-synonymous SNP is one of the most markedly differentiated genetic variants in the genome in terms of allele frequency differences between populations [14]–[16], according to data from HapMap [17], the Human Genome Diversity Panel (HGDP) [18] and the 1000 Genomes Project [16]. Extreme population differentiation is a signature of local positive selection [15], [19]–[21], but genomic scans for targets of natural selection based on other criteria, such as extended long haplotypes [22]–[24] or selective signatures in the allele frequency spectrum [25], have failed to identify ZIP4 as a candidate gene for positive selection. To date, whether this variant has evolved under positive selection or neutrality, and its potential functional significance, has not been examined.
In the work reported here, we had three main objectives: (i) to investigate evolutionary explanations for the extreme population differentiation of the ZIP4 Leu372Val polymorphism by use of coalescent simulations; (ii) to test for functional differences in cellular zinc transport between the alleles of the Leu372Val polymorphism using a heterologous expression system; and (iii) to discuss potential selective forces behind this possibly adaptive event and their implications for zinc homeostasis in modern humans. We have extensively characterized the extreme geographical differentiation of the Leu372Val substitution and provide evidence that it has been subject to a nearly complete but mild selective sweep in Sub-Saharan Africa. Our simulations show how the extreme pattern of population differentiation, yet absence of other classical signatures of positive selection, can be explained by directional selection accompanied by the effects of a recombination hotspot near the polymorphic adaptive site. Additionally, our data demonstrate in vitro functional differences between the two human polymorphic alleles at codon 372 of the human ZIP4 transporter in surface protein expression, basal intracellular levels of zinc and zinc uptake. We hypothesize that the reduction in intracellular zinc levels mediated by the Val372 allele may have been advantageous in Sub-Saharan Africa, possibly by restricting access of a geographically restricted pathogen to this micronutrient, and that other possible secondary consequences for disease risk and health may result from the differential activity of the ZIP4 alleles.
Five common non-synonymous SNPs are known in the human ZIP4 gene (Table 1): Glu10Ala (rs2280839), Ala58Thr (rs2280838), Ala114Thr (rs17855765), Thr357Ala (rs2272662) and Leu372Val (rs1871534). However, only the latter two SNPs show elevated levels of population differentiation in the 1000 Genomes Phase1 sequencing data when comparing the Yoruba from Ibadan, Nigeria (YRI) with either the Han Chinese from Beijing, China (CHB) or Utah residents of Northern and Western European origin (CEU). As shown in Figure 1A and 1B, their FST values fall above the 99.999 th percentile of the genome-wide FST distributions between CEU-YRI (with FST values for rs2272662 and rs1871534 of 0.48 and 0.98, respectively) and between CHB-YRI (with FST values of 0.51 and 0.98, respectively). We therefore verify that the Leu372Val substitution encoded by SNP rs1871534 is the non-synonymous polymorphism exhibiting the most extreme allele frequency differences in the human ZIP4 gene. Next, we genotyped the 51 populations from the Human Genome Diversity Panel (HGDP) and compiled additional allele frequencies for this position in worldwide populations from the Alfred database [26], [27]. Additionally, we obtained new data from a Pygmy population from Gabon and North African populations of Western Sahara, Morocco, and Libya. These new data confirm that the Leu372 variant is the most common allele outside of Africa, and provide a more detailed picture of the geographical allele frequency distributions of this non-synonymous polymorphism (Figure 1C and Table S1). Overall, the Val372 variant showed the highest frequencies in Sub-Saharan Africa, with populations such as the Ibo or the Yoruban people exhibiting the most extreme derived allele frequencies worldwide (0.99 and 0.96, respectively). Interestingly, two presumably early-branching groups in Sub-Saharan Africa, the Pygmy and the San people, showed opposing trends in the derived allele frequency (0.94 and 0.0, respectively). Even though the small sample size from the San (only six individuals) means that a population frequency of up to 0.221 cannot be excluded (with p = 0.05 based on assuming Hardy-Weinberg equilibrium and a binomial approach), such divergent tendencies in these two Sub-Saharan populations are maintained. Given the elevated levels of population differentiation of the SNP rs2272662, we also genotyped the HGDP panel for the Thr357Ala polymorphism. However, compared with the Leu372Val substitution, the derived allele at this non-synonymous SNP displayed intermediate frequencies worldwide (Figure S1 and Table S1) and less extreme allele frequency differences between populations.
Given the allele frequency differences observed in the Leu372Val polymorphism between the two early human branches in Africa and the uncertainty associated with the low coverage of the Neanderthal genome draft sequence [28], we resequenced the corresponding orthologous positions for rs1871534 and rs2272662 in an additional Neanderthal sample, labeled SD1253 and excavated at El Sidrón site in Spain [29]. The two positions were amplified in a multiplexed reaction, along with a diagnostic Neanderthal mitochondrial DNA (mtDNA) fragment, to monitor contamination in the PCR reaction. For the L16230-H16262 diagnostic mtDNA fragment, 64 clones were generated (Figure S2), all of which show the Neanderthal-specific 16234T-16244A-16256A-16258G haplotype [28]. This again supports the very low level of contamination in this particular sample. For the orthologous positions of the human rs1871534 and rs2272662 SNPs, 19 and 14 sequences were successfully obtained, respectively. With the exception of one clone in the second position, all sequences showed the previously inferred ancestral alleles, in agreement with the reads present for the Vindija individuals 33.16 (one read for each position), 33.25 (two for rs1871534 and none for rs2272662) and 33.26 (two and one, respectively) (Figure 2). The successful resequencing of this Neanderthal individual, together with published reads from additional Neanderthals [28] and from the Denisovan individual [30], strongly suggests that the Leu372 variant (encoded by the C allele in rs1871534) is the ancestral human form, which is also in agreement with the chimpanzee state (Figure 2). Together with the extreme population differentiation pattern, these results suggest that a selective sweep may have taken place in Sub-Saharan Africa, where the derived variant is nearly fixed.
Next we examined the complete genomic region around ZIP4 (Figure 3) in the 1000 Genomes sequencing data. Whereas we found a cluster of three strongly elevated FST scores between CEU and YRI in the neighboring SNPs rs1871535 (intronic), rs1871534 and rs2272662 (further suggesting directional selection in a specific geographical region), in both populations there was a clear absence of extreme values in neutrality statistics such as Tajima's D or Fay and Wu's H (Figure S3). Notably, no other polymorphism in the flanking region of the human ZIP4 displays the high levels of population differentiation of the Leu372Val substitution. Interestingly, in both African and non-African populations there is a recombination hotspot in the ZIP4 gene, which could have reduced any signature of selection on the surrounding linked variation, thereby explaining the apparent lack of significant departures from neutrality. To further investigate this possibility, we carried out coalescence simulations under a variety of recombination and selection scenarios using a well-established demography [31]. As shown in Figure 3D, the observed values for FST and most of the different neutrality statistics cannot be explained by neutral evolution or positive selection with a constant recombination rate. Instead, this atypical pattern of extreme population differentiation, yet seemingly neutral Tajima's D and other neutrality statistics, showed a higher recovery in simulations with directional selection on the derived allele in Sub-Saharan African populations in the context of the observed recombination landscape, including the hotspot (Figure 3D and 3E, Figure S4). In a more formal evaluation of the results, we quantified the empirical probability for each scenario and neutrality test as well as for different combinations of tests by using composite scores encompassing at least three complementary signatures of positive selection: (i) site frequency spectrum, (ii) population differentiation, and (iii) haplotype structure. The scenario of “weak selection (s = 0.005) + hotspot” is the most likely among the different ones tested (Table S2). Moreover, all the empirical likelihoods calculated for the different composite scores indicate that the proposed scenario of “weak selection (s = 0.005) + hotspot” is more likely than the neutral scenario (Table 2). Therefore, our simulation results indicate that the atypical patterns of selection in the ZIP4 gene can indeed be explained by positive selection having acted upon the Val372 allele in Sub-Saharan African populations and that recombination has erased further accompanying signatures of the selective sweep. Selection coefficients lower than the ones tested (3.0%, 1.0%, 0.5%) further dilute the signal of selection in the site frequency spectrum based neutrality tests (results not shown), but require such long duration times of the sweep that would substantially predate the population split between African and Eurasian populations.
We observed that the Leu372Val polymorphism affects a highly conserved amino acid (Figure 4) and that the same codon position has been altered in acrodermatitis patients carrying missense mutations Leu372Arg [32] and Leu372Pro [8]. Moreover, both PolyPhen [33] and SIFT [34] algorithms predict functional effects for the Leu372Val substitution (see Table 1). These observations led us to test the Leu372Val polymorphism for a possible functional change in the ZIP4 transporter, using transiently transfected HeLa cells. To be able to control for possible haplotypic effects between the two most highly differentiated non-synonymous SNPs in the ZIP4 transporter, we also considered variation at the Thr357Ala polymorphism in the functional analyses. Furthermore, we introduced the pathological mutations Leu372Arg and Leu372Pro in the Ala357 background of the human ZIP4 gene and analyzed them as well. The pathological impact of the Leu372Pro mutation on ZIP4 protein biology and function has already been evaluated in the mouse ZIP4 protein [10], but not the Leu372Arg mutation. Besides providing confirmation of their impact in the context of the human gene, the use of these pathological mutations provided us with an extreme phenotype to which to compare the phenotype associated with the ZIP4 non-synonymous polymorphisms. In all cases, functional analyses were carried out to determine effects on expression, subcellular localization, and zinc transport.
As shown in Figure 5, human ZIP4 proteins carrying the Leu372Pro and Leu372Arg mutations showed an absence of surface protein expression (P<0.001, one way ANOVA versus the Ala357-Leu372 isoform), consistent with the known causal role of these variants in the zinc deficiency disorder, acrodermatitis enteropathica. Interestingly, the derived Val372 variant also showed significantly decreased surface expression, but to a much lesser extent, and independently of the Thr357Ala substitution (P<0.05 in both Ala357 and Thr357 backgrounds; one way ANOVA versus the Ala357-Leu372 isoform). Overall, the Leu372Val substitution had a highly significant effect on surface expression (ANOVA, p = 0.00021), while there was no effect ascribable to the Thr357Ala replacement (p = 0.579). Western blot analysis of all isoforms revealed a remarkable decrease in detection of the Ala357-Pro372 isoform (Figure S5A). However, the reduced expression of this isoform was not due to a defect in the construct sequence but to a higher protein degradation rate, as shown in Figure S5B. Further analysis showed that the Ala357-Leu372 and Ala357-Val372 isoforms do not differ in protein degradation rate. Therefore, the differences in the surface expression experiment must be due to a different trafficking pattern of these variants. In this sense, co-localization of ZIP4 with calnexin (a protein present in the lumen of the endoplasmic reticulum) indeed showed that those proteins presenting lower surface expression were partially retained in the endoplasmic reticulum (Figure S6).
Zinc transport analysis of the different ZIP4 isoforms was performed in two ways. First, we quantified basal zinc content with FluoZin-3 in HeLa cells overexpressing the various ZIP4 variants during a 24-hour period (Figure 6A), and second, we recorded intracellular zinc uptake upon perfusion with an external solution containing 200 µM Zn2+ (Figure 6B). Our results show that basal zinc content in cells overexpressing pathological variants Pro372 and Arg372 did not differ from surrounding non-transfected HeLa cells. On the contrary, all common ZIP4 variants (Ala357, Thr357, Leu372 and Val372) promoted increased intracellular zinc levels. However, and in agreement with their reduced surface expression, Val372 variants (in both Ala357 and Thr357 backgrounds) presented lower basal zinc content compared to Leu372 (P<0.01 and P<0.05, respectively; one way ANOVA versus the Ala357-Leu372 isoform; Figure 6A). As shown in Figure 6B, cells overexpressing the pathological Leu372Arg and Leu372Pro mutations did not uptake zinc, consistent with their inability to traffic to the plasma membrane. Zinc uptake mediated by the Val372 variants was also consistent with their reduced membrane expression; i.e. the Val372 variants in both Ala357 and Thr357 backgrounds presented significantly lower maximum transport (Tmax) compared to the Leu372 variant (P<0.01 in each case; Figure 6B). However, the time to reach half-maximal transport (t1/2) showed no significant difference, indicating that transport kinetics were not markedly different among the four common variants (Figure 6). Overall, these results support the idea that the Val372 variant does not disturb the kinetics of the ZIP4 transporter but leads to lower zinc uptake transport due to reduced surface expression.
Our study was triggered by the observation of extreme population differentiation between Sub-Saharan African and non-African populations involving the Leu372Val polymorphism in the ZIP4 gene, unaccompanied by any other signals of a classic hard sweep, such as long extended haplotype homozygosity, in either population (Figures S3, S7 and S8). By interrogating and compiling allele frequencies in more than 100 worldwide human populations, we further characterized the extreme population differentiation of the Leu372Val polymorphism and confirmed that this result was not an artifact of allele switching [15]. Given the worldwide distribution of the human derived and ancestral alleles (confirmed by sequencing a Neanderthal and phylogenetic conservation), we conclude that this sweep must have taken place within Africa, probably in Sub-Saharan Africa, and not outside the African continent. Notably, the extreme population differentiation of the Leu372Val polymorphism represents the top fourth region within the global genome-wide FST distribution between CEU-YRI obtained from the 1000 Genomes Project data. The only CEU-YRI FST values that are more extreme all involve well-known examples of local geographical adaptation in humans: the SLC24A5 and SLC45A2 genes (with an FST of 0.9826 and 0.9765, respectively), which have been associated with light skin pigmentation in Europeans; and the DUFFY gene (with an FST of 0.9765), which provides resistance to the malaria pathogen Plasmodium vivax. Moreover, with the notable exception of DUFFY FY*O allele [35], [36], most of the extreme FST values obtained when comparing Africans with non-Africans are usually attributed to local adaptation outside of Africa. Our detection of such a rare signature of natural selection in the African continent is therefore quite remarkable. Interestingly, it is congruent with a recent study that has found only limited evidence for classical sweeps in African populations, which is likely due to a combination of limitations of the currently used methodology and specific characteristics of African population history [37].
Notably, we observed a nearly complete but mild selective sweep for the Val372 variant in Africa, which involves three SNPs with extremely high population differentiation, whereas most other commonly used tests for selection show values not even close to genome-wide significance. Our coalescent simulations indicate that this unusual pattern might be explained by local positive selection in combination with an observed recombination hotspot of moderate strength. At approximately 7 cM/Mb, the recombination rate is only around 7-fold higher than the genomic background, but the hotspot is extended over 3–4 kb. Therefore, a similar number of recombination events may accumulate over time corresponding to a more typically sized hotspot of 1 kb and a recombination rate of around 25 cM/Mb. To our knowledge, this is the first example of a nearly complete selective sweep that is obscured by the effect of a recombination hotspot. It is compatible with earlier theoretical observations that instances of weaker selection in the presence of recombination may not always have an influence on polymorphism statistics [38] and with the observed effect of recombination on the partial sweep around the malaria-related β-globin gene [39]. Because of the unclear effects of the recombination hotspot, it was not possible to estimate the age of the sweep using linkage disequilibrium decay related methods (e.g. [40]). It is likely that a mild selection pressure would have needed a long time to reach the extreme population differentiation values observed, indicating this may be an ancient event. The fact that the high frequency of the Val372 allele is restricted to Sub-Saharan African populations suggests that the selection process started after the Out of Africa expansion of modern humans (i.e. sixty thousand years ago). Alternatively, it is also possible that the bottleneck in the Out of Africa expansion did not sample the Val372 allele, which in turn could explain its absence in most non-African populations. This implies that the Out of Africa event is not a hard upper limit for the age of the selection process.
Other more complex evolutionary scenarios cannot be entirely ruled out, and could warrant a more detailed investigation. For example: (i) selection acting on standing genetic variation, in the sense that the Val372 variant was already segregating when it came under the influence of local selection; (ii) additional directional selection against the Val372 allele in non-African populations; (iii) selection favoring the Leu372 variant on multiple, geographically independent origins mostly in non-African populations, in addition to positive selection on the Val372 variant in Africa; and (iv) ‘gene surfing’ of any of the two variants on the wave of a population range expansion [41]. However, we consider it is unnecessary to invoke such complex scenarios in preference to the simpler one we propose based on coalescent simulations. Moreover, back-and-forth migrations between Sub-Saharan African, Northern African and Middle Eastern populations after the first Out-of-Africa wave of migration [42] could easily explain the observed low-intermediate allele frequencies in Middle Eastern populations without invoking additional selection events.
In the absence of additional linked functional variants in the region, we infer that directional selection has acted on the ZIP4 gene. This conclusion is supported by: (i) the disease phenotype of acrodermatitis enteropathica, which involves extreme and potentially lethal zinc deficiency and is caused by, among others, diverse mutations at amino acid position 372 in ZIP4 [43]; (ii) the absence of cellular zinc transport in Leu372Arg and Leu372Pro acrodermatitis mutants; (iii) the finding that the Val372 variant leads to reduced zinc transport at the cellular level; and finally (iv) the conservation of this amino acid position across diverse species (Figure 4). Furthermore, we infer that the Leu372Val substitution was the functional site targeted by selection due to its location in the predicted center of selection (highest FST), and since it is the only putative functional polymorphism in the ZIP4 gene. Of the other two polymorphic variants with somewhat high allele frequency differences between populations, the Thr357Ala substitution (rs2272662) does not show any functional effect and the intronic rs1871535 cannot be associated with any known regulatory function (according to information on DNAse I hypersensitivity clusters, CpG Islands and transcription factor binding sites available from the ENCODE data (http://genome.ucsc.edu/ENCODE [44]). Therefore, both rs1871535 and rs2272662 are likely to be neutral. Other non-synonymous polymorphisms with intermediate allele frequencies in the ZIP4 gene (Glu10Ala, Ala58Thr, and Ala114Thr) have very low FST scores and are therefore not considered candidate variants for selection.
Our functional results in transfected HeLa cells indicate that the Val372 form of the ZIP4 receptor has lower relative cell surface expression, despite no expected differences in mRNA expression and protein synthesis. Interestingly, we found that this decreased expression translated into reduced zinc transport of the derived Val372 variant at the cellular level. That is, we observed differences in the maximal transport (Tmax) with no significant differences in the transport kinetics (T1/2) between Leu372 and Val372. The functional results observed in transfected HeLa cells are likely to be transferable to other epithelial cells, in accordance with independent experiments showing an effect of acrodermatitis variants at position 372 on surface expression (in CHO cells) and on zinc transport (in HEK293 cells) when using mouse cDNA [10]. However, the critical function of ZIP4 in knockout studies has been shown to primarily affect intestinal zinc uptake [13].
In contrast to the Leu372Pro and Leu372Arg acrodermatitis mutations, which served as controls and showed an almost complete absence of zinc transport, both the Leu372 and Val372 variants must be capable of carrying out zinc transport in the normal range of concentrations, given their high frequency in the healthy population. The consequences of this difference in zinc transport at the organ and organismal level are currently unclear, although there is a strong indication that this variant may indeed be phenotypically relevant. For example, a similar non-synonymous mutation in the porcine homologue of ZIP4 leads to non-pathogenic reduced tissue concentrations of zinc in piglets [45].
Could the concept of “nutritional immunity” [46], [47] involving zinc explain a putative selective force in Sub-Saharan Africa? According to this hypothesis, the human host restricts access to certain micronutrients, so that pathogens become less virulent. This is a well-known mechanism of immune defense mediated by iron metabolism [48], and there are indications that zinc metabolism could have a similar function [47], [49]. For example, hypoferremia and hypozincemia are both part of the acute phase response to infection and both seem to be influenced by a different zinc transporter from the same family, ZIP14 [50]. We speculate that the selective force behind the extreme FST pattern of the Leu372Val substitution may be related to pathogens or infectious diseases. It is known that decreased zinc uptake mediated by ZIP4 leads to decreased zinc concentrations in the major organs, as shown in a mouse knockout model [13]. While the phenotypic effect of the Val372 allele in humans is currently unknown, we conjecture that the in vitro difference may indeed translate into physiological differences, possibly leading to a slightly decreased uptake of dietary zinc. Fittingly, there is suggestive evidence that African genetic ancestry may involve lower serum levels of zinc [51], as African-American children have a fourfold risk of zinc deficiency compared to Hispanic children. This result would suggest that African ancestry may be associated with lower serum zinc levels, although these results may be biased due to differences in lifestyle, socio-economic status etc., and this observation would need to be confirmed by controlled studies. Alternatively, lower zinc concentrations mediated by the Leu372Val substitution in the enterocyte cells could facilitate early diarrheal episodes during a digestive infection in order to reduce the pathogen load on the luminal surface [52], [53]. Similarly, the lower level of expression of the ZIP4 isoform carrying the Val372 variant could also be advantageous if any parasite uses the ZIP4 receptor to enter enterocytes. Furthermore, the selective force may be related to pre-historic differences in dietary zinc due to lifestyle or to local levels of zinc concentrations in soil and the food chain.
No large-scale ethnic comparisons related to serum or tissue zinc concentrations are available. To our knowledge, rs1871534 has not been tested in case-control studies in African populations related to one of the numerous existing infectious diseases like malaria, trypanosmias or Lhassa fever. It is therefore possible that important evidence for a possible selective force has been missed. In future research, the inclusion of additional cell lines, and genotype-phenotype association studies in diverse ethnic populations may help to clarify further phenotypic consequences of this non-synonymous polymorphism. Genotype-phenotype association studies should involve African-American or East African populations in which the Val372 allele is segregating at intermediate frequencies. Candidate phenotypes and traits to interrogate could be serum zinc concentrations, zinc content in hair and nails, serum zinc concentrations after controlled zinc supplementation, and a range of disease traits, especially diseases with an elevated risk in different populations, for example, diverse types of cancer in African Americans. As this SNP was not included in the commonly used Affymetrix and Illumina SNP arrays with up to one million variants (although it is included in several of the latest arrays), potential clinically relevant associations may have been missed. Interestingly, common polymorphisms in other zinc transporters show genome-wide associations with disease traits, such as a non-synonymous variant in the zinc efflux transporter ZnT8 (SLC30A8) and diabetes incidence [54], as well as a regulatory variant in the zinc influx transporter ZIP6 (SLC39A6) and survival in esophagal cancer [55].
The identification of a high-frequency derived allele polymorphism in the ZIP4 zinc transporter gene (SLC39A4), combined with a more complete picture of worldwide allele frequencies and in-depth coalescent simulations, is consistent with a long lasting selective event in Sub-Saharan Africa driven by a moderate selection coefficient. This event did not leave the typical footprint of a selective sweep with long haplotypes or detectable neutral deviations in the allele frequency spectrum of the surrounding region, most likely because of the presence of a moderate recombination hotspot. Through functional experiments we have verified the Leu372Val substitution as the likely causal site. Given that two functionally different alleles of this key component of cellular zinc uptake are distributed so divergently across worldwide populations, our results may point to functional differences in zinc homeostasis among modern human populations with possible broader relevance for health and disease.
The G and C alleles at rs1871534 (Leu372Val) have been swapped in various public sources such as HapMap (http://www.hapmap.org) or dbSNP (http://www.ncbi.nlm.nih.gov/SNP) that report conflicting allele frequencies in populations with a similar geographical origin. This situation led us to repeat the genotyping of this SNP in the Human Genome Diversity Panel (HGDP-CEPH) [18]. We also genotyped rs2272662 (which causes the Thr357Ala substitution) because, within the ZIP4 gene, it shows the second highest allele frequency differences between CEU and YRI HapMap populations and allele frequencies were not available at the worldwide level. The rs1871534 and rs2272662 loci were genotyped in the H971 subset [56] of the HGDP-CEPH [18], representing 51 worldwide populations, and in an additional population from Africa: Pygmies from Gabon (N = 39)[57]. We also genotyped rs1871534 in North African populations from Western Sahara (Saharawi, N = 50), Morocco (Casablanca, N = 30; Rabat, N = 30; Nador, N = 30) and Libya (Libyans, N = 50). Genotyping was performed using Taqman assays C__11446716_10 and C__26034235_10 on an Applied Biosystems Light Cycler (7900HR), according to standard protocols. Additional genotypes for rs1871534 were obtained from the Alfred database (http://alfred.med.yale.edu) [26], [27].
Informed consent was obtained for all human samples analysed and genotyping analyses were performed anonymously. The project obtained the ethics approval from the Institutional Review Board of the local institution (Comitè Ètic d'Investigació Clínica - Institut Municipal d'Assistència Sanitària (CEIC-IMAS) in Barcelona, Spain.
The El Sidrón Neanderthal sample SD1253 has been used in many paleogenomic studies due to its high endogenous DNA content and low contamination levels [28], [58]–[62], attributable in part to having been extracted using an anti-contamination protocol [63]. In addition, it has the advantage of having been dated to 49,000 years ago [64], prior to the arrival of modern humans to Europe. The two orthologous positions for rs1871534 and rs2272662 were amplified using a two-step PCR protocol [59] in a multiplexed reaction along with a diagnostic Neanderthal mitochondrial DNA (mtDNA) fragment. After visualizing the PCR products in a low-melting temperature agarose gel, the bands were excised, purified and cloned using the TOPO-TA cloning kit (Invitrogen). Inserts of the correct size were sequenced on an ABI3730 XL capillary sequencer (Applied Biosystems).
Simultaneous coalescent simulation of recombination hotspots and selection were carried out using Cosi v1.2 [31], [65]. As the underlying neutral demography, we used the best-fit model of Shaffner et al. [31], [65] with slight modifications (Table S3), similar to a previously used approach [66]. In particular, the migration frequencies were set to zero and the time points of the European and African population bottlenecks were moved back to 3,300 generations before present to accommodate the long sweep times resulting from the lowest selection coefficient we used (0.5%). The sweep was shifted back 350 generations to retain the final population expansions with the advantage of (i) a better approximation to the fitted model, and (ii) the generation of sufficient singletons when compared to the 1000 Genomes Phase1 data. Subsequent thinning of the simulated data was performed by removing 48% of singleton positions across all populations to account for the underestimation of singletons in 1000 Genomes data. This correction step yielded a much improved (although not perfect) unfolded site frequency spectrum as displayed by the derived allele frequencies (DAF) and a FST distribution that closely matched the empirical data from 1000 genomes (Figure S9). Specifically, we compared the empirical FST and DAF distributions from the 1000 genomes data against the original demographic “best-fit” model [31] and two models adapted to allow for different selective sweeps (the one from [66] and that applied in the current study). As seen in Figure S9, our modified model matched the empirical data as well as or better than the other demographic models.
For each subsequent simulation, we used either the recombination landscape including hotspots from the YRI population provided by the 1000 Genomes Consortium and based on HapMap 2 trio data (http://1000genomes.org) or alternatively a constant recombination rate of 8.17×10−9, which was calculated as the mean recombination rate in the 100 kb window surrounding ZIP4. Simulations had a length of 100 kb, were run in 500 replicates for each scenario and sample sizes were set to 176 chromosomes for Sub-Saharan Africans and 194 chromosomes for Europeans. Regions under positive selection were modeled using a single causal variant that rose to an allele frequency of 0.98 corresponding approximately to that observed today in YRI. We simulated three different selection coefficients (0.5%, 2% and 3%) that led to different durations of the sweep: 2,938 generations (∼60,000–85,000 years for generation times of 20 and 29 years, respectively;[67]), 1,469 generations (∼30,000–43,000 years), or 458 generations (∼10,000–13,000 years).
Empirical probabilities and likelihoods for the different selection statistic values observed in ZIP4 were estimated under each simulated selection scenario (see Table 2). Firstly, the empirical percentile in which each observation was found was estimated for each test (FST, dDAF, Tajima's D, Fay and Wu's H, Fu Li's D and XP-EHH) and scenario (neutral + constant recombination, neutral + hotspot recombination, low selection + constant recombination, medium selection + constant recombination, high selection + constant recombination, low selection + hotspot recombination, medium selection + hotspot recombination, high selection + hotspot recombination). This percentile was then subtracted from one if it was higher than 0.5 and multiplied by two to mimic a two-tailed test. Thus, if the observed value was found at the median of the simulated distribution, it yielded a probability of one. By contrast, if it was found in a tail of the distribution, it yielded a probability close to zero. For each scenario, we computed the combined empirical probability for several set combinations of observed neutrality test values by multiplying each corresponding empirical probability (Table S2). Each combination contained at least one neutrality statistic capturing each of the three main signatures of selection explored (population differentiation, haplotype structure or the site frequency spectrum). Next, empirical likelihoods were estimated as the ratio of the combined empirical probability under each selection scenario over the same probability under neutrality only for the hotspot recombination landscape observed in ZIP4 (Table 2). Likelihoods for the different combinations of statistics containing dDAF in Table S2 were nearly identical to the equivalent combinations obtained with FST (data not shown). As a conservative decision given the high correlation between FST and dDAF, we do not present the likelihood of any combination including both statistics. It is important to point out that any of the currently available human demographies in combination with coalescent simulators have relatively severe limitations mainly (i) in terms of the number of included populations (e.g. African populations) (ii) the accuracy and timing of the demographic events and (iii) the option to include selective sweeps as well as a defined recombination landscape. Therefore it is clear that the complexities of possible evolutionary scenarios (as discussed in the main text) are beyond what can be modeled by current approaches.
Neutrality tests on simulated and the 1000 Genomes population data were performed as described by Pybus et al. [68] and using the 1000 Genomes Selection Browser (http://hsb.upf.edu). Briefly, Tajima's D, Fu and Li's D and Fay and Wu's H were calculated using a sliding window approach with 30 kb windows and approximately 3 kb offset. FST [69] and XP-EHH [70] between CEU and YRI were calculated for each polymorphic position.
Human ZIP4 cDNA encoding the long isoform of the protein and the Ala357 and Leu372 variants was cloned into pcDNA 3.1 (+) expression vector together with a hemagglutinin (HA) tag at the carboxyl terminus as described previously [71]. The Leu372Pro and Leu372Arg mutants, as well as the Thr357Ala and Leu372Val polymorphisms, were introduced via site-directed mutagenesis following standard conditions (QuikChange II XL; Stratagene; see Table S4 for complete human cDNA and primers used in the mutagenesis). The six human ZIP4 isoforms obtained (i.e. Ala357-Leu372, Ala357-Val372, Thr357-Leu372, Thr357-Val372 as well as Ala357-Pro372 and Ala357-Arg372) were confirmed by sequencing with the ABiPrism 3.1 BigDye kit before their use in transfection experiments. HeLa cells were cultured in DMEM plus 10% FBS and, subsequently, each of the various ZIP4 forms were transiently transfected using polyethyleneimine as the transfection reagent (PolySciences).
For the cell surface expression experiments, live cells were incubated with anti HA (1∶1000) in DMEM without serum for 1h at 37° before fixation with 4% paraformaldehyde. After blocking for 30 min (1% BSA, 2% FBS in PBS), cells were incubated with a secondary antibody (1∶2000) for 45 min in the blocking solution. For the total cell expression experiments, cells were permeabilized with 0.1% Triton in PBS for 10 min after fixation. Following blocking for 30 min (1% BSA, 2% FBS in PBS), cells were incubated in the blocking solution with anti HA (1∶1000) for 1 h 30 min, washed with PBS and incubated with the secondary antibody (1∶2000) for 45 min. Images were acquired using an inverted Leica SP2 confocal microscope with a 40×1.32 Oil Ph3 CS objective. Expression was quantified by measuring chemiluminescence with a plate reader (24-well plates) using peroxidase-linked anti-mouse antibody (GE Healthcare) as a secondary antibody and SuperSignal West Femto reagent as a substrate (Thermo scientific). Data are presented as the ratio between surface expression and total expression of the transporter. Statistical significance was tested using standard ANOVA.
Cells were transiently transfected with the various ZIP4 isoforms plus empty ECFP vector for 24–36 h. Cytosolic Zn2+ signal was determined in CFP-positive cells loaded with FluoZin3 2.5 µM (Invitrogen) in a solution containing 140 mM NaCl, 5 mM KCl, 1.2 mM CaCl2, 0.5 mM MgCl2, 5 mM glucose, 10 mM HEPES, 300 mosmol/l, pH 7.4 for 20 min. Cytosolic [Zn2+] increases are presented as the difference with respect to the basal signal of emitted fluorescence (510 nm) after adding 200 µM ZnSO4 in a continuous perfusion bath. The kinetics of the various isoforms were calculated using a sigmoidal non-linear regression. In the same set of experiments, basal cellular Zn2+ content was estimated as the difference in FluoZin intensity between transfected cells and non-transfected cells before adding Zn2+ to the bath. Flourescence intensity was measured using an Olympus IX70 inverted fluorescence microscope, controlled by Aquacosmos software (Hamamatsu).
|
10.1371/journal.pmed.1002815 | Malaria morbidity and mortality following introduction of a universal policy of artemisinin-based treatment for malaria in Papua, Indonesia: A longitudinal surveillance study | Malaria control activities can have a disproportionately greater impact on Plasmodium falciparum than on P. vivax in areas where both species are coendemic. We investigated temporal trends in malaria-related morbidity and mortality in Papua, Indonesia, before and after introduction of a universal, artemisinin-based antimalarial treatment strategy for all Plasmodium species.
A prospective, district-wide malariometric surveillance system was established in April 2004 to record all cases of malaria at community clinics and the regional hospital and maintained until December 2013. In March 2006, antimalarial treatment policy was changed to artemisinin combination therapy for uncomplicated malaria and intravenous artesunate for severe malaria due to any Plasmodium species. Over the study period, a total of 418,238 patients presented to the surveillance facilities with malaria. The proportion of patients with malaria requiring admission to hospital fell from 26.9% (7,745/28,789) in the pre–policy change period (April 2004 to March 2006) to 14.0% (4,786/34,117) in the late transition period (April 2008 to December 2009), a difference of −12.9% (95% confidence interval [CI] −13.5% to −12.2%). There was a significant fall in the mortality of patients presenting to the hospital with P. falciparum malaria (0.53% [100/18,965] versus 0.32% [57/17,691]; difference = −0.21% [95% CI −0.34 to −0.07]) but not in patients with P. vivax malaria (0.28% [21/7,545] versus 0.23% [28/12,397]; difference = −0.05% [95% CI −0.20 to 0.09]). Between the same periods, the overall proportion of malaria due to P. vivax rose from 44.1% (30,444/69,098) to 53.3% (29,934/56,125) in the community clinics and from 32.4% (9,325/28,789) to 44.1% (15,035/34,117) at the hospital. After controlling for population growth and changes in treatment-seeking behaviour, the incidence of P. falciparum malaria fell from 511 to 249 per 1,000 person-years (py) (incidence rate ratio [IRR] = 0.49 [95% CI 0.48–0.49]), whereas the incidence of P. vivax malaria fell from 331 to 239 per 1,000 py (IRR = 0.72 [95% CI 0.71–0.73]). The main limitations of our study were possible confounding from changes in healthcare provision, a growing population, and significant shifts in treatment-seeking behaviour following implementation of a new antimalarial policy.
In this area with high levels of antimalarial drug resistance, adoption of a universal policy of efficacious artemisinin-based therapy for malaria infections due to any Plasmodium species was associated with a significant reduction in total malaria-attributable morbidity and mortality. The burden of P. falciparum malaria was reduced to a greater extent than that of P. vivax malaria. In coendemic regions, the timely elimination of malaria will require that safe and effective radical cure of both the blood and liver stages of the parasite is widely available for all patients at risk of malaria.
| Multidrug-resistant malaria results in recurrent parasitaemia, a cumulative risk of anaemia, and progression to severe and fatal disease.
Whilst artemisinin combination therapies (ACTs) and intravenous (IV) artesunate can reduce morbidity and mortality associated with P. falciparum malaria, they have no activity on the hypnozoite stages of P. vivax, which can relapse weeks to months following an initial infection and sustain ongoing transmission of the parasite.
In Papua, Indonesia, antimalarial resistance has emerged in both P. falciparum and P. vivax. We used a prospective malariometric surveillance network to investigate the differential impact of a universal policy of ACTs for uncomplicated malaria and IV artesunate for severe malaria on the morbidity and mortality attributable to P. falciparum and P. vivax.
After controlling for population growth and changes in treatment-seeking behaviour, the incidence of malaria after policy change fell by about 60% for P. falciparum and 40% for P. vivax.
There was a 50% fall in the proportion of patients with malaria requiring admission to hospital and a 30% fall in malaria-related mortality. Bed occupancy due to admission with malaria fell by 25%.
Whilst there was a small decrease in the absolute incidence of P. vivax infections over the 9-year study period, the proportion of malaria cases and malaria-attributable deaths to P. vivax increased with time.
The results highlight the importance of highly effective blood schizontocidal antimalarial drugs in reducing the overall burden of drug-resistant malaria.
The rising proportion of malaria due to P. vivax emphasizes the need for safe and effective drug regimens that clear both the blood and liver stages of P. vivax in malaria elimination efforts in coendemic regions.
| Prompt and effective treatment of malaria reduces morbidity and limits onward transmission of the Plasmodium parasite [1,2]. Large-scale use of highly efficacious antimalarial treatment regimens has contributed to significant reductions in P. falciparum malaria in many malaria-endemic regions [3,4]. P. vivax is more difficult to cure than P. falciparum because it forms dormant liver stages (hypnozoites) that are intrinsically resistant to standard schizontocidal drugs. Unless patients are treated with an effective drug regimen that clears both the blood and liver stage of the parasite, these hypnozoites can reactivate periodically, causing recurrent blood-stage infections (relapses) and ongoing transmission [5]. Malaria treatment campaigns that do not include radically curative primaquine regimens for patients infected with P. vivax may have only a modest effect on the number of cases of P. vivax malaria and thus are likely to be associated with an increase in the proportion of malaria due to this parasite compared to P. falciparum [6–8]. When primaquine radical cure is included in national guidelines, it is usually prescribed without prior testing for glucose-6-phosphate dehydrogenase (G6PD) deficiency. To mitigate the risks of drug-induced haemolysis, many countries recommend a 15-mg daily dose administered over 14 days despite evidence showing that 30 mg daily is more effective [9]. When supervised, a 14-day regimen of primaquine can reduce the risk of P. vivax relapse by more than 85% [10,11]; however, in most endemic settings, daily supervision of such a prolonged treatment regimen is not practical [12], and this can result in a significant reduction in primaquine adherence and effectiveness [13–15]. In a large-scale observational study of patients with vivax malaria in Papua, Indonesia, the effectiveness of unsupervised primaquine was estimated to be only 10% [16].
Malaria endemicity in Papua, Indonesia, varies from hypo- to hyperendemic for P. falciparum and P. vivax [17]. In the early 2000s, clinical trials and ex vivo drug-susceptibility testing demonstrated high-grade resistance to chloroquine and sulphadoxine + pyrimethamine in endemic P. falciparum strains and chloroquine resistance in P. vivax strains [18–20]. Frequent, recurrent parasitaemia is more likely in the setting of high-grade drug resistance and is associated with a cumulative risk of chronic anaemia, severe malaria, and mortality [21,22]. Therefore, in March 2006, Indonesian national antimalarial treatment guidelines were changed to an artemisinin combination therapy (ACT) (dihydroartemisinin plus piperaquine [DP]) for uncomplicated malaria due to any Plasmodium species and intravenous (IV) artesunate for severe malaria. At the same time, policy for the use of primaquine in patients with P. vivax infections was changed from a total dose of 3.5 mg/kg over 14 days to a higher dose of 7 mg/kg over 14 days. Information regarding the treatment changes was distributed widely via health professionals and community leaders. In Mimika District, located in southern Papua Province, hospital and community surveillance systems were put into place prior to the policy change to allow an assessment of the subsequent changes in malaria-attributable morbidity and mortality due to either P. falciparum or P. vivax. Previous analyses from the same location have quantified the burden of malaria in the hospital, the epidemiology of malaria in the community, and local treatment-seeking behaviour [17,22,23]. The current analysis used routinely collected surveillance data collected over a 9-year period (2004 to 2013) to investigate the temporal trends in malaria morbidity and mortality before and after the change in antimalarial treatment policy and the relative impact of this intervention on the burden of P. falciparum compared with P. vivax.
The geography, climate, and demographics of Mimika District and its capital, Timika, have been described previously [17,22]. Briefly, Mimika District lies in south central Papua, eastern Indonesia, and covers an area of 21,522 km2; it has 12 subdistricts and 85 villages (Fig 1). The region has fragmented forest ranging from extensive coastal lowlands to high mountainous environments. Timika has a growing population of native Papuans and Indonesian migrants, estimated to be 120,457 in 2004 and increasing to 196,401 in 2013 [24].
Malaria transmission is perennial with minimal seasonal variation and is normally restricted to the lowland areas below 1,600-m elevation, where most of the population now resides. There are three primary mosquito vectors: Anopheles koliensis, two members of the A. farauti complex, and A. punctulatus; all of these vectors are both exo- and endophilic and are primarily opportunistic in host-seeking behaviour. In 2005, the point prevalence of parasitaemia was estimated to be 16.3%: 46% due to P. falciparum, 39% due to P. vivax, and 11% due to mixed P. falciparum/P. vivax infections [17]. Local P. vivax strains have a typical equatorial relapse periodicity of 3–4 weeks [16,25].
Clinical trials conducted in Timika in 2004 and 2005 demonstrated failure of chloroquine and sulphadoxine + pyrimethamine for the treatment of uncomplicated falciparum malaria with a recurrence rate following combination treatment of 48% at day 28 [18]. Recurrence of P. vivax at day 28 post chloroquine monotherapy was even higher at 65% [18]. Ex vivo studies confirmed high-grade resistance to these drugs [19].
Formal healthcare facilities in the district include Rumah Sakit Mitra Masyarakat (RSMM), a 110-bed hospital funded by a local mining company providing healthcare free of charge to indigenous Papuan communities and at a nominal cost to non-Papuan Indonesians. Additionally, there are 12 government-funded community primary health clinics (puskesmas) and 10 clinics administered by the local mining company. A government-funded hospital (Rumah Sakit Umum Daerah [RSUD]) opened at the end of 2008 but did not begin seeing substantial numbers of malaria patients until 2010. Antimalarial treatment can also be bought at a wide range of regulated and unregulated private sector clinics and facilities in Timika [23].
During the early 2000s, patients with uncomplicated malaria presenting to community clinics were treated with chloroquine plus sulphadoxine + pyrimethamine if they had falciparum malaria and with chloroquine plus low-dose primaquine (total dose 3.5 mg/kg) if they had vivax malaria. At the RSMM hospital, patients with uncomplicated malaria were treated with oral quinine, whereas patients with severe malaria received IV quinine. Those with P. vivax malaria also received unsupervised low-dose primaquine [16]. In March 2006, the policy for treatment of uncomplicated malaria due to any Plasmodium species at all public community clinics and the hospital was changed to DP, a regimen with a risk of P. falciparum recrudescence of 4.4% and P. vivax recurrence of 10% by day 42 [26]. At the same time, treatment of severe malaria was changed from IV quinine to IV artesunate [27], and the dose of unsupervised primaquine was doubled to 7 mg/kg divided over 14 days. The new unified antimalarial treatment policy was adopted by RSUD hospital upon opening. Healthcare providers disseminated information regarding the antimalarial policy change and the benefits of DP to clinics and to communities via village leaders.
All patients presenting to one of the formal sector community (puskesmas) clinics with symptoms consistent with malaria had capillary blood collected for blood film examination or had a rapid diagnostic test prior to antimalarial treatment. Between April 2004 and December 2009, weekly reports on the number of blood film examinations and the number of patients treated for malaria were collated by the district health authority. These reports were aggregated by the species of infection within four age bands: <1 year, 1 to 5 years, 5 to 10 years, and those older. Malariometric surveillance data were not collected from healthcare facilities in the private sector.
Cluster-randomized, cross-sectional surveys to determine treatment-seeking behaviour were conducted in 2005 and again in 2013, using an identical sampling strategy [17,23]. In 2005, during the pre–policy change period, 45.7% (349/764) of patients with malaria presented to public sector facilities and would have been detected by the surveillance network, but in 2013, 6 years after policy change, this figure had risen to 67.3% (66/98; p < 0.001) [28]. The shifts in treatment-seeking behaviour were apparent in all age groups. In the first household survey, 32.3% (10/31) of members who died of any cause within the preceding year did so at RSMM hospital compared to 26.3% (5/19) in 2013 (p = 0.656).
All patient presentations to RSMM between 2004 and 2013 (whether to the outpatients department, emergency department, or inpatient wards) were recorded by hospital administrators in a QPro database. Patients were identified using a unique hospital reference number. Demographic data and the clinical diagnoses assigned by the attending physician were collected. Drug prescriptions and results of full blood-count analyses from RSMM’s Coulter counter were recorded in separate databases and identified by the same individual hospital reference number. RSMM policy dictates that all outpatients with fever or other signs or symptoms consistent with malaria and all inpatients regardless of presentation have a thick film prepared for malaria microscopy. Thin films and rapid diagnostic tests were done after-hours when the laboratory was closed. At RSUD hospital, a malaria register was initiated in January 2010, documenting aggregated data on basic demographic details, infecting Plasmodium species, fulfilment of clinical criteria for severe malaria, admission status, and malaria-related death in patients at the hospital.
Since 2002, vector-control activities in the region have been provided predominantly by the Public Health Malaria Control (PHMC) programme. These have included twice-yearly indoor residual spraying (IRS) and distribution of long-lasting insecticide-treated bed nets (LLINs) covering 10%–20% of households. Vector-control activities remained relatively constant from 2004 until mid-2013, when a large-scale IRS campaign and bed net distribution was commenced.
PHMC maintained entomological surveillance at five routine ‘sentinel’ sites representative of key locations in lowland Mimika from 1996 onwards. At each site, human-landing collections (HLCs) of mosquitoes were conducted at least once per week by 2 to 5 trained collectors for 5 to 10 hours during evening hours. Captured mosquitoes were examined and identified by microscopy and their species recorded based on key morphological characters. Data were combined to derive a human biting index over time. Automated daily rainfall was recorded at Kuala Kencana township, representing one of the 12 subdistricts in Mimika. Routine entomology and meteorological data were available from January 2004 until June 2009.
The analysis was conducted according to an a priori statistical plan (S1 Text) and is reported according to RECORD (S1 RECORD Checklist). Additional multivariable regression analyses, requested at statistical review, were also undertaken of key outcomes controlling for population size, vector biting, and monthly time trends. Data from the hospitals, community clinics, and meteorological and entomological surveillance were aggregated by calendar month. The surveillance period was divided into four periods: pre–policy change (April 2004 to March 2006), early transition (April 2006 to March 2008, an equal interval to that observed before policy change), late transition (April 2008 to December 2009, corresponding to the end of the community and entomology surveillance), and post transition (January 2010 to December 2013, corresponding to the period between the opening of the new RSUD hospital and the end of the study period) (S1 Fig, S1 Table).
Estimated malaria incidence rates were derived for the pre-policy and the early and late transition periods using absolute numbers of malaria cases from both the hospital and community surveillance, the estimated population at the time (assuming linear growth between the censuses), and estimates of the proportion of febrile patients seeking treatment within the malariometric surveillance system, obtained from the two household surveys in 2005 and 2013 (S1 Fig). In a sensitivity analysis, the incidence of malaria was also derived assuming no change in treatment-seeking behaviour. Analyses of malaria-related morbidity and mortality were limited to RSMM data, as the necessary information was not collected from the other sites. To account for monthly time trends, vector biting, and population growth, Poisson regression analyses of the data pre-2009 were performed to estimate the adjusted incidence rate ratios (IRRs) for falciparum and for vivax malaria for the late transition period versus the pre–policy change period. Similar analyses were undertaken for hospital admissions using binomial regression to estimate the risk ratio.
All graphing and statistical analysis was done in STATA version 15.1 (StataCorp, College Station, TX, United States). Temporal trends in outcomes were presented graphically over the entire study period. Comparisons of outcomes before and after policy change were made using medians (with interquartile ranges [IQRs]), proportions (n/N with absolute differences and binomial 95% confidence intervals [CIs]), incidence rates (per 1,000 person-years [py] with IRRs and Poisson 95% CIs). Comparisons of prospectively collected community and hospital surveillance data were restricted to the pre-policy and late transition periods to ensure inclusion of the community surveillance (which ended in December 2009) and to avoid bias associated with the opening of RSUD (which began admitting significant numbers of malaria patients in January 2010). p-Values were not presented, since numbers of cases from the community clinics and hospitals were large and statistical significance was achieved even in the absence of clinical significance.
Ethical approval for this study was obtained from the Health Research Ethics Committees of the University of Gadjah Mada, Indonesia (KE/FK/544/EC), and Menzies School of Health Research, Darwin, Australia (HREC 10.1397).
Between April 2004 and June 2009, the mean daily rainfall was 27.8 mm with small peaks in mid-2005, mid-2007, and mid-2008 (Fig 2, S1 Data). Over the same period, 204,968 human-hours of nighttime mosquito collections were conducted; the estimated mean number of anopheline bites per py was 145 (range 28 to 534) (Fig 2, S1 Data). There was a large peak in the number of mosquitoes caught throughout much of 2007; the mean estimated number of bites per py during this period was 238 (range 34 to 534). Overall, A. koliensis accounted for 85.0% (6,992/8,222) of mosquitoes captured by HLC, compared to 11.6% (956/8,222) for A. farauti species complex and 3.3% (274/8,222) for A. punctulatus.
Data were gathered from 12 community outpatient facilities over a period of 69 months (April 2004 to December 2009). A total of 671,386 blood films were examined, of which 193,566 (28.8%) were positive for malaria; 98,530 (50.9%) were due to P. falciparum, 87,632 (45.3%) were due to P. vivax, 3,308 (1.7%) were due to P. malariae, and 4,096 (2.1%) were mixed species infections (Table 1). Slide positivity for P. falciparum declined from 14.8% (37,304/251,286) before treatment policy change to 13.0% (25,362/194,792) in the late transition period, a difference of −1.8% (95% CI −2.0% to −1.6%). Over the same period, the slide positivity for P. vivax increased from 11.5% (28,872/251,286) to 14.8% (28,861/194,792), difference = 3.3% (95% CI 3.1%–3.5%). The corresponding proportion of malaria infections due to P. vivax rose from 44.1% (30,444/69,098) to 53.3% (29,934/56,125), difference = 9.2% (95% CI 8.7%–9.8%) (Fig 3, S2 Data). In multivariable analyses controlling for population size, vector biting, and monthly time trends, the IRR for the late transition period compared with the pre–policy change period was 0.70 (95% CI 0.67–0.74) for falciparum malaria and 1.02 (95% CI 0.97–1.07) for vivax malaria.
The proportion of all blood films read that were positive for P. falciparum gametocytes fell steadily over the study period from 1.4% (3,491/251,286) pre–policy change to 0.7% (1,419/194,792) in the late transition period, a difference of −0.7% (95% CI −0.7% to −0.6%) (Fig 3). Over the same time interval, there was an increase in the overall gametocyte slide positivity for P. vivax, which rose from 1.9% (4,700/251,286) to 2.3% (4,403/194,792), a difference of 0.4% (95% CI 0.3%–0.5%).
Data from RSMM were available for 117 months (April 2004 to December 2013). Overall, there were 1,054,674 patient presentations to the hospital, of which 196,380 (18.6%) were associated with malaria, 100,078 (51.0%) with P. falciparum, 65,306 (33.3%) with P. vivax, 5,097 (2.6%) with P. malariae, 120 (0.06%) with P. ovale, and 25,779 (13.1%) with mixed species infections. In total, 27,890 (14.2%) of the patients with malaria required admission to hospital, and 595 (0.3%) died in hospital. In the posttransition period, 22.5% (27,480/122,232) of malaria diagnosed at a tertiary facility was at the newly opened RSUD hospital (S3 Data); therefore, before-and-after comparisons at the RSMM hospital were only made between the pre–policy change era and the late transition period.
Before policy change, the most commonly prescribed blood schizontocide at RSMM hospital was oral quinine (20,364/24,538; 83.0%) (Fig 4); thereafter, DP was prescribed in 88.6% (139,002/156,902) and artesunate-amodiaquine in 6.0% (9,442/156,902) of malaria cases. Patients requiring IV therapy were prescribed quinine in 83.1% (3,858/4,641) of cases before policy change, but thereafter, 98.5% (16,414/16,661) were treated with IV artesunate. Before policy change, 63.5% (4,455/7,019) of patients with P. vivax malaria were treated with a 14-day primaquine regimen, and after policy change, this rose to 71.1% (51,344/72,196).
The proportion of all presentations to RSMM that were related to malaria rose from 16.5% (28,789/174,289) before policy change to 18.3% (34,117/186,312) in the late transition period (difference of 1.8% [95% CI 1.5%–2.0%]) (Fig 5). This was driven by a rise in the proportion of outpatients with malaria, which increased from 13.6% (21,044/155,029) to 17.3% (29,331/169,391), a difference of 3.7% (95% CI 3.5%–4.0%). Over the same period, the proportion of inpatients with malaria fell from 40.2% (7,745/19,260) to 28.3% (4,786/16,921; difference of −11.9% [95% CI −12.9% to −11.0%]), and the proportion of patients with malaria requiring admission fell from 26.9% (7,745/28,789) to 14.0% (4,786/34,117; difference of −12.9% [95% CI −13.5% to −12.2%]) (Fig 6, Table 1 and S3 Data). In multivariable analyses comparing the pre–policy change and late transition period, after controlling for monthly trends, the proportion of inpatients with malaria decreased by 0.56-fold (95% CI 0.51–0.60), and the proportion of patients with malaria requiring admission fell by 0.82-fold (95% CI 0.75–0.90).
Of the patients admitted with malaria, the median length of stay decreased from 3 days (IQR 2–4) to 2 days (IQR 2–4), and this was associated with a fall in the median total monthly inpatient bed occupancy due to malaria from 1,033.5 days (IQR 897–1,251.5) pre–policy change to 769.5 days (IQR 685.5–856) post policy change (Fig 6). Overall, the proportion of malaria cases due to P. vivax mono- or mixed species infection rose from 32.4% (9,325/28,789) to 44.1% (15,035/34,117), a difference of 11.7% (95% CI 10.9%–12.4%) (Fig 7). Malaria due to any Plasmodium species collectively accounted for 57.7% (4,388/7,603) of all severe anaemia at the hospital before policy change and 41.7% (1,928/4,627) in the late transition period, a difference of −16.0% (95% CI −17.8% to −14.2%) (Fig 7).
Before policy change, malaria accounted for 15.5% (137/886) of all deaths at RSMM, with 0.48% (137/28,789) of patients who presented with malaria dying during their hospital encounter. In the late transition period, the corresponding figures were 10.4% (100/961; difference = −5.1% [95% CI −8.1 to −2.0]) and 0.29% (100/34,117; difference = −0.18% [95% CI −0.28 to −0.08]) (Fig 6, S3 Data). Over the same period, there was a significant fall in the mortality attributable to P. falciparum (0.53% [100/18,965] versus 0.32% [57/17,691]; difference = −0.21% [95% CI −0.34 to −0.07]), but this was not apparent for P. vivax (0.28% [21/7,545] versus 0.23% [28/12,397]; difference = −0.05% [95% CI −0.20 to 0.09]) or nonmalarial disease (0.51% [749/145,500] versus 0.57% [861/152,195], difference = 0.05% [95% CI −0.002 to 0.10]). Overall, P. vivax accounted for 15.3% (21/137) of malaria-related deaths pre–policy change and 28.0% (28/100) in the late transition period, a difference of 12.7% (95% CI 2.0%–23.3%) (S3 Data).
Assuming that 30% of deaths in the region occurred at RSMM hospital, the overall malaria-attributable mortality rate in the population of Mimika District fell from 1.90 (95% CI 1.73–2.08) per thousand py before the policy change to 1.29 (95% CI 1.15–1.43) per 1,000 py in the late transition period, a rate difference of −0.61 (95% CI −0.83 to −0.39) per 1,000 py.
When community and hospital surveillance data were combined, a total of 295,971 cases of malaria were diagnosed between April 2004 and December 2009. The overall incidence of malaria was 406 (95% CI 404–409) per 1,000 py before policy change, 372 (95% CI 370–374) per 1,000 py in the early transition period, and 351 (95% CI 349–354) per 1,000 py in the late transition period (Table 2, S4 Data). Assuming that 45.7% of patients with malaria were detected by the surveillance network pre–policy change and 67.3% in the late and posttransition periods, the overall incidence of P. falciparum fell from 511 to 249 per 1,000 py (IRR = 0.49 [95% CI 0.48–0.49]), whereas the incidence of P. vivax fell from 331 to 239 per 1,000 py (IRR = 0.72 [95% CI 0.71–0.73]). In a sensitivity analysis assuming no shift in treatment-seeking behaviour, the incidence of P. falciparum fell from 234 to 168 per 1,000 py (IRR = 0.72 [95% CI 0.71–0.73]), whereas the incidence of P. vivax rose from 152 to 161 per 1,000 py (IRR = 1.06 [95% CI 1.05–1.08]).
In March 2006, Indonesia was the first malaria-endemic country to adopt a unified schizontocidal treatment policy for malaria due to any Plasmodium species: DP for uncomplicated malaria and IV artesunate for severe malaria. These changes came on a background of failing treatment regimens due to high-grade multidrug resistance in both P. falciparum and P. vivax species. In Mimika District, southern Papua, the uptake of the new policy was rapid, with the new treatment regimens adopted into practice in most public health facilities within a month. In this high-transmission setting, we found that the implementation of highly effective antimalarial treatment regimens was associated with a marked reduction in both malaria-related morbidity and mortality. The incidence of malaria and the proportion of malaria requiring admission to hospital fell by one-half, bed occupancy of patients with malaria fell by 26%, and malaria-related mortality fell by one-third. Associated with these changes, the proportion of patients with malaria attributable to P. vivax increased from 41% to 54%, and the proportion of malaria-related deaths attributable to P. vivax rose from 15% to 28%.
Our study highlights the complexity of defining temporal changes in the burden of malaria at a population level over a long period of time [1,2,29]. During the 9 years of surveillance, there was a substantial increase in the local population, fluctuations in rainfall and vector mosquito numbers, and marked changes in healthcare provision and treatment-seeking behaviour. Our comprehensive surveillance quantified patient numbers at the only lowland hospital (until late 2009) and all public and mine-supported clinics, which collectively diagnosed and treated almost half a million cases of malaria over the study period. Although these clinics offered healthcare for free or at a nominal cost, our treatment-seeking surveys suggested that initially only 46% of patients with malaria sought treatment in the public sector. However, following the change in policy, there was a significant shift in behaviour, with 67% of patients seeking treatment in the public sector where they could access DP, a drug perceived to be highly effective compared to previously available treatments [28]. This shift in behaviour, along with the rise in the total population and an increase in vector numbers in 2007 (Fig 2), may have contributed to the initial surge in malaria cases and slide positivity observed in the community and hospital outpatients department in the early transition period but a subsequent fall in these metrics in the late transition period as the impact of ACT on P. falciparum began to manifest (Fig 3). Conversely, in the latter part of the study, a new public hospital facility (RSUD) outside of the initial surveillance network was opened and assumed care of approximately 20% of malaria inpatients, acting to artificially decrease the burden of malaria at RSMM hospital. To control for these confounding factors, the overall temporal trends by month are presented, but the comparisons pre–and post–policy change were restricted conservatively to the period immediately before policy change and the late transition period, prior to the opening of the new hospital facility.
The replacement of failing treatment regimens with highly efficacious schizontocidal treatment has potential to reduce malaria-related morbidity and mortality and decrease the risk of recurrent parasitaemia and ongoing transmission [30]. In Papua in 2005, the efficacy of DP against P. falciparum and P. vivax was greater than 95%, with antimalarial efficacy sustained against both species throughout the study period [26,31]. In the same year, a clinical trial of patients with severe malaria at the RSMM hospital demonstrated that IV artesunate reduced associated mortality by 35% compared to IV quinine [27]. Our analysis highlights that, following implementation of both of these artemisinin-based treatment strategies in April 2006, there was a significant reduction in malaria-related morbidity and mortality, which was most apparent at the RSMM hospital. Whilst outpatient numbers actually increased over the study period, both the absolute number of malaria admissions and the proportion of malaria patients requiring admission fell (the latter from 27% to 14%). This was associated with a marked reduction in total bed occupancy due to malaria, shorter admission times, and a lower risk of severe anaemia- and malaria-related mortality. In total, 264 bed days per month were made available at the hospital for the treatment of other diseases. These findings are consistent with African studies that have shown that the most prominent impact of enhanced malaria control activities is a reduction in severe malaria and mortality [32,33]. The variation in malaria morbidity was less marked in the community, where absolute numbers of patients remained high and where there was only a modest fall in malaria prevalence from 16.3% to 12.2%. However, after accounting for population growth and shifts in treatment-seeking behaviour, the estimated overall incidence of malaria in the community also fell significantly.
A consistent finding in both the hospital and community setting was the marked increase in the proportion of malaria caused by P. vivax. Policy change was associated with a differential variation between species in the overall cases of malaria, severe disease, and gametocyte carriage. At the start of the study, P. vivax accounted for 32% of all malaria at the hospital and 44% in the community. By the late transition period, P. vivax was the predominant cause of malaria, accounting for 54% of all malaria cases. Whereas the overall risk of mortality in patients presenting with P. falciparum fell from 0.53% to less than 0.25% in the posttransition period, there was no fall in mortality associated with P. vivax, possibly reflecting a higher likelihood of concomitant nonmalarial morbidities in severely ill patients with P. vivax malaria [34,35]. In the community, the proportion of patients with P. falciparum gametocytes on blood film examination halved (from 1.4% to 0.7%), whereas the proportion of patients with P. vivax gametocytes remained unchanged at about 2% (Fig 3).
P. vivax is less amenable than P. falciparum to control by enhanced or scaled-up antimalarial treatment efforts [6,7,36]. Mass drug administration that does not include antirelapse therapy has little effect on P. vivax [3]. There are several biological reasons for this refractoriness. Firstly, P. vivax gametocytes appear early during the course of an infection and are therefore more likely than P. falciparum gametocytes, which appear later, to have been transmitted to mosquitoes prior to antimalarial treatment. Secondly, failure to sterilize the liver of hypnozoites can result in multiple subsequent relapses and thus much greater transmission potential from a single inoculation of P. vivax as compared with P. falciparum. Thirdly, immunity to P. vivax develops early in endemic regions because of the high force of infection from relapses [22,37]. This results in a large pool of asymptomatic patients who probably still harbour gametocytes and therefore remain infectious to mosquitoes [38,39].
Although DP provides posttreatment prophylaxis against P. vivax recurrence for up to 42 days, it has minimal effect on P. vivax relapses thereafter. At the same time as the change in blood schizontocidal treatment policy, the recommended dose of primaquine for radical cure of P. vivax infections was revised from a total dose of 3.5 mg/kg to 7 mg/kg to treat the relatively primaquine-tolerant P. vivax strains in Papua [40]. If safely and effectively delivered, high-dose primaquine regimens should produce a significant reduction in P. vivax transmission. However, in Timika, the provision of unsupervised high-dose primaquine combined with DP has, at best, a modest effect on the likelihood of representation to hospital with vivax malaria [16]. We postulate that nonadherence to unsupervised primaquine is one of the most likely explanations for the minimal decline in P. vivax incidence in the region.
Our study has several important strengths. The multifaceted surveillance system incorporated prospectively collected community and hospital data along with before-and-after cross-sectional surveys, providing a means of checking the consistency of the findings across various settings and thus increasing confidence in the internal validity of our results. Data on nonmalaria presentations to hospital enabled us to control the hospital data for population growth by presenting the proportional burden of malaria over time. High-quality microscopy services both at the hospital and in the community clinics are maintained by accredited microscopists, whose performance is reviewed regularly [22]. Consistent procedures for examination and reporting of blood films therefore support the validity of longitudinal assessments of Plasmodium species distributions observed in our analysis. Individualized clinical data at RSMM with linkage to pharmacy and haematology data enabled a very large-scale assessment of the temporal trends in malaria-related morbidity (in addition to incidence) in the hospital population.
Our study also has some significant limitations. Estimates of the local population were imprecise because of large transient migrant groups who were excluded from censuses. This was an issue particularly in the latter years, when the population at risk of malaria may have been underestimated; thus, our estimates of the reduction in malaria incidence are likely to be conservative. The estimated change in the proportion of malaria patients detected by the surveillance network was derived from two assessments of treatment-seeking behaviour in 2005 and 2013 [28]. In a sensitivity analysis, assuming no shift in treatment-seeking behaviour, the estimated reduction in the incidence of falciparum malaria was only 28%, with a 6% rise in the incidence P. vivax malaria. However, changes in treatment-seeking behaviour should not have confounded estimates of the proportion of malaria due to P. vivax or the reduction in hospital-related morbidity. The surveillance system captured changes in human–mosquito attack rates and climate variability but did not record concurrent vector-control interventions. Throughout most of the study period, formal vector-control activities were supported by the mine-supported PHMC program. These activities remained relatively constant, apart from a large bed net distribution and IRS program in urban Timika that commenced in 2013, towards the end of the posttransition period. Changes in vector-control measures were therefore unlikely to have influenced to a significant extent the reduction in malaria incidence between the pre–policy change and late transition periods. Finally, in view of the variations in treatment-seeking behaviour and healthcare provision, an a priori decision was made to compare the malarial outcomes between the pre-policy and the late transition periods. Although comparison of dichotomised temporal data can result in loss of power and in type I error, multivariable regression analyses of unconstrained monthly data confirmed the observed trends and demonstrated that they were still apparent after controlling for population growth and vector biting.
In summary, in Papua, Indonesia, a change in antimalarial treatment policy from failing drugs to highly efficacious artemisinin-based treatment for both uncomplicated and severe malaria was associated with a modest reduction in the overall incidence of malaria but a significant reduction in malaria-related hospital admissions and mortality. In this area coendemic for both P. falciparum and P. vivax, there was a marked increase in the proportion of malaria attributable to P. vivax. Additional scale-up of the existing treatment strategy is likely to result in further reductions in P. falciparum transmission; however, in order to reduce P. vivax transmission significantly, antirelapse therapy will need to be delivered more effectively. Novel strategies are being developed to improve primaquine adherence through community education campaigns and directly observed supervision of treatment [41]. The availability of point-of-care testing for G6PD deficiency raises the possibility of introducing short-course high-daily-dose primaquine regimens and tafenoquine that will facilitate further adherence to a complete course of treatment [42,43]. In coendemic regions, access to safe and effective radical cure of malaria for all patients at risk of malaria will be critical for the timely elimination of malaria.
|
10.1371/journal.pntd.0003321 | Th1-Biased Immunomodulation and Therapeutic Potential of Artemisia annua in Murine Visceral Leishmaniasis | In the absence of vaccines and limitations of currently available chemotherapy, development of safe and efficacious drugs is urgently needed for visceral leishmaniasis (VL) that is fatal, if left untreated. Earlier we reported in vitro apoptotic antileishmanial activity of n-hexane fractions of Artemisia annua leaves (AAL) and seeds (AAS) against Leishmania donovani. In the present study, we investigated the immunostimulatory and therapeutic efficacy of AAL and AAS.
Ten-weeks post infection, BALB/c mice were orally administered AAL and AAS for ten consecutive days. Significant reduction in hepatic (86.67% and 89.12%) and splenic (95.45% and 95.84%) parasite burden with decrease in spleen weight was observed. AAL and AAS treated mice induced the strongest DTH response, as well as three-fold decrease in IgG1 and two-fold increase in IgG2a levels, as compared to infected controls. Cytometric bead array further affirmed the elicitation of Th1 immune response as indicated by increased levels of IFN-γ, and low levels of Th2 cytokines (IL-4 and IL-10) in serum as well as in culture supernatant of lymphocytes from treated mice. Lymphoproliferative response, IFN-γ producing CD4+ and CD8+ T lymphocytes and nitrite levels were significantly enhanced upon antigen recall in vitro. The co-expression of CD80 and CD86 on macrophages was significantly augmented. CD8+ T cells exhibited CD62Llow and CD44hi phenotype, signifying induction of immunological memory in AAL and AAS treated groups. Serum enzyme markers were in the normal range indicating inertness against nephro- and hepato-toxicity.
Our results establish the two-prong antileishmanial efficacy of AAL and AAS for cure against L. donovani that is dependent on both the direct leishmanicidal action as well as switching-on of Th1-biased protective cell-mediated immunity with generation of memory. AAL and AAS could represent adjunct therapies for the treatment of leishmaniasis, either alone or in combination with other antileishmanial agents.
| Visceral leishmaniasis (VL) is a fatal, vector-borne tropical disease that affects the poorest sections of the society. The currently available drugs are toxic, expensive and have severe side effects. The problem is further compounded by emergence of VL-HIV co-infection and occurence of PKDL after apparent cure. Thus, alternate therapeutic interventions are needed in the absence of vaccines and mounting drug resistance. VL is also characterized by severe depression of cell-mediated immunity that complicates the efficiency of chemotherapeutic drugs. Restoration of the dampened immune system coupled with antileishmanial effect would be a rational approach in the quest for antileishmanial drugs. Plant derived secondary metabolites have been recommended for the containment of antiparasitic disease including leishmaniasis that synergistically aid in lifting the immune suppression. We previously reported in vitro antileishmanial activity of n-hexane fractions of Artemisia annua leaves (AAL) and seeds (AAS) that was mediated by apoptosis. In this study, we found significant reduction in liver and spleen parasite burden of Leishmania donovani infected BALB/c mice upon oral administration of AAL and AAS with concomitant immunostimulation and induction of immunological memory. The immunotherapeutic potentiation by AAL and AAS with no adverse toxic effects validates their use for treatment of this debilitating disease.
| Protozoal infections are a worldwide health problem, particularly in the third world countries [1]–[2], and account for approximately 14% of the world's population, who are at risk of infection. Leishmaniasis is considered by the WHO as one of the six major infectious diseases, with a high incidence and ability to produce deformities [3]–[4]. Therefore, finding a safe, effective and affordable treatment for such neglected tropical syndromes is a major concern and of high priority [3]. There are two main forms of leishmaniasis: cutaneous, characterized by skin sores; and visceral, which affects the internal organs (e.g. the spleen, liver, and bone marrow). Visceral leishmaniasis (VL) is the more severe form, causing significant morbidity and mortality, if left untreated. In the current scenario, the disease is associated with the high cost of treatment and poor compliance. In addition, drug resistance, low effectiveness and poor safety have been responsible for retarding the treatment efficacy of current chemotherapy [5]. Concomitant infection with malaria or pneumonia increases the fatality of the illness if not diagnosed and treated in time. The problem of leishmaniasis has been worsened due to parallel infections in AIDS patients [6]–[7]. In the absence of a credible vaccine, there is an urgent need for effective drugs to replace or supplement those in current use. The pentavalent antimony compounds, which constitute the first line of drugs for treatment of leishmaniasis were developed before 1959. The resistance to these drugs is now widespread in Bihar, India where 50–65% patients fail to be treated successfully with normal dose schedule of these first line drugs [8]. The new drugs that have become available in recent years for the treatment of VL are AmBisome, the excellent but highly expensive liposomal formulation of Amphotericin B (AMB) and the oral drug miltefosine, which has now been registered in India. The toxicity of these agents and the persistence of side effects even after modification of the dose level and duration of treatment are, however, severe drawbacks. Drug combinations like miltefosine/paromomycin and SbIII/paromomycin are also ineffectual, as Leishmania donovani is known to easily develop resistance [9]. In spite of rapid advances in synthetic chemistry that promises to offer new drugs, natural products continue to play an important role in therapy: Of the 1,184 new drugs registered between 1981 and 2006, 28% were natural products or their derivatives. Another 24% of the new drugs had pharmacophores (i.e., functional groups with pharmacological activity) derived from natural products [10]. Thus, a good starting point to find anti-parasitic natural products would be traditional medicinal plants that have been employed to treat infections, in Asia, Africa or America [11]. For both good scientific reasons and strong pragmatism, the WHO also advocates the use of traditional medicines for the treatment of these tropical diseases [12]. In the quest for new antileishmanial agents with negligible adverse effects, it was thus imperative to focus on alternative systems of medicine [7] including anti-parasitic plant extracts or secondary metabolites derived from them, as an alternative to synthetic drugs.
It is well documented that a defective cell mediated immune response marks the progression of leishmaniasis and restoration of cellular immunity is critical to disease control [13]. The CD4+ as well as CD8+ T cells have been implicated in resolution of infection [14]. The Th1/Th2 dichotomy of CD4+ T cells is also evident in murine VL where the active diseased state is marked by a predominance of Th2 response whereas protection or cure is denoted by a strong Th1 response [15]. Further, recovery from the disease and resistance to reinfection is attributed to generation of long lasting immunological memory, which is dependent upon parasite specific memory T cells [16]. There is substantial evidence signifying that the immune system synergistically promotes the therapeutic efficacy of antiparasitic drugs [17]. Therefore, antileishmanial drugs that can quickly reverse the immune suppression of the infected host and polarize the response towards Th1 phenotype with generation of immunological memory, besides killing the parasites, are desirable.
In the context of our study, many traditional medicinal plants have been shown to possess dual antileishmanial and immunopotentiating activities validating their use in folk medicine [18], [19]. Artemisia annua (Asteraceae), a well-known traditional medicinal plant, has been extensively used as antimalarial [20]–[21] and anticancer agent [22]. Recently the in vitro and in vivo efficacy of artemisinin (one of the constituents of A. annua, A. indica and A. dracunculus) against hepatocellular carcinoma [23] and experimental VL has been reported [24]. Flavonoids of A. annua have been linked to beneficial immunomodulatory activities in subjects affected from parasitic and chronic diseases [25]. The in vitro and ex vivo leishmanicidal activity of the A. annua leaves (AAL) and seed extracts (AAS) has been evaluated previously against L. donovani promastigotes and intracellular amastigotes by our group [26]. In the present study, we have explored the immunotherapeutic potential of AAL and AAS against VL in L. donovani infected BALB/c mice.
Female BALB/c mice aged 6–8 weeks and weighing 20–25 g were used in the present study after prior approval from the Jamia Hamdard Animal Ethics Committee (JHAEC) for the study protocol (Ethical approval judgment number is 459). JHAEC is registered under the Committee for the purpose of supervision and control of experiments on animals (CPCSEA). All animals were individually housed in the Central Animal House of Jamia Hamdard as per internationally accepted norms. The mice were kept in standard size polycarbonate cages under controlled conditions of temperature (23 ± 1°C), humidity (55 ± 10%), 12:12 h of light and dark cycle and fed with standard pellet diet (Ashirwad Industries, Chandigarh, India) and filtered water (ad libitum).
Fresh A. annua leaves and dried seeds with floral parts were collected from the Herbal Garden of Jamia Hamdard, washed, air-dried and ground separately and extracted with n-hexane as described previously [26]. The n-hexane extract of leaves (AAL) and seeds (AAS) were concentrated to dryness under reduced pressure at 35°C using a rotary evaporator and the semisolid paste further concentrated in a vacuum dessicator. Dosing solutions were prepared aseptically in dimethyl sulphoxide (DMSO, cell culture grade), and diluted further in PBS (0.02 M phosphate buffered saline, pH 7.2) to achieve a final DMSO concentration not exceeding 0.2%, which is non-toxic. All reagents including AAL and AAS were free of lipopolysaccharide (0.2 ng/ml endotoxin) as determined by the Limulus amoebocyte lysate assay.
Leishmania donovani (MHOM/IN/AG/83) promastigotes were grown in M199 medium, supplemented with 10% FBS, 2 mM glutamine, 100 units ml−1 penicillin, and 100 µg ml−1 streptomycin sulfate at 22°C. Late stationary phase promastigotes were obtained after incubation of the parasites for 4–5 days with starting inoculum of 1×106 parasites ml−1 [27].
Stationary phase L. donovani promastigotes were used to infect 6 to 8-weeks old BALB/c mice (2 ×107/animal) through tail vein. Ten weeks post infection, parasite burden was confirmed in three arbitrarily selected animals; after which, mice were randomly assigned into seven groups of 10 mice each (A–F). Group A – Control infected mice without any treatment (INF); Group B - Vehicle control mice that received normal saline orally (VC). Test fractions and compounds were administered to three groups orally: Group C (AAL); D (AAS); and E artemisinin (ART). These groups received three doses (50/100/200 mg/kg body weight {b.w.}) daily for ten consecutive days. Group F - received Amphotericin B (AMB, 5 mg/kg b.w. on alternate days over a 10 day period, intravenously) and served as the positive control. Ten days post treatment, 5 mice per group were euthanized by carbondioxide asphyxiation, liver and spleen parasite burden determined from giemsa-stained multiple impression smears, and expressed as Leishman-Donovan Units (LDU) that was calculated as the number of parasites per 500 nucleated cells x organ weight in mg [28]. Percent reduction of parasite burden was calculated as: (LDU of infected control - LDU of treated mice)/LDU of infected control mice × 100. Cure or protection correlated with a reduction in hepato-splenomegaly and elimination of parasites to negligible levels [28]. Fourteen days-post treatment; the remaining 5 mice per group were sacrificed for evaluating the immunological response.
Freeze-thawed leishmanial antigen (FT) was prepared as reported previously [29]. Briefly, stationary-phase promastigotes, harvested after the third or fourth passage in liquid culture, were washed four times in cold 1× PBS and resuspended at a cell density of 2×108 cells ml−1. The preparation was frozen and thawed at 80°C (30 min) and 37°C water bath (15 min), alternately for 6 cycles, and stored at −70°C until use. Soluble leishmanial antigen (SLA) was prepared as reported previously [30]. In brief, the freezing-thawing cycles were repeated ten times, and the suspension finally centrifuged (5250 × g, 4 °C, 10 min). The supernatant containing soluble leishmanial antigen (SLA) was harvested and stored at −70°C until use. The protein content in FT and SLA was measured by the method of Lowry et. al. [31].
The delayed-type hypersensitivity (DTH) response in control infected and treated mice was determined as an index of cell-mediated immunity. The response was evaluated by measuring the difference in footpad swelling at 24 h, 48 h and 72 h following intradermal inoculation of the test footpad with 50 µl (800 µg ml−1) of FT compared to the PBS-injected contra-lateral footpad [14].
Proliferation of splenic and lymphatic lymphocytes as an index of cell mediated immune (CMI) response, was evaluated in spleen and lymph nodes (axilliary, inguinal and popliteal) by trypan blue dye exclusion as well as by carboxyfluorescein succinimidyl ester (CFSE) staining. Spleens from different groups of mice were homogenized and the erythrocytes lysed with lysis buffer (20 mM Tris, pH 7.4 containing 0.14 M NH4Cl) at room temperature, 10 min. After centrifugation (1400 × g, 4°C, 10 min), the cells were washed with PBS and resuspended in complete RPMI-1640 medium (supplemented with 25 mM HEPES (pH 7.4), 50 mM 2-mercaptoethanol, 100 U ml−1 penicillin, 100 µg ml−1 streptomycin and 10% FBS). Alternately, the homogenous suspension of lymph nodes was washed and resuspended in complete RPMI 1640 medium. The viability of both splenic and lymphatic lymphocytes as determined by trypan blue dye exclusion [24] exceeded 95%. For assessment of proliferation, the spleen (5 × 106 cells ml−1) or lymph node (2 × 106 cells ml−1) cells were cultured at 37°C for 48 h in a humidified atmosphere containing 5% CO2 in the presence of 10 µg ml−1 SLA or Con A (5µg ml−1). Proliferation was ascertained by direct counting of viable cells after trypan blue dye exclusion [32].
Alternatively, for assessment of proliferation by CFSE dilution, the lymphocytes isolated from treated, infected and naïve mice were stimulated with SLA (10 µg ml−1) as described above. The lymphocytes (5 ×106 cells ml−1) were incubated with 1 µM CFSE. After 48 h, the cells were washed twice with PBS and finally resuspended in PBS. The cells were acquired in a BD LSR II flow cytometer following which the cell population was assessed and contour plots generated after appropriate gating [33].
Nitric oxide (NO), a major microbicidal molecule killing intracellular Leishmania, is released during conversion of L-arginine into citrulline by nitric oxide synthase that is activated by the Th1 subset of CD4+ T cells. The nitrite, the primary, stable and non-volatile product of NO was quantified as an indirect correlate of NO production. The 48 h culture supernatants of peritoneal macrophages of differently treated, infected and naïve mice was analyzed for nitrite contents in the presence or absence of SLA (10 µg ml−1) and, in parallel re-stimulated with AAL and AAS (50 µg ml−1) as described previously [34]. Briefly, the Griess reagent (1% sulfanilamide and 0.3% N-(1-naphthyl) ethylenediamine dihydrochloride in 5% H3PO4) was added to the culture supernatant at a 1:1 ratio and incubated for 15 min at room temperature. The optical density (OD) was determined at 550 nm using an ELISA reader. Sodium nitrite (NaNO2) diluted in culture medium was used to generate a standard curve.
Mice were bled at the time of treatment and at 10 days post treatment, and sera stored at −70°C until use. The specific serum IgG isotype antibody (Ab) response was measured by conventional enzyme-linked immunosorbent assay (ELISA) [14]. Briefly, wells of ELISA plates (Nunc, Roskilde, Denmark) were coated with FT (25 µg ml−1) and incubated overnight at 4°C. After washing three times with buffer (20 mM PBS, pH 7.2 containing 0.05% Tween 20), the wells were blocked with 1% BSA for 2–3 h at room temperature. The plate was washed and mice sera at 1,000-fold dilution was added, followed by washing and incubation with isotype-specific goat anti-mouse IgG1 and IgG2a antibody (Sigma Aldrich) at 4°C overnight. The wells were then washed and incubated at 4°C overnight with peroxidase-conjugated rabbit anti-goat IgG (Sigma Aldrich). The wells were washed and incubated with substrate solution (o-phenylenediamine dihydrochloride, 0.8 mg ml−1 in 0.02 M phosphate-citrate buffer, pH 5.0, containing 0.04% H2O2) for 30 min, and the absorbance read on an ELISA plate reader at 490 nm.
The Th1 (IFN-γ) and Th2 (IL-4, IL-10) cytokine concentrations in the sera and culture supernatant of lymphocytes from different groups of mice were measured by a bead-based multiplex assay [35]-[36]. This assay used microspheres as the solid support and allowed simultaneous quantification of cytokines in a flow cytometer according to the manufacturer's instructions. Briefly, serum, culture supernatants from SLA-stimulated (10 µg ml−1) lymphocytes or the cytokine standards were mixed with equal volume of antibody-coated capture beads and subsequently incubated with biotin-conjugated secondary antibody mixture (anti-mouse) for 2 h at room temperature in the dark. Beads were then washed (400 × g, 4°C, 5 min) and the supernatant was discarded carefully, leaving approximately 100 µl sample in each tube. This was repeated once, and the samples were further incubated with streptavidin–PE for 1 h at room temperature in the dark. After two further centrifugation steps as mentioned above, the beads were resuspended in assay buffer and read on a BD FACS Calibur (BD Biosciences) and analyzed with Cell Quest software. The data were processed using BD CBA software, with results based on a standard concentration curve.
Lymphocyte phenotyping was performed as described previously [37]. The spleens (1/3 of the organ) from differently treated and untreated BALB/c mice were placed in PBS and stored on ice prior to preparation of single cell suspension. The splenic erythrocytes were lysed as described above. After centrifugation (1400 × g, 4°C, 10 min), the cells were washed with FACS buffer (PBS containing 1%FBS). The cell suspensions were refrigerated (4°C) pending staining with antibodies. For each sample, 2 × 106 cells were stained with anti-CD4-FITC and anti-CD8-PE antibodies for 15 min on ice. The cells were then washed and resuspended in PBS for flow cytometric analysis which was performed on a LSR II flow cytometer equipped with DIVA software (Becton Dickinson).
Flow cytometry was performed for intracellular analysis of IFN-γ- producing CD4+ and CD8+ T lymphocytes at the single-cell level. Splenocytes from treated and untreated infected mice were stimulated with 10 µg ml-1 SLA for 24 h. Brefeldin A (10µg ml-1) was added to the culture and incubated for 1 hr. The cells were washed with FACS buffer and stained with APC and PE conjugated anti-CD4 and anti-CD8 antibody, respectively, washed and fixed with 100 µl of intracellular fixation buffer and permeabilized with permeable solution (BD Pharmingen). The cells were subsequently stained with FITC-conjugated anti-IFN-γ or isotype-matched control monoclonal antibodies, and analyzed on a flow cytometer following acquisition. The CD4+ and CD8+ T cells were gated individually for determining the population of FITC positive IFN-γ- producing cells [14].
Splenic cells from differently treated and untreated BALB/c mice were suspended in RPMI-1640 medium after removing the red blood cells with lysis buffer as described above. Cells (1 × 107 cells ml−1) were washed thrice and incubated for 1 h at 37°C on petri plates. After removing the non-adherent T and B cells, the adherent macrophages were collected and washed with FACS buffer. To quantify the expression of co-stimulatory molecules (CD80 and CD86) on CD11b+ and F4/80+ cells, 2 × 106 macrophages from each sample were stained with PE-labeled anti-CD80, FITC-conjugated anti-CD86 and APC- labeled anti-CD11b or PE-Cy5- labeled anti-F4/80 monoclonal antibodies on ice for 15 min and washed with PBS. Cells were acquired on a BD LSR II flow cytometer equipped with DIVA software (Becton Dickinson) [38].
Spleen cells were isolated from differently treated and infected BALB/c mice, washed with FACS buffer and incubated for 30 min at 4°C with the following fluorochrome-conjugated anti-mouse antibodies: CD8-APC, CD62L-PE and CD44-FITC (BD Pharmingen), and then fixed with 2% paraformaldehyde. Cell acquisition was performed with a BD LSR II flow cytometer [16].
Hepatic and renal functions of BALB/c mice were evaluated in treated and untreated mice as described previously [36]. Fourteen days post-treatment, mice were bled and sera were separated by centrifugation (5000 × g, 4°C, 2-3 min) and stored at -70°C until use. The hepatic and renal functions was assessed by measuring the levels of serum glutamic oxaloacetic transaminase (SGOT), serum glutamic pyruvic transaminase (SGPT), alkaline phosphatase (ALP), urea and creatinine using commercially available kits (Span Diagnostics Ltd.).
All the in vitro experiments were performed at least in triplicate. A minimum of five mice per group were used for in vivo experiments. The statistical significance of differences between groups was determined as described in the figure legends using ANOVA followed by Tukey's test by graph pad prism 5 software. P value of <0.05 was considered statistically significant. Error bars represent the standard error of the mean (SEM). Results are from one of three representative experiments.
AAL and AAS administered to 10 weeks infected BALB/c mice at 200 mg/kg b.w. for 10 consecutive days caused 95.45±2.05% and 95.84±1.95% (P<0.001) reduction of parasite burden in spleen and 86.67±2.53% and 89.12±1.92% (P<0.001) in liver, respectively (Fig. 1A & B) at 10 days post treatment. At 100 mg/kg b.w., AAL and AAS induced 88.58±1.23% and 85.08±6.92% protection in spleen and 72.99±7.2% and 80.27±1.25% in liver, respectively. The lowest dose of AAL and AAS (50 mg/kg b.w.) used in this study, caused more than 70% decrease in parasite load in spleen and approximately 50% in liver. ART was comparatively less effective since even the higher dose (200 mg/kg b.w.), could lower the parasite burden in liver as well as spleen by only 50%. With AMB (5 mg/kg b.w.), parasite elimination in liver and spleen was 94.02± 1.81% and 98.09±2.44%, respectively. AAL and AAS treatment (200 mg/kg b.w.) also resulted in significant reduction (48.84% and 45.35% respectively) in spleen weight compared to infected controls (Fig. 1 C & D) that was comparable with AMB (52.33%).
Since cure of leishmaniasis is associated with an effective immune response, we investigated the possible immunological alterations induced by the treatment of AAL, AAS and ART in L. donovani-infected BALB/c mice at cure. Leishmanial antigen (FT)-specific IgG1 and IgG2a isotype levels were assessed in the sera of mice at 10 days post-treatment. Control-infected animals exhibited significantly higher IgG1 than IgG2a levels (P ≤ 0.001) compared to treated groups (Fig. 2). The highest IgG2a/IgG1 ratio was found in AAL (2.09) and AAS (1.92) -treated mice at 200 mg/kg b.w., followed by AMB treatment group (1.56). In case of mice treated with ART, the IgG1 levels were significantly higher than IgG2a, resulting in decreased IgG2a/IgG1 ratio (0.75) even at the higher dose (200 mg/kg b.w.).
Chemotherapeutic intervention and cure is generally associated with the acquisition of a DTH response and consequently “classical” cell-mediated immunity [39]. Hence, we investigated FT-induced DTH responses in infected BALB/c mice at 10 days post-treatment with AAL, AAS, ART and AMB. AAL and AAS treated mice showed the strongest DTH response at 200 mg/kg. b.w.; a significant increase in footpad thickness was observed at 24 h (0.45±0.07 and 0.48±0.14, respectively) as compared with the INF (013±0.004) control group, followed by 100 and 50 mg/kg b.w. (Fig. 3). Whereas AMB (0.24±0.06) and ART (0.20±0.03) treated mice showed a marginal levels of DTH response. There was almost no change in DTH response at 48 versus 24 h; however, the DTH reactivity waned at 72 h in all the groups (Fig. 3).
Active VL is characterized by marked T-cell anergy toward leishmanial antigens [40]-[42]. By direct enumeration under microscope, we observed a significant proliferative response of SLA-stimulated splenocytes and lymphocytes from mice at 10 days post-treatment with AAL and AAS. Maximum effect was found at 200 mg/kg b.w. followed by 100 and 50 mg/kg b.w., whereas AMB and ART treatment showed marginal levels of SLA-specific lymphoproliferation in splenic and lymph node cells (Fig. 4A & 4B).
Alternately, lymphoproliferative capacity of lymphocytes after treatment with different groups was assessed by CFSE labeling. The percentage of normal cells that underwent division in spleen and lymph nodes was 17.8% and 16.6%, respectively. AAL and AAS treated (200 mg/kg bw) groups exhibited the highest lymphoproliferative response in spleen (32.1% and 34.2%) and lymph nodes (31.4% and 36.8%). AMB and ART treated groups induced low levels of lymphoproliferation (22.2% and 20.1%) in spleen as well as lymph nodes (21.0% and 19.8%) that was slightly higher than that observed in spleen (18.6%) and lymph nodes (17.9%) of infected control group (Fig. 4C & 4D).
The effect of AAL and AAS on macrophage function was assessed by measuring the amount of Nitric oxide (NO) produced by peritoneal macrophages of treated mice. Griess reagent was used to measure the nitrite levels, the stable end product of NO metabolism. The nitrite concentration (µM) was determined by extrapolation from a standard curve generated with sodium nitrite. In macrophages of AAL and AAS treated mice, a dose dependent NO production was observed upon in vitro re-stimulation with SLA followed by AAL and AAS stimulation and un-stimulation. Higher levels of nitrite were produced in AAL (11.18± 0.81µM) and AAS (11.33±0.63µM) treated mice (200 mg/kg/b.w.) after re-stimulation with SLA as compared to INF (3.11±0.09µM) control (Fig. 5). In contrast, AMB and ART treatment induced low nitrite levels (6.05±0.20 and 5.18±0.09µM, respectively) in peritoneal macrophages.
To evaluate the immune alterations, Th1 (IFN-γ) and Th2 (IL-4 and IL-10) signature cytokines in serum and culture supernatants were estimated by bead-based multiplex assay. Mice treated with AAL and AAS (200 mg/kg b.w.) induced significantly elevated levels of serum IFN-γ (2771±50.91 and 3033±396.69 pg ml−1) and reduced levels of IL-4 (4020±91.92 and 3961.5±208.24 pg ml−1) and IL-10 (4231.5±459.27 and 4077.5±35.0 pg ml−1) compared to untreated infected controls INF (low IFN-γ; 1236±12.37, high IL-4; 5696±79.9 and high IL-10; 5049±101.47) (Fig. 6A). AMB and ART induced low levels of these cytokines compared to INF. Similar pattern of Th1 and Th2 cytokines was observed in the culture supernatant of lymphocytes from mice treated with AAL and AAS and that was significantly high compared with infected (INF) control (Fig. 6B).
It is well established that MHC class II-restricted CD4+ T cells are dominant during the development of immunity against Leishmania [43]. However, a few studies point to an essential role for CD8+ cells in immunity to primary infection with L. major [44] and also in the induction of long-term, vaccine-induced resistance against many intracellular pathogens [13]. A low population of CD4+ (8.2%) and CD8+ (5.1%) T cells were detected in the spleens of mice with established L. donovani infection (Fig. 7). The population of CD4+ and CD8+ T cells increased 10 days after treatment with 50 mg/kg b.w. of AAL (14.4% and 10.9%) and AAS (16.8% and 12.7%), respectively. The increase was slightly more at a treatment dose of 100 mg/kg b.w. of AAL (18.4% and 13%) and AAS (18.9% and 15.4%). The CD4+ and CD8+ T cell population was, however, highest at 200 mg/kg b.w. treatment with AAL (20.7% and 15.2%) and AAS (21.8% and 16.12%) (Fig. 7). These findings demonstrate a prominent inclination toward Th1 effector function and the involvement of both CD4+ and CD8+ T cells at cure with AAL and AAS treatment. These responses in case of ART and AMB treated groups were negligible.
Both CD4 and CD8 T cells are source of IFN-γ and are essential for resolution of leishmaniasis [13], [43]. In infected mice, a low frequency of CD4 (14.43±0.40%) and CD8 (12.59±0.58%) T cells secreting IFN-γ was detected which was elevated by AMB treatment (CD4 20.13± 0.71%, CD8 18.02± 0.45%). However, the maximum induction of IFN-γ-producing CD4 and CD8 T cells was observed after AAL (32.05±0.55%, 27.16±0.42%) and AAS (33.37±0.74%, 28.09±0.41%) treatment at 200 mg/kg. b.w. In case of ART (200 mg/kg. b.w) treatment no significant increase in the frequencies of IFN-γ producing CD4 (16.41±0.46%) and CD8 (14.92±0.37%) T cells were observed (Fig. 8A & 8B).
Ligands on antigen presenting cells (APCs) called co-stimulatory molecules interact with specific receptors on T cells providing the second signal which then leads to activation of the antigen stimulated T cells. CD80 and CD86 present on APCs are essential for the activation of lymphocytes and the secretion of cytokines. The expression of CD80 and CD86 co-stimulatory molecules was analyzed individually on CD11b (monocytes) and F4/80 (macrophages). CD80 and CD86 expression and co-expression was particularly enhanced on F4/80 gated cells in comparison to CD11b positive cells. AAL and AAS treatment (200 mg/kg.b.w) significantly up-regulated the expression of CD80 (20.3% and 21.13%) and CD86 (14.6% and 16.0%) as well as co-expression of CD80/86 (21.9% and 26.0%) on F4/80 cells. AMB treatment was less effective in up-regulating CD80 (7.9%) and CD86 (7.1%) expression and CD80/86 (13.9%) co-expression, which was followed by ART. ART induced the lowest level of CD80 (5.6%) and CD86 (6.6%) expression and CD80/86 (10.3%) co-expression (Fig. 9A) Similar pattern of CD80 and CD86 expression and CD80/86 co-expression was modulated on CD11b positive cells. AAL and AAS treatment induced maximum up-regulation of CD80 (11.7% and 15.7%) and CD86 (12.6% and 15.4%) expression along with CD80/86 (13.1% and 17.0%) co-expression in comparison with infection control where as in AMB and ART treated groups, the effect was less pronounced (Fig. 9B).
Resistant to Leishmania re-infection is attributed to generation of memory T cells in the host [16]. CD44 CD62L expression was low in infected control group (10.1%), which was up-regulated after treatment with AAL (13.8%) and AAS (15.1%) at 200 mg/kg b.w. demonstrating resolution of infection and generation of memory. AMB (9.7%) and ART (9.4%) treatment exhibited negligible effect on generation of memory T cells. Further the percentage of effector memory cytotoxic T cells (CD44high CD62Llow) was also increased following treatment with AAL (37.1%), AAS (38.8%) and AMB (37.3%) and no such up-regulation was evident in case of ART (27.9%) treatment (Fig. 10).
Estimation of ALP, SGOT and SGPT for liver dysfunction and urea and creatinine for renal dysfunction was done ten days post-administration of AAL, AAS and ART in normal BALB/c mice (Table 1) as well as infected and treated mice (Table 2). AAL, AAS and ART (up to 200 mg/kg b.w.) treated group demonstrated normal levels of serum enzymes, indicating no in vivo toxicity. The fractions thus proved to be non-toxic in BALB/c mice used in antileishmanial screening.
In the absence of effective vaccines, emerging resistance against current chemotherapeutic drugs or their combinations, and a stigma of being an AIDS-defining illness, improved therapy for leishmaniasis remains desirable. Plant extracts represent a natural library of potentially bioactive molecules that can activate intrinsic leishmanicidal mechanisms. In our earlier studies, we reported potent anti-leishmanial activity of AAL and AAS with selective elimination of the parasites without affecting host macrophages. The leishmanicidal effect was mediated by programmed cell death. α-amyrinyl acetate, β-amyrine and precursors of artemisinin were the major constituents in AAL and cetin, EINECS 211-126-2 and artemisinin precursors in AAS [26]. In an effort to realize the full therapeutic potential of AAL and AAS, in the present study, we have explored the efficacy of AAL and AAS against VL using L. donovani infected BALB/c mice. The major findings emerging from this study are that AAL and AAS (200 mg/kg b.w) result in maximum clearance of parasites (85 to 90%) from the liver and spleen of infected BALB/c mice as compared to untreated infected controls. Significant reduction (88%–96%) in spleen weight was also observed with AAL and AAS. While, only marginal numbers of parasites were cleared from the liver and spleen upon treatment with artemisinin (ART) even at higher dose (200 mg/kg b.w.). Similar therapeutic effect has also been reported with the extracts of Tinospora sinensis [45], Aloe vera [27], Actinopyga lecanora [46] and with the essential oil of Chenopodium ambrosioides [47] and Bixa orellana [48].
VL is characterized by a variety of immunopathological consequences in man. The most remarkable of these are depression of CMI response and B cell activation [49]. As an index for CMI, DTH, a type IV hypersensitivity reaction was measured in treated mice. DTH develops when antigen activates sensitized TDTH cells resulting in secretion of IFN-γ and IL-2 [50] that promotes enhanced phagocytic activity of the recruited macrophages for effective killing of the parasites. DTH reaction is thus important in host defense system against Leishmania parasites. The importance of a positive DTH response in human leishmaniasis is illustrated by the fact that apparent clinical cure in the absence of a positive DTH response is often predictive of a relapsing infection [51]. Our results demonstrated that the DTH response was depressed in L. donovani infected BALB/c mice. However, treatment with AAL and AAS stimulated maximum DTH response at 24 h while negligible levels of DTH were induced with ART. Elicitation of DTH response has also been observed with Asparagus racemosus extracts [52] and Prunus cerasus treatment in infected BALB/c mice.
Resistance against Leishmania infection remains largely associated with a polarized Th1 and an insufficient Th2 response. Cytokines such as IFN-γ and IL-4 direct immunoglobulin class switching in B cells to IgG2a and IgG1, respectively [53], as an indirect correlate of T helper subsets potentiated. Thus, IgG2a and IgG1 levels indirectly reflect the Th1/Th2 responses and hence their relative production is used as a surrogate marker for the induction of protective (Th1) or deleterious (Th2) type of immune responses. To assess the immunological status of the mice upon treatment, we evaluated serum levels of parasite-specific IgG1 and IgG2a. L. donovani infection in BALB/c mice resulted in increased IgG1 and decreased IgG2a levels. However, treatment with AAL and AAS showed three-fold decrease in IgG1 with approximately two-fold increase in IgG2a levels as compared to infected controls. ART treatment did not reveal any significant difference in isotype levels. Our data reflecting higher levels of IgG2a over IgG1 thus indicate that Th1-mediated protective immunity is generated by AAL and AAS treatment. Our results comply with the reports of Sachdeva et al., [52] who showed decrease in IgG1 coupled with increase in IgG2a levels upon treatment of L. donovani infected BALB/c mice with A. racemosus in combination with cisplatin. Aqueous extract of A. racemosus has also been reported to result in significant increase in antibody titers [15] and upregulation of Th1 and Th2 cytokines [54], suggesting Th1/Th2 adjuvant activity.
Bhattacharjee et al., [55] reported that the expression of Th1 signature cytokines (IFN-γ and IL–2) is protective for VL whereas expression of Th2 cytokines viz. IL-4 and IL-10 increases during infection. Gomes et al. [56] showed that orally administered Kalanchoe pinnata selectively suppress IgG and IL-4 and up-regulates IFN-γ production in murine VL. Our studies are in agreement with these observations as AAL and AAS treatment generated a protective immunity through induction of IFN-γ and decline in IL-4 and IL-10 in serum as well as culture supernatants of spleen cells. The percentage of CD4 and CD8 T cells producing IFN-γ also increased after AAL and AAS treatment as depicted by intracellular staining. Further, AAL and AAS (200 mg/kg b.w.) stimulated strong lymphoproliferative responses in lymph nodes as well as spleens, which was observed by CFSE dilution and trypan blue dye exclusion. The increased levels of IFN-γ correlated with the strong proliferative response and activation of Th1 subset of CD4+ T cells.
Efficiency of chemotherapy in leishmaniasis is also impaired due to suppression of immune functions during the course of infection [57]. Disease outcome of VL is associated with various immunological dysfunctions. Successful chemotherapy requires strong cellular responses based on CD4+ and CD8+ T cells. Experimental mouse models of VL show that CD8+ T cells are important in control of L. donovani/L. infantum infection in the liver, through their ability to produce IFN-γ and/or their cytolytic activity [58]. Moreover, CD8+ T cells, together with CD4+ cells, are required to control and prevent reactivation of VL in mice [59]. Asparagus racemosus and Prunus cerasus have also been reported to enhance the percentage of CD4+ and CD8+ T cells in spleen of naive [60] and L. donovani infected BALB/c mice upon subsequent treatment [50]. The results of the present investigation revealed that the percentage of CD4+ and CD8+ T lymphocytes in spleens of L. donovani infected BALB/c mice were greatly augmented by AAL and AAS at 200mg/kg b.w. as compared to untreated infected controls as well as ART treated mice. The therapeutic implication of AAL and AAS in VL was further exploited by scoring the memory differentiation markers CD44 and CD62L. AAL and AAS induced generation of immunological memory as characterized by expression of CD62Llow and CD44high on CD8+ T lymphocytes.
The stimulatory Th1/Th2 balance is dictated by the presence of other costimulatory stimuli simultaneously acting on T cells and antigen-presenting cells (APCs) that play crucial roles in eliminating intracellular pathogens. The optimal activation of naive T cells is achieved by occupancy of T-cell receptor (TCR) by the peptide-MHC complex displayed on the surface of APCs, delivery of co-stimulatory signals, and the presence of pro-inflammatory cytokines [61]. Ligation of CD28 with CD80 and CD86 is known to induce the secretion of IL-6 and IFN-γ by DCs for T and B cell activation, proliferation, and differentiation [62]. CD80 and CD86 expression has been reported to be down modulated in certain diseases [63]-[64]. The expression and co-expression of CD80 and CD86 was analyzed on CD11b+ and F4/80+ cells. Treatment with AAL and AAS significantly enhanced CD80 and CD86 expression and co-expression on both CD11b+ and F4/80+ cells however maximum expression was observed in case of F4/80+ population. Thus, our results suggest the potential of AAL and AAS in activating the APCs through co-stimulatory signals that eventually help in the generation of effective immune response by secreting various signaling molecules like IFN-γ for subsequent activation, proliferation, and differentiation of lymphocytes. ART treated mice did not show significant expression of CD80 and CD86 co-stimulatory molecules.
Macrophages can be activated by different signals leading to their development into functionally distinct subsets with different disease outcomes. Thus, appropriate activation of macrophages is crucial for eliminating this intracellular pathogen. Macrophage stimulation is mediated by the products of Th1 and NK cells in particular, IFN-γ, which stimulates macrophages to produce inducible nitric oxide synthase (iNOS, also known as NOS2), an enzyme which catalyzes L-arginine to generate NO and citrulline [65]. NO is a toxic molecule that plays a major role in killing intracellular Leishmania parasites by the production of reactive oxygen species and generation of peroxinitrite. Such metabolites can cause protein, lipid and nucleic acid oxidation [66]. The function of NO in the leishmanicidal activity of activated macrophages has been demonstrated both in vitro and in vivo [67]. Our data demonstrate that AAL and AAS treatment in L. donovani infected BALB/c mice induces high levels of nitrite in SLA-stimulated macrophages (Fig. 5A) as compared to infection control as well as ART treated mice, suggesting that the inhibitory effect of the AAL and AAS on infection index is mediated by NO. In the absence of SLA, the NO production was muted but after re-stimulation with AAL and AAS, the NO production was upregulated.
The impairment of kidney function and deterioration of liver function by chemotherapeutic agents is recognized as the main side effect and the most important dose limiting factor associated with their clinical use. There is a continuous search for agents, which provide nephro- and hepatic- protection against the renal and liver impairment induced by chemotherapeutic drugs for which allopathy offers no remedial measures. In the current study, AAL, AAS and ART were administered (50, 100 and 200mg/kg b.w.) in normal and L. donovani infected BALB/c mice. It was found that serum levels of SGOT, SGPT, ALP, urea and creatinine in treated mice were comparable to those in naive mice, indicating absence of nephro- and hepato-toxicity.
Taken together, our findings indicate that treatment of infected mice with AAL and AAS significantly decreased the hepatic and splenic parasite load with reduction in spleen weight. AAL and AAS caused increased production of Th1 cytokines (IFN-γ) and concomitant decrease in Th2 signature cytokines (IL-4 and IL-10). The Th1 subset potentiation was also evident from class switching in B cells to produce higher levels of IgG2a over IgG1 and significant elicitation of DTH. AAL and AAS also resulted in higher CD4+ and CD8+ T cell numbers, lymphoproliferation, up-regulation of co-stimulatory molecules (CD80 and CD86) on APCs and generation of NO. Cure as well as resistance against L. donovani infection was due to the parasite killing by AAL and AAS that was mediated by immunopotentiating effects shifting Th1/Th2 balance in favour of the host with induction of cell-mediated immunity as postulated in Fig. 11. AAL and AAS may emerge as prospective antileishmanial therapy that may be administered alone or synergistically with current chemotherapeutic drugs, owing to their safety and ability to enhance disease healing Th1 immune responses.
|
10.1371/journal.pcbi.1005471 | The preferred nucleotide contexts of the AID/APOBEC cytidine deaminases have differential effects when mutating retrotransposon and virus sequences compared to host genes | The AID / APOBEC genes are a family of cytidine deaminases that have evolved in vertebrates, and particularly mammals, to mutate RNA and DNA at distinct preferred nucleotide contexts (or “hotspots”) on foreign genomes such as viruses and retrotransposons. These enzymes play a pivotal role in intrinsic immunity defense mechanisms, often deleteriously mutating invading retroviruses or retrotransposons and, in the case of AID, changing antibody sequences to drive affinity maturation. We investigate the strength of various hotspots on their known biological targets by evaluating the potential impact of mutations on the DNA coding sequences of these targets, and compare these results to hypothetical hotspots that did not evolve. We find that the existing AID / APOBEC hotspots have a large impact on retrotransposons and non-mammalian viruses while having a much smaller effect on vital mammalian genes, suggesting co-evolution with AID / APOBECs may have had an impact on the genomes of the viruses we analyzed. We determine that GC content appears to be a significant, but not sole, factor in resistance to deaminase activity. We discuss possible mechanisms AID and APOBEC viral targets have adopted to escape the impacts of deamination activity, including changing the GC content of the genome.
| The APOBEC family of cytidine deaminases are important enzymes in most vertebrates. The ancestral member of this gene family is activation induced deaminase (AID), which mutates the Immunoglobulin loci in B Cells during antibody affinity maturation in jawed vertebrates. The APOBEC family has expanded particularly in the mammals and in primates, where they have evolved roles in restriction of viruses and retrotransposons. Biochemical studies have established that AID preferentially targets “hotspots” such as AGC and avoids “coldspots” such as CCC. Other APOBECs have evolved distinct hotspots. For example, APOBEC3G, which targets retroviruses including HIV, has evolved to target the motif CCC as a hotspot, but it is unclear why. Here we ask why the AID/APOBEC cytidine deaminases evolved their particular mutational hotspots. Our results show that a wide range of unrelated genes including mammalian LINE1 ORFs and non- mammalian (ancestral-like) viruses are highly susceptible to mutations in APOBEC hotspots and less susceptible to the hypothetical non-APOBEC hotspots. On the other hand, mammalian viruses tend to exhibit low susceptibility to the same APOBEC hotspots, suggesting these viruses have co-evolved to minimize potential damaging mutations, and that the native GC content plays a large role in this behavior.
| The AID/APOBEC family of cytidine deaminases have important functions in both intrinsic and adaptive immunity. AID is expressed primarily in germinal center B cells as part of the adaptive immune response [1], whereas the APOBECs act primarily in the intrinsic immune response in various cell types (reviewed in [2]). These mutagenic enzymes act mostly upon single-stranded DNA [3] converting Cytosine in DNA or RNA to Uracil, or for AID, a methylated Cytosine to Thymine in vitro [4]. In the absence of further editing, the resulting U-G mismatch in DNA will often be replicated over to create a C to T transition mutation [5]. AID in particular relies on downstream non-canonical DNA repair pathways to introduce further mutations [6]. Although AID-mediated mutations occur at rates approximately 106 higher than background [7, 8], they occur almost entirely in the Immunoglobulin genes that code for the B-cell receptor [9, 10]. At the same time, DNA editing mechanisms can potentially target the host genome (reviewed in [11]) and evidence of AID and APOBEC-mediated mutagenesis has been identified in many human cancers [12]. Thus, there appears to be a tradeoff between the benefits and the potential for self-damage of APOBEC-mediated mutations, yet the details on how this tradeoff is achieved both on a biochemical and evolutionary level are still not well understood.
AID has been proposed to be the most ancestral member of the APOBEC family [13] and is found in all jawed vertebrates. Duplication and diversification of AID have created the APOBEC family, which edit both RNA and DNA. All mammals additionally have at least 4 APOBEC genes, APOBECs1-4 [14, 15]. APOBEC1 deaminates the mRNA of the Apolipoprotein B gene (ApoB) at a specific site to introduce a stop codon in humans and mice [16, 17]. Zebrafish APOBEC2 appears to maintain embryonic development and other functions in development including retina regeneration [18–20], while the role of human APOBEC2 is currently unknown. The function of the APOBEC4 gene also still not known [2].
APOBEC3 has diversified considerably during mammalian evolution. Whereas mice have a single APOBEC3, humans have seven APOBEC3 variants (A-H) [21]. The APOBEC3 genes participate widely in innate immunity by mutating retrotransposons [22], exogenous viruses [23], and endogenous viruses [24].
Human APOBEC3s differ in their restriction capabilities. Perhaps the best understood example of APOBEC restriction is against HIV, which is targeted by human APOBEC3G. The HIV genome in turn encodes a counter-defense in the form of the vif gene, which encodes a protein that targets APOBEC3G for degradation. Experiments in vitro using vif-deficient viruses showed that HIV was highly restricted and mutated by APOBEC3G [25], and later APOBEC3F [26]. Although APOBEC deamination appears effective in inhibiting HIV in the absence of vif, other studies suggest that HIV has evolved its own deamination hotspots, thus co-opting APOBEC mutagenesis to increase the likelihood of beneficial mutations [27]. Thus, there is a conflict between HIV adopting APOBEC3G mutation via hotspots compared to the ability of APOBEC to induce lethal hypermutation, although it is not currently established to what extent HIV favors either. Anti-viral activity for human APOBEC3G and other mammalian APOBEC3s has been observed across a wide range of viruses including Human Papillomaviruses such as HPV-16 [28], Adeno-associated virus (AAV) [29], Torque Teno Virus (TTV) [30], Human Herpes-Simplex Virus 1 (HSV-1)[31], Human T-lymphotropic virus 1 (HTLV-1) [32] and Simian Foamy Virus (SFV) [33]. AID has been shown to deaminate the Hepatitis B Virus (HBV) [34] and C virus (HCV) in vitro [35]. In addition, APOBECs are thought to restrict human endogenous retroviruses (HERV) [24]. Furthermore, human APOBEC3s are capable of mutating the Long Terminal Repeat (LTR) retrotransposons such as the LINE1 (L1) element [36]. In particular, human APOBEC3B, human APOBEC3C, and human APOBEC3F strongly inhibit L1 retrotransposition in vitro, while human APOBEC3G and human APOBEC3H weakly inhibit this same activity [37]. This phenomenon has also been demonstrated with APOBEC3s in reptiles [38]. Homologs of human APOBEC3B have been shown to restrict L1 retrotransposition in other primates as well, where greater L1 activity was correlated with lower human APOBEC3B expression levels and diversity [39].
AID and the human APOBEC3s differ not only by their targets which include exogenous and endogenous viruses, but also their preferred deamination context in DNA or RNA. The relation between these preferred contexts and their biological targets is still poorly understood. Although members of the APOBEC family all mutate C to U, individual preferences vary. AID, for example, preferentially deaminates the hotspot WRC (W = A/T, R = A/G, C is mutated) while avoiding the coldspot SYC (S = G/C, Y = C/T). APOBEC3G, on the other hand, preferentially mutates CCC, an AID coldspot [6, 40]. Additionally, in the context of mutations targeting HERV, it has been shown that the +1 position is also important for APOBEC3F and APOBEC3G leading to more complex motifs, respectively TTCA and CCCA [41]. Preferences for AID/APOBEC targeting are summarized in Table 1. Interestingly, individual APOBEC preferences can be changed via minimal amino acid changes in a hotspot recognition loop [42].
The AID/APOBEC enzymes must maintain a delicate balance of ensuring deamination is sufficient for key functions such as antibody diversification and restriction of viral and retrotransposon activity, while avoiding detrimental targeting of the host genome. Achieving this balance is accomplished via several mechanisms. Firstly, individual AID/APOBECs are expressed in distinct cell types. For example, AID is expressed primarily in Germinal Center B cells, whereas APOBEC3G is expressed in several cell types, particularly T cells [53]. Additionally, there may be intracellular regulation. Thus, in Germinal Center B-cells, AID is mainly found in the cytoplasm but its activity levels, are modulated in part by its Nuclear Localization Signal (NLS) [54]. Although APOBEC3B also contains an NLS domain, this type of regulation has still not been investigated for other APOBECs [55]. Although AID is actively targeted to the antibody (Immunoglobulin) loci in B-cells, it also has the potential to mutate many “off-target” locations throughout the entire genome [56]. These targets include known oncogenes such as Bcl6, Myc, and others. Mistargeted AID is also linked to poor prognosis in Chronic lymphocytic leukemia (CLL) [57] and other B Cell Lymphomas [58]. Furthermore, the mutation pattern of human APOBEC3B, NTC, has been strongly correlated to so-called “kataegis” mutations in breast cancer, characterized by short inter-mutation distances [12].
Although the mechanism of action and physical structure of several APOBECs are reasonably well characterized, it is not well understood how their specific hotspots were established during APOBEC evolution. We investigate the hypothesis that the APOBECs evolved their mutational preferences to control the impact of deaminase activity by evolving preferences that induce the most damage to their intended targets such as exogenous virus genomes, while causing the least damage to the host genes. We describe a bioinformatics-based approach examining the impacts of motifs, both known and hypothetical, on various genomes. We demonstrate that the current APOBEC preferences are shaped by their ability to restrict their targets, though these observed vulnerabilities may be mediated in large part by the GC content of the native sequence.
Our motivation is to quantify the impact of deamination at motifs, both known and hypothetical, across many different potential targets. Previous structural descriptions of APOBEC mutagenesis have indicated the role of a recognition pocket in the -2 and -1 positions on the single-stranded DNA substrate preceding the targeted Cytosine [59]. We therefore examined the effects of mutation at each of the 16 different NNC hypothetical hotspots, referred throughout as NNC motifs (N = any nucleotide) to NNT, consistent with the AID/APOBEC deamination mechanism in the absence of further repair. We measure a genome’s susceptibility to mutation by extending a method first introduced by Karlin [60] (see: “Calculating susceptibility” section in Methods for a detailed definition, including discussion of necessary controls). Briefly, the model we use assesses the impact of a particular hotspot or motif (for example, AGC) on a given gene coding sequence by first counting the number of observed motifs in both orientations, then counting the number of nonsynonymous (changing the amino acid) mutations on the sequence that would occur by mutating C to T in every observed motif (or G to A in the reverse orientation) (Fig 1A). In quantifying a coding sequence’s susceptibility to mutations in a particular motif, we assumed that nonsynonymous mutations have more potential to be deleterious. Given the complexity of nucleotide sequence evolution, there are many factors this simple model does not account for, such as the GC content of the genome, or the impact of hypothetical mutations other than from C to T. To address these considerations, we also evaluate several alternative models (see: “Alternative models and controls corroborate observed susceptibilities” section below).
To illustrate our method, we show an example in Fig 1. Here, the target sequence has 4 AGC motifs, with two on the forward strand, and two on the reverse strand (appearing as the reverse complement of AGC, GCT) of which 3 cause nonsynonymous mutations and 1 is synonymous when mutated from C/G to T/A. This sequence therefore has a nonsynonymous mutation fraction of 0.75, equivalent to the probability that a mutation at an AGC site causes an amino acid change (Fig 1B).
This fraction was then compared to a null model consisting of 1000 randomized sequences that preserve the amino acid sequence, but where synonymous codons are chosen randomly. For each randomized sequence in the null model, the nonsynonymous mutation fraction was calculated. Under the model, if a gene is susceptible to a particular NNC motif, then it would have a higher incidence of nonsynonymous mutations than the majority of randomized versions from the null model. We therefore refer to the percentile of randomized sequences having a lower nonsynonymous mutation fraction than the wild-type sequence as that gene’s susceptibility for a given motif (Fig 1C). It is in fact equivalent to a P-value for susceptibility. Thus, a particular motif is susceptible to mutation if the percentile is close to 1. Conversely, a gene with a low susceptibility, closer to 0, can be considered resistant to that same motif. As an example, if we are considering AGC to AGT mutations, if a gene has a nonsynonymous mutation fraction of 0.75 but in the null model, 1 out of 4 permutations have an even lower nonsynonymous mutation fraction, this gene’s sequence would be at the 25th percentile for the nonsynonymous mutation fraction of the AGC motif (Fig 1D).
We next examined how mutational susceptibility of retrotransposons and viral genomes was affected by different APOBEC preferences. We defined a gene set as a group of genes with a common biological function, for example, a set of housekeeping genes, or the genes in a particular virus. Starting with the individual susceptibility measures for each gene, we calculated the average susceptibility for each gene set (Fig 2A). As an example, Fig 2B shows the susceptibility values for the Epstein-Barr Virus (EBV), a herpesvirus that is tropic to B-cells, potentially exposing it to AID, which has a strong preference for mutating AGC motifs.
We reasoned that under a null model, if susceptibility is equivalent to a P-value, then random percentiles would assume a uniform distribution from 0 to 1, and would have a median and mean of 0.5. In contrast, the average susceptibility of the EBV genome is approximately 0.2 and for the set of human housekeeping genes it is 0.4 (Fig 2C), suggesting human housekeeping genes are slightly less resistant to the impact of AGC deamination. As each coding sequence susceptibility score can be treated as a P-value, to assess the susceptibility of each gene set statistically we combined the individual P-values using an unweighted Z-transform approach, which is more powerful than the standard Fisher method [61]. We found this approach was capable of detecting significance even for small gene sets, such as HIV or Adeno-Associated Virus 2 (AAV2), which have only 7 and 8 genes respectively but show significant results for several motifs (S1 Table). In the example of Fig 2B, the susceptibility of the EBV genome to the AGC motif was statistically lower than a null median of 0.5 (P < 10−49, S1 Table) and in general EBV displays statistical resistance to 13 of the 16 NNC motifs we examined (using a cutoff of P < 3.125 × 10−3 for significance, representing Bonferroni corrected P = 0.05). We applied the same z-score method to determine the statistical vulnerability of a gene set, defining it as the combination of the scores (1-susceptibility) to obtain a P-value (S2 Table).
To quantify the impact of deamination on various genomes we analyzed a diverse set of genomes, including viruses previously reported as AID or APOBEC targets, sets of housekeeping genes from mammalian (mouse and human) genomes and, as controls, genomes from viruses of non-vertebrate hosts that have evolved in the absence of AID/APOBECs (see Methods section “Data Sources”, and Table 2). For each of these gene sets we were interested in quantifying the impact of C>T mutations at the reported hotspots for AID/APOBEC (Table 1). Since we wanted to gain insight into why these particular hotspots evolved, we considered all motifs of the form NNC (N = any nucleotide), which includes 10 hotspots targeted by AID/APOBEC and 6 hypothetical hotspots such as GCC that have not evolved specificity within the AID/APOBEC family. We measured susceptibility values (as described above) for every gene within each gene set, calculating the mean susceptibility in each case. This was repeated for all 16 NNC motifs, thus obtaining 16 susceptibility values for each gene set.
We observed that one of these gene set clusters (indicated by the red box labeled “Resistant” in Fig 3) has high overall resistance (i.e., low susceptibility) to NNC motifs. This resistant cluster includes many viral genomes, such as Human Herpesvirus 1 [31], and hepatitis C Virus [35]. Curiously, this cluster includes the Epstein-Barr Virus, suggesting that this virus may have evolved resistance to mutations by both APOBECs and AID. Although to our knowledge no studies have demonstrated a direct mechanism of AID restricting EBV experimentally, the virus is B-cell tropic and would presumably be exposed to AID deamination in germinal centers and to an extent in extrafollicular compartments [62]. Gene sets of host (mouse and human) housekeeping genes also show overall resistance to mutations in NNC motifs. Housekeeping genes will, by definition, be co-expressed with AID and those APOBEC genes wherever they are expressed. Although little is known about the mechanisms by which our host genomes avoid APOBEC mutations, the evidence from AID in B-cells shows that mutations indeed occur genome-wide but are repaired, although the repair process is imperfect [56]. Our analysis here suggests that this mutational behavior (perhaps including APOBEC) may have created selective pressures on the genome to minimize this potentially hazardous activity.
The cluster of gene sets in the blue box of Fig 3 and labeled as Vulnerable had higher mean susceptibility scores across all NNC motifs. The cluster includes a human endogenous retrovirus (HERV), the human (LINE-1) and mouse retrotransposon (LINE-1 and MusD) coding regions, as well as the control genomes, which include viruses of invertebrate hosts (oyster, plants) that would not have co-evolved with AID or APOBEC. Surprisingly, there are several virus gene sets that are targeted by APOBEC and that we did not initially expect to be vulnerable, including HIV, Simian Foamy Virus (SFV), Hepatitis B Virus, Murid Herpesvirus 4 and Human Papillomavirus 16. HIV’s apparent vulnerability is surprising since HIV is a well-known target of human APOBEC3G. However, as described in the Introduction, HIV encodes vif, an APOBEC3G inhibitor, which may obviate the need for sequence-level avoidance. Other unknown defense mechanisms for these targeted viruses may also explain their apparent vulnerability. Mouse APOBEC3 is incapable of impairing MHV68 (Murid Herpesvirus 4) viral replication, for example, whereas several human APOBEC3s are capable of restriction, suggesting an alternate, unknown anti-mAPOBEC defense may be present [63]. Furthermore, recent work on HIV has shown that APOBEC3G may in fact help the virus mutate and create escape variants [64], which may in turn explain its relatively high level of susceptibility. A similar explanation may hold for human papillomavirus (we analyzed HPV16), which recent studies have shown to be extensively mutated in vivo and in vitro [28]. In light of these phenomena, our results suggest that even if a virus targeted by APOBEC is found to be susceptible to APOBEC hotspots, the virus itself may contain an alternate method of escape such as vif, subversion of APOBEC for evolutionary purposes, or other factors.
Considering now the hierarchical clustering results for the 16 NNC motifs (vertical dendrogram of Fig 3) we observed that many existing (or “canonical”) AID/APOBEC hotspots, including 7 of the 10 AID/APOBEC hotspots are contained within the same cluster (right-most 7 motifs in Fig 3, boxed in orange). This cluster contains 3 of the 4 canonical AID hotspots AGC, AAC, and TGC (only excluding TAC), as well as 3 of the canonical human APOBEC3B NTC hotspots (ATC, GTC, and TTC, only excluding CTC), and the mouse APOBEC3 hotspot TCC, but excludes the human APOBEC3G hotspot CCC. In light of the contrast observed, the existence of the two clusters, resistant and vulnerable, suggests that AID/APOBEC hotspots cause high mutational damage to exogenous viruses and native transposable elements. At the same time, we have observed that many of the genes targeted (intentionally in the cases of viruses and unintentionally in the case of host housekeeping genes) show resilience to the impact of these mutations at hotspots. This resilience may have evolved as an avoidance strategy to minimize mutations if the hotspots that APOBEC evolved do not cause high collateral damage to the host genome.
GC bias is an important evolutionary constraint in many viruses, and may play important APOBEC-independent roles such as protection from insertions [65]. Although we demonstrate later that the GC content of the gene sets considered correlates strongly with susceptibility (see next section “GC content of genomes predicts deaminase hotspot susceptibility”), we sought to quantify the extent to which native GC content contributed to the observed susceptibilities. To that end, for each of the gene sets considered, we generated new susceptibility scores that controlled for the GC content by defining a new null model that used the GC content of the gene set being analyzed to select codons (see Methods). Using the same unweighted z-score calculation as described above to determine statistical significance, we observed that for the high GC-content EBV genome, 9 out of the 16 NNC trinucleotide contexts that we examined were still statistically resistant (compared to 13 out of 16 in the uncorrected case). For the vulnerable OsHV genome, 6 out of 16 NNC trinucleotides were statistically vulnerable (compared to 10 out of 16 in the uncorrected case). We confirmed these trends again by recalculating susceptibilities for each of the gene sets, now using the GC-corrected null model (S2 and S3 Tables). We visualize these trends with another heatmap similar to Fig 3 which maintains the ordering of the columns and rows for ease of comparison (S1B Fig). Although the trends in susceptibility are similar to the non-GC-corrected case, the differences between vulnerable and resistant gene sets are reduced, suggesting that GC content explains a great deal, though not all, of the observed resistance to APOBECs.
Additionally, there is a question of whether the total number of nonsynonymous mutations should be minimized, or whether it is more appropriate to examine the nonsynonymous mutation fraction, which normalizes the number of motifs. There are two possible scenarios to consider for the frequency of mutation. In one scenario, a gene may only be exposed to the mutagen a few times during its activity. This is observed in B-cells, where AID is exposed to a gene possibly only once per cell cycle [8]. In this scenario, because the number of mutations is limiting, the probability that a mutagenic event causes an amino acid change is reduced by minimizing the fraction of nonsynonymous mutations, i.e. the number of nonsynonymous mutations divided by the total number of motifs. If a mutagen is exposed to a sequence only rarely, as in the case of AID, a coding sequence could minimize the probability of each mutagen changing its sequence by having many hotspots in the silent position. In the second scenario, mutations are not rate-limiting and the absolute number of hotspots causing nonsynonymous mutations would need to be minimized rather than the fraction. We found that tests conducted under the second scenario show similar results, albeit with slightly weaker trends (see section below “GC content of genomes predicts deaminase hotspot susceptibility“). Thus, our results are clearest in the first scenario (fraction of nonsynonymous mutations) assuming it is more indicative of susceptibility to mutagenesis. We conclude that examining the fraction is more instructive in determining trends (S2A Fig).
Additionally, as a control, we applied this model to different mutation profiles rather than the standard C to T. We chose every type of transition mutation (NNA to NNG, NNT to NNC, and NNG to NNA, which is complementary to NNC to NNT). In the cases of NNA to NNG and NNT to NNC (S2B and S2C Fig), we observe trends that are the opposite of our findings as described in the main text. Averaged across all 16 hypothetical NNA -> NNG hotspots, the susceptibility is lower in the oyster herpes virus by 0.18 (P = 0.011, 2 tailed t-test) relative to the housekeeping genes, though still prevalent but not as strong for the L1 elements (difference = 0.10, P = 0.24). Similarly, with all 16 NNT to NNC T to C mutations, L1 (difference = 0.12, P = 0.048) and oyster herpes viruses (difference = 0.198, P = 0.0063) show lower susceptibility to hotspots relative to human housekeeping genes. This suggests that mutagenesis at NNC to NNT hotspots is preferable for host genomes compared to these other patterns of mutation. We applied the same analysis as we did in Fig 3, to the hypothetical mutations of NNG to NNA (S2D Fig). Because we consider motifs on both strands, NNG to NNA is equivalent to CNN to TNN. For NNG to NNA we observed some similarity to the NNC to NNT case. Again there appear to be groups of resistant and vulnerable gene sets and these overlap somewhat with the NNC to NNT case, although the clusters are less clearly separated. If the average susceptibility of each gene set (rows in Figs 3 and S2D) were the key feature separating the vulnerable and resistant clusters, then we would expect similar clusters for both NNC to NNT and CNN to TNN. However, the clusters in each case are not the same suggesting that hotspot (column-) specific features that are unique to each cluster, are also important.
Although considering all 12 possible types of mutation is beyond the scope of this study, we did consider the three transition mutations other than NNC to NNT (S2B–S2D Fig) and one transversion (NNC to NNG, S2E Fig) that is directly comparable to Fig 3. Concerning the latter case, we found some similarities to the NNC to NNT transition case in that all control and retrotransposon gene sets clustered together in a predominantly vulnerable group. Clearly, some similarity is to be expected since most mutations at the third position of a codon are synonymous, and the NNC contexts are shared between the two cases.
When we analyzed susceptibilities using the total number of nonsynonymous replacements rather than the fraction of replacements, we observed greater variability across the 16 NNC motifs (S2A Fig). In spite of this, at the highest level the same two clusters of vulnerable and resistant gene sets emerge. However, under this assumption many of the gene sets in the resistant cluster, such as the mammalian housekeeping genes and the Epstein-Barr virus genome here show high vulnerability to GC rich motifs. These motifs include the APOBEC3G hotspot CCC, and the hypothetical hotspots GCC and GGC, all three of which cluster together. These gene sets simultaneously show resistance to the AT rich motifs ATC, TTC, TAC, and AAC, which are also clustered together. Varying GC content of the gene sets may explain these discrepancies. A gene with a very high GC content would contain more GC-rich motifs and fewer AT-rich motifs (and therefore a higher raw nonsynonymous mutation count) by chance. However, it is surprising that the trends are similar to our observed susceptibilities (Fig 3), since our definition of susceptibility examines only the nonsynonymous mutation fraction instead of raw count, “normalizing” for the GC content. Our analysis from here on thus examines the model of susceptibility that calculates the percentile of nonsynonymous mutation fraction compared to random sequences.
It is not immediately clear how GC content will also affect the nonsynonymous mutation fraction that is used to calculate susceptibility (Fig 3). However, a gene set with varying GC content may adopt a codon usage that inherently favors APOBEC hotspots at positions that cause synonymous mutations. We therefore compared the GC content of our vulnerable and resistant gene sets and found a strong statistical difference, with resistant gene sets having much higher GC content (Fig 4A, 2-tailed t-test, P <1.0−5). Furthermore, we observed that the correlation of susceptibility with GC content was strong for some motifs and not others. For many of the 16 motifs we examined, a higher GC content correlated strongly with low susceptibility to that motif as demonstrated by very negative correlation coefficients (Fig 4B). For example, the AT-rich hotspot TTC showed a strong correlation (Fig 4C), but for the GC-rich hotspot CCC the trend is not as apparent (Fig 4D). There is zero correlation between susceptibility of the motif GGC (which is not an observed AID/APOBEC hotspot) and GC content. Finally, to further corroborate these trends we have assessed the correlations between GC content and the alternate definition of motif susceptibility which considers the absolute number of nonsynonymous mutations rather the fraction. We observe again that a low GC content confers stronger susceptibility (negative correlation) to motifs of the WNC (W = A/T, N = any nucleotide) motif (Fig 4E), consistent with our findings to be discussed below (see section: “Retrotransposons and viruses show great discrepancy in resilience to biological deaminase hotspots.”). Our analysis suggests that GC content is strongly predictive of existing APOBEC hotspot susceptibility, with the exception of the human APOBEC3G hotspot CCC.
The results shown in Fig 3 suggested that the current APOBEC hotspot preferences may have evolved to mutate particular targets while minimizing self-damage. To explore this further, we asked which of the 16 possible NNC motifs were best in discriminating host against parasitic (viral, transposable element) genomes. We proceeded by performing all pairwise comparisons for each gene set in the resistant cluster to each gene set in the vulnerable cluster, noting the differences in the susceptibility scores for each of the 16 NNC motifs. Since we observed 10 resistant gene sets and 12 vulnerable gene sets (Fig 3), this resulted in 120 individual comparisons. For each comparison the 16 NNC motifs were sorted by the difference in susceptibility of the resistant to the vulnerable genome (resistant—vulnerable). As an example, when we compared human housekeeping genes (resistant) to L1 retrotransposon elements (vulnerable) we discovered that the housekeeping gene sets had an average susceptibility to the motif ATC of 0.39, whereas the average susceptibility for the L1 elements was 0.872 (S4 Table). Thus, with a difference of -0.482, the APOBEC3A/B hotspot ATC shows the strongest contrast out of any of the 16 motifs and is ranked the highest, whereas the motif GAC has a difference of +0.321 and is ranked the lowest. The rankings suggest that ATC, rather than GAC, might evolve to be a preferable motif due to its capability for damaging L1 ORFs more so than impacting host genes.
By analyzing all 16 motifs across 120 comparisons, we found that the rankings of human APOBEC3B (NTC, N = any nucleotide) and AID (WRC, W = A/T, R = A/G) hotspots in particular, were significantly lower than expected by chance (see Methods, bootstrapped P = 0.0071 and P = 0.0002 respectively). We further calculated, for each motif, the average difference across all pairwise comparisons of vulnerable and resistant gene sets, again ranking each motif by the overall difference in susceptibility. For 9 out of the 16 motifs we examined, the susceptibility of the vulnerable genes was strongly lower than that of the resistant gene sets to a similar degree, a difference that was highly significant (t-test, P < 2.2 × 10−16 for all 9 motifs, Fig 5A columns shown in orange box). Curiously, these 9 motifs which demonstrate the greatest difference in mean susceptibilities also include all 8 motifs defined by the motif WNC (W = A/T). Although WNC itself is not necessarily a canonical preference seen in an existing cytidine deaminase, this set of 9 significant motifs still contains considerable overlap with existing cytidine deaminases, including all AID hotspots (WRC, where R = A/G), three of the four possible APOBEC3A/B hotspots (ATC, TTC and GTC), the human APOBEC3C hotspot (TTC) and the murine APOBEC3 hotspot (TCC).
As an alternative method to confirm the significantly lower susceptibility of these 9 motifs (WNC and GTC), we compared the average rank of the 16 NNC motifs across all of our 120 pairwise comparisons, noting that the average rank of the 9 motifs were all statistically lower than expected by chance, whereas the average ranks of the other 7 motifs were higher than this (S5 Table, BH-corrected P < 10−4, one-sample MWU). Additionally, although it was not possible to include degenerate motifs such as WRC and NTC in our original analysis due to the non-independence with other observed motifs such as ATC, we did calculate susceptibility scores for these two motifs separately, as described next.
The 9 motifs we identified here as having lower susceptibility (Fig 5A) includes all 7 motifs that we previously indicated as distinguishing APOBEC-resistant and vulnerable gene sets based on the clustering of motif susceptibility (Fig 3, orange box). These 7 motifs cluster primarily as a result of their biological ability to distinguish vulnerable and resistant gene sets. Additionally, the two degenerate motifs of respectively AID (WRC) and hAPOBEC3B (NTC) showed large differences between vulnerable and resistant motifs as well (shown in white in Fig 5A). Both results suggest that the motifs that most strongly distinguish APOBEC-resistant and vulnerable genomes tend to be the existing hotspots, rather than hypothetical hotspots not associated with AID or APOBEC.
We observed that broadly, the vulnerable cluster (Fig 3, blue box) contains two qualitatively distinct families of genes, notably viruses and transposable element clusters. Similarly, the resistant cluster includes endogenous host genes such as the housekeeping gene set, and also viruses such as EBV and HSV-1, that have likely co-evolved with the APOBECs and may have acquired low APOBEC susceptibility as a consequence of this co-evolution. We assumed that the preferred APOBEC hotspots would be those that maximally damage their intended targets in intrinsic immunity, i.e. retrotransposons and vulnerable viruses, while minimizing damage to the host genome. With this assumption in mind, we looked more closely at two particular comparisons relevant to APOBEC evolution in humans. We first examined human housekeeping genes against human L1 elements, as the L1 elements are a major biologically significant target for APOBECs [39, 66]. Secondly, we compared two viruses: the Epstein-Barr virus (Human Herpesvirus 4) from the resistant cluster, against the Ostreid Herpesvirus, which we used as a control genome since its host is invertebrate and therefore not expressing any APOBEC genes, and is categorized in the vulnerable cluster.
In the first comparison (Human housekeeping vs Human L1) we found that the top 3 motifs in terms of susceptibility differences are all APOBEC3B TC hotspots: ATC, GTC, and TTC (Fig 5B, S4 Table). Interestingly, the observed difference does not arise because human housekeeping genes are particularly resistant to these motifs, since they have mean susceptibilities of 0.39, 0.51, and 0.43 respectively, which is close to that expected by chance (0.5). However, retrotransposons are particularly susceptible to these hotspots, with susceptibilities ranging from 0.87 to 0.91 for these three hotspots, where 1 represents maximum susceptibility. These high values suggest that human APOBEC3B hotspots may have evolved due to a particularly high capability of inducing nonsynonymous mutations in L1 elements, rather than low susceptibility to these hotspots in the host genome. We verified this effect for the AID hotspot WRC on some human non-housekeeping genes as well, by comparing the susceptibilities of our human housekeeping genes to a set of B-cell specific genes [67]. We used the list of B-cell specific genes listed in this study to obtain their coding sequences from the UCSC genome browser and calculated their susceptibilities, using the longest sequence if multiple isoforms were available for a single gene. For the B-cell specific genes, the susceptibilities to the 4 WRC hotspots were even lower than housekeeping genes: (AGC: difference = 0.10, P = 0.0075), (AAC: difference = 0.1, P = 0.000038), (TAC: difference = 0.12, P = 8.8 × 10−9), (TTC: difference = 0.07, P = 0.00313), suggesting that B-cell genes are under even more pressure to avoid deleterious mutations at AID hotspots than other genes.
In the second case (Fig 5C, S6 Table), we compared the APOBEC-resistant Epstein-Barr virus (EBV) to the Ostreid Herpesvirus (OsHV). The most highly ranked motifs in terms of susceptibility differences again included many A/T-rich motifs, with 8 out of the 9 top motifs again having the motif WNC. Furthermore, the two motifs showing the strongest difference are the hotspots ATC and TTC (collectively the motif WTC), which are human APOBEC3A/APOBEC3B hotspots. Many of the AID hotspots that EBV might be expected to evolve resistance to (given that EBV is a B-cell tropic virus) also show strong contrast, with the hotspots TAC and AAC ranked 4th and 5th and TGC and AGC being ranked 8th and 9th respectively. These differences, in contrast to the comparison between housekeeping genes and L1 elements, arise not only due to the high susceptibility of OSHV to motifs such as ATC (0.82) and TTC (0.83), but also due to EBV’s strong resistance, i.e. low susceptibility (0.21 for ATC, 0.277 for TTC). These low susceptibilities to APOBEC hotspots in EBV, together with our observation that AID/APOBEC hotspots again show the strongest contrast between resistant and vulnerable viruses, suggest that viral genomes such as EBV have adopted resistance to this particular form of mutagenesis.
In general, we note that motifs of the form WNC, and more specifically, WTC (W = A/T) appear to show the largest difference in vulnerabilities between housekeeping genes and retrotransposons, as well as between resistant virus and vulnerable genes. While WNC itself is not necessarily a canonical AID or APOBEC hotspot, it is a superset of the well-characterized AID hotspot WRC [1]. Furthermore, WTC overlaps with the hotspot NTC seen in human APOBEC3A/APOBEC3B. Thus, the fact that many of the motifs with the largest differences correspond to existing AID/APOBEC hotspots supports a hypothesis that the ability of AID/APOBECs to act effectively against viruses and retrotransposons while avoiding damage to the host is an evolutionary advantage for the host.
Many factors play an important role in shaping virus nucleotide evolution [68, 69]. We next describe how we investigated in particular the effects of codon bias, GC content, CpG dinucleotide frequency, selective pressure, and AID hotspot targeting of a hypothetical or actual virus sequence to validate our model.
We constructed new gene sets of 250 genes of completely random open reading frames that are 1500 nucleotides long, beginning with the methionine codon, ending with a stop codon, and with random non-stop codons in between. For the random codons, each nucleotide was chosen randomly, with either 40%, 50%, or 60% GC content. We calculated the susceptibility scores of these random sequences to validate susceptibility independent of evolutionary effects. Under uniform distribution of C and G (50% GC content), meaning each of the four bases is equally likely to be chosen, we observe that the median susceptibility to the canonical hotspots NTC (hA3B), CCC (hA3G), and WRC (AID) is 0.5 (S3A Fig). With random sequences of varying GC content, however, the susceptibility changes. Randomized sequences of low (40%) GC content showed high susceptibility to the hotspots NTC and WRC, and similarly, high GC content sequences showed low susceptibility (60%). We reaffirmed this by creating new randomized sequences of the resistant, high-GC content EBV and vulnerable, low-GC content OsHV, replacing each codon in each sequence with a synonymous one weighted by the GC parameter (a similar weighing description is given in the Methods under “Calculating Susceptibility”, though here it is applied to the baseline sequences rather than the null model) and observed similar trends, suggesting a strong role of wildtype GC content in explaining APOBEC resistance. We note, however, that this effect is not present for the hotspot CCC. We corroborated these effects by looking at the effect of GC content on codons for existing protein sequences. To determine the effects of native GC content on determining APOBEC hotspot resistance, we additionally calculated susceptibility under a GC-controlled null model (see Methods), weighing the GC parameter to be the mean GC content of the entire gene set. This meant that each gene sequence would now be compared to random synonymous sequences with a chosen GC content. Using this model, we found that changing the GC content of the random sequences affects the observed susceptibilities, reaffirming the importance of GC content in determining susceptibility. If a baseline (wild type) coding sequence is being compared to random sequences with higher GC content, the baseline sequence will tend to show higher susceptibilities to NTC and WRC hotspots (suggesting that lower GC content is more vulnerable to deamination). Despite these trends, the vulnerable OsHV sequence still has higher susceptibilities to NTC and WRC than EBV when GC is corrected (S3B Fig), although are not significantly so (NTC P-val: 0.18, WRC P-val: 0.34, BH corrected t-test). L1 elements, however, have their apparent high susceptibilities to NTC (the hotspot that hAPOBEC3B has which restricts them) neutralized when corrected for in this manner. Thus, high impact of these hotspots on L1 may be a beneficial effect of their inherent low GC content. There is little change in the susceptibilities of HIV to the NTC and WRC degenerate motifs (NTC P-val: 0.93, WRC P-val: 0.82) when we apply this correction, suggesting that while low GC genomes can be resistant to AID/APOBEC hotspots, there can be genomes for which their GC content does not contribute strongly to their vulnerability.
Additionally, we obtained codon usage tables of the coding sequences of OsHV, EBV, and our human housekeeping genes and L1 elements using Emboss cusp (http://www.bioinformatics.nl/cgi-bin/emboss/cusp). When we compared OsHV to EBV, we regenerated sequences in each genome using the codon usage table of the other and calculated the susceptibility of the new sequence. We observed that OsHV, when shuffled with codons with probability proportional to the codon usage bias of EBV, adopts a susceptibility much closer to that of EBV, and vice versa (S3C Fig). We observe similar effects when we exchange the codon biases of HK and L1 elements (two mammalian gene sets that we assume would be differentially targeted by APOBECs) and calculate susceptibilities (S3C Fig). We conclude that the nature of APOBEC hotspot susceptibility is also influenced by codon bias for both virus and vertebrate genes. However, note that the results described above using the GC-corrected model in effect creates a model with codon bias according to wildtype GC content. Thus, it is most likely that our results here using codon bias changes (S3C Fig) are correlated with those using GC-content changes (S3B Fig).
Additionally, codons within the same amino acid can be classified as “intra-codon” or “inter-codon” depending on whether they are a NNC with C to T at the silent position or not. For example, AGC is an intra-codon for serine, since AGC to AGT does not change its amino acid, but AGA would be an inter-codon. Statistical tests of all intra- and inter-codons comparing the vulnerable OsHV and resistant EBV show that OsHV has statistically fewer intra-codons than EBV (difference = 0.182, Fisher Test P < 2.2 × 10−16, 2-tailed Fisher test), providing additional evidence for the role of codon bias in hotspot susceptibility. Finally, our findings on codon bias were corroborated by the trinucleotide distribution of resistant and vulnerable gene sets (“Clustering algorithm”, under Methods).
Selective pressure was also modeled by taking the original EBV sequence, a resistant genome, and inducing mutations at different positions with probability weighed to either favor nonsynonyous or synonymous mutations, according to a parameter we use to approximate the dN/dS ratio. dN/dS represents the ratio of the number of observed mutations at nonsynonymous sites per site, to the number of mutations at synonymous sites per site. Specifically, mutations occurring in the first two positions of a random codon almost always represent nonsynonymous mutations and mutations in the third position are typically (but not always) synonymous. Therefore, we mutated positions along the EBV sequence with probability weighed by a parameter. This parameter is the ratio of mutation at the first two positions to the third, and acts under the assumption that each site is equally likely to be mutated and that this ratio of mutations is correlated to dN/dS. For example, if this parameter is equal to 2, then a mutation at each position of the codon is equally likely, since 2/3 of the mutations would be nonsynonymous, affecting one of the first two positions, and the remaining mutations would affect the third nucleotide, which is often synonymous. The corresponding value for the dN/dS ratio would be closer to 1, as there are twice as many nonsynonymous sites as synonymous. Under these assumptions, dN/dS is roughly half of our selective pressure parameter. For values of this parameter being 0.25 (red), 1 (green), 4 (blue), and 8 (purple) and for up to 500 mutations over time we calculated susceptibilities for all 16 NNC motifs (S3D Fig). Over time, averaged across all these motifs, the low susceptibility changes to a higher susceptibility, which may be explained by a change in GC content to neutral (or 50%) as each of the 12 mutation types (from one base to another) are equally likely. If dN/dS is high, however, the change is smaller. Indeed, the susceptibilities between the observed EBV sequences and EBV after 500 mutations when our selective pressure parameter = 8 (approx. dN/dS = 4) is not statistically significant except for the motifs GCC and GGC (GCC: P = 0.0004694, GGC: P = 0.0053; G2-tail t-test, BH corrected). The difference however between EBV after 500 mutations under high selective pressure (parameter = 8, approx. dN/dS = 4) and low (parameter = 0.25, approx. dN/dS = 0.125) is significant for many motifs including AAC (P = 0.0019), ACC (P = 0.0018), ATC (P = 0.0018), CGC (P = 0.0019), GAC (P = 0.025), TAC (P = 0.00039), TCC (P = 0.0377), and TTC (P = 0.0018). Based on our simulations, high selective pressure appears to play a modest role in maintaining APOBEC hotspot resistance. We note that many mammalian herpesviruses display resistance (Fig 3), and if APOBEC deamination is deleterious to survival of the virus, then there would be evolutionary pressure on the virus to change in such an environment. However, this may be mediated simply by changing the GC content of the virus.
Finally, there are other models of APOBEC/AID activity at hotspots. One of these is the S5F model of Yaari et al. [70] which inferred the mutability and substitution profile during somatic hypermutation among all 1024 possible 5-mers that targeted silent positions and which are therefore assumed to be independent of the effects of selection. Although somatic hypermutation includes not only AID activity, but also non-canonical Base Excision Repair and Mismatch Repair, the model is arguably the most comprehensive available for somatic hypermutation. To investigate the impacts of this activity on a gene’s susceptibility over time, we used this model and induce mutations at each position to the vulnerable OsHV and HIV genomes, resistant EBV genome, and random sequences, with probabilities proportional to the S5F model. This model includes the effects of base excision and mismatch repair which induce mutations beyond that which occur at hotspots, so we also simulate mutations using the same frequencies but only at C to T. We induce a number of mutations equal to up to 60% of the sequence length (to account for different gene lengths). Our results show that when we consider susceptibility to the AID hotspot WRC, AID targeting decreases susceptibility to WRC motifs (S4A Fig), highlighting the possibility that mutations at non-hotspot sites can decrease the susceptibility of a gene. In the case where we apply S5F, but only mutating C to T (and thus also G to A on the opposite strand), we mutate up to 60% of C and G positions instead per gene. Under this mutational model, that ignores the DNA repair mechanisms that act downstream of AID, we observe the opposite trend, namely that susceptibilities to the WRC motif instead increase as mutations accumulate (S4B Fig).
To extend this analysis beyond AID and somatic hypermutation, we also modeled the activity of two other APOBECs. We performed a simple simulation of APOBEC mutation at TC (hAPOBEC3B) and CCC (hAPOBEC3G) hotspots and then calculated the changes in susceptibility to those motifs accordingly. When we mutate the resistant EBV sequence at the hotspot NTC, susceptibility increases to a high level (S4C Fig). Interestingly, we did not see a change in susceptibility to the other hotspot CCC (S4D Fig) for all genomes that we examined. This suggests that regardless of the number of hotspots that are at nonsynonymous or synonymous positions, susceptibility to hotspots can be altered by GC content. The fact that CCC susceptibility, a hotspot which appears neutral to GC content (S3A Fig), does not change by this mechanism lends support for this idea.
Finally, the CpG dinucleotide has biological significance as a marker for DNA methylation and therefore gene activation and inactivation, especially in mammals [71]. We calculate another susceptibility measure that properly controls for this frequency (See: “Controlling for CpG dinucleotide content” under Methods). As shown in S5 Fig, when we consider this, most of the gene sets clustered similarly to the uncorrected method (Fig 3). As far as clustering of the motifs is concerned, we obtained one cluster of 6 motifs (rightmost dendrogram of S5 Fig), containing all 4 ANC motifs including two AID hotspots, AGC and AAC, which strongly overlaps with the cluster of motifs obtained without correcting for CpG (as described above for Fig 3, which contains 3 of the 4 ANC motifs, 3 of 4 TNC motifs, and GTC). These results suggest that the impact of these motifs is still similar when measuring susceptibility controlled for CpG. The gene set clusters also remain the same as the clusters without controlling for CpG content, with the exception of two formerly vulnerable gene sets, MHV68 and HERV which now cluster within the resistant gene sets. These gene sets, respectively a mammalian virus and set of retrotransposons, are co-expressed in environments alongside APOBECs, and in the case of HERV, is strongly edited by them [24]. Since accounting for CpG does not appear to affect the results very strongly, we do not control for this motif in the analyses, although this observation suggests that adapting CpG usage can be a mechanism of reducing effective susceptibility for a subset of viruses, particularly HERV.
Altogether our results suggest that resistance to APOBEC hotspots is a natural consequence of GC content, and provides an explanation of how, under selective pressure, a coding sequence may adopt resistance to APOBECs. These findings are consistent with the results described above (Fig 4) where we found a strong correlation between a genome’s GC content and its motif susceptibility. High GC content, for example, is a feature of several herpesviruses, and a previous study proposes that GC content may play a previous protective role against retrotransposon insertion in these viruses [65]. While we are not suggesting that gene sequences always adopt or evolve their GC content solely for the purposes of APOBEC resistance, we do show that GC content adjustment may confer additional defense against the consequences of restriction by most APOBECs, with the notable exception of the hAPOBEC3G hotspot CCC.
We have introduced a method of quantifying the susceptibility of gene coding sequences to mutational hotspots. We used this method to examine differences in susceptibilities across many host and pathogen genomes and tested both known and hypothetical cytidine deaminase targets. We observed that the known hotspots, particularly the APOBEC hotspots, have a high capacity for inducing non-synonymous mutations in retrotransposons and viruses, while minimizing damage to host genes. Although it needs further investigation, our data suggests a possibility that this selectiveness has been important in shaping AID/APOBEC hotspot preferences during vertebrate evolution.
Susceptibility is potentially influenced by many non-independent factors, including GC content, amino acid usage, and codon bias. Our basic measure of susceptibility is neutral to GC content and codon bias, allowing us to quantify the extent that these factors can evolve to influence susceptibility and its counterpart, resistance. In some cases GC content, via its effects on codon usage, strongly influences susceptibility, while in other cases the effect is less clear (Fig 4A–4D). We expect that a C/G rich hotspot such as the APOBEC3G hotspot CCC should be correlated to the genome’s GC content and indeed, many GC-rich herpesviruses such as HSV-1 (68% GC) and EBV (60% GC) show high susceptibility to CCC (as well as to the hypothetical hotspots GGC and GCC) when we use an alternate definition of susceptibility which considers the absolute number of nonsynonymous C to T mutations, rather than the fraction (S2A Fig). However, the situation is complex since we found cases where GC-rich genomes showed resistance to even GC-rich motifs such as CGC (S2A Fig). We conclude that evolving the optimal fraction of nonsynonymous motifs is more important than the absolute number in order to resist a broad variety of APOBEC mutations, since it is the fraction that is minimized across the resistant gene sets. Alternatively, the motif count (as opposed to the fraction) may still play a role in achieving susceptibility and it is possible a virus may evolve a high GC content to reduce the potential damage of A/T rich hotspots such as TC (human APOBEC3A/APOBEC3B) and TTC (human APOBEC3F). Additionally, tetranucleotide motifs for APOBEC3F and APOBEC3G have been observed [41] where the +1 position is also important, and although we have done the analysis for NNC motifs only, future work might include more complex motifs as well.
We have controlled for the frequency of the CpG motif (see under Methods: “Controlling for CpG dinucleotide content”, S5 Fig), however an extended analysis (that we leave for future work) might also consider the possibility that resistance is a consequence of conservation of the CpG motif itself. CpG sites in vertebrate hosts will often be methylated and are prone to spontaneous deamination. At particular loci, deamination of CpGs may be disadvantageous and evolutionarily selected against. Interestingly, AID and some APOBEC enzymes including APOBEC1can deaminate methylated Cs [4]. Thus, there may be some interplay between spontaneous and AID/APOBEC mediated deamination of methylated CpG sites. Given the additional constraints imposed by the coding region sequence, this interplay is potentially quite complex, and we therefore leave the corresponding analysis for future work.
A high GC content appears to confer resistance to deamination by many of the naturally occurring cytidine deaminases (including hAPOBEC3B, characterized by the hotspot TC and AID, characterized by the hotspot WRC). Additionally, the diversification of APOBEC3 hotspots observed in primates may favor a wider spectrum of intrinsic immune responses, as we have shown that by the same metric, genomes with high GC content remain vulnerable to hAPOBEC3G’s CCC hotspot. However, even when accounting for the endogenous GC content of the gene sets considered, we observed similar, albeit less significant overall trends (S3 Table). Thus we have shown here that there is an advantage to having a high GC content in that it tends to minimize the impact of most APOBEC hotspots. In the future deaminase resistance might also be validated experimentally, for example, by examining the viral fitness of vulnerable or low-GC content herpesviruses in APOBEC-deficient cell lines.
One of our important observations was that retrotransposon ORFs from the L1 families had very high susceptibility scores, with some having average susceptibility as high as 0.9. As our definition of susceptibility is based on the fraction of mutations that cause amino acid changes, one potential explanation for high susceptibility is that past mutations in hotspots have more frequently caused synonymous mutations rather than nonsynonymous ones. This bias would (as expected) cause the remaining hotspots to become enriched for potential nonsynonymous mutations. Also, it would be incompatible with the observation that APOBEC restriction of retrotransposons is usually deaminase-independent, as has been shown in cell culture experiments [72]. However, this observation is in dispute, as in vivo and in vitro assays have shown that hAPOBEC3A can deaminate exposed single-stranded LINE1 DNA [73]. Recent results show that older retrotransposons are often heavily edited for deactivation and that high retrotransposon editing can be beneficial in terms of genome diversification and potential exaptation [66, 74]. Additionally, a recently published study provides phylogenetic and experimental evidence suggesting that even throughout primate APOBEC3A evolution, L1 elements were unable to evolve resistance to APOBEC3A restriction, which may explain our high observed susceptibilities of L1 ORFs to APOBEC3 hotspots [75]. Thus, a high L1 ORF hotspot susceptibility may be beneficial to the host by driving exaptation and new functionality.
One possible implication that arises from our results is that the evolved hotspot preferences may have in turn shaped the genomes of viruses that target those cell types where the deaminases are expressed, suggesting a possible arms race (past or ongoing) between viruses and the vertebrate intrinsic immune systems. It is important to note that while the susceptibility measure may reflect the evolutionary impact of deaminases, it might not always implicate AID/APOBEC as an active agent. The low susceptibilities to canonical hotspots that we have observed in virus genomes that are likely to have ongoing exposure to AID/APOBEC, such as EBV and HSV-1 (Fig 3) suggest that this feature may be evolutionarily advantageous.
We observed that HIV was a vulnerable gene set with high susceptibilities, yet the HIV vif protein neutralizes APOBEC3G, targeting it for degradation, which may abrogate the need to evolve resistance at the genome level. Similarly, another herpesvirus we examined, MHV 68 (a strain of Murid Herpesvirus 4), showed high susceptibility to APOBEC hotspots including mouse APOBEC3 hotspots defined by the motif TTC. A recent study showed that the mouse APOBEC3 does not restrict this virus either in cell culture or in vivo [63], suggesting the possible existence of an as yet unidentified APOBEC evasion mechanism. MHV68, alongside the transposon-like HERV elements, also show reduced susceptibility when we control for CpG dinucleotide motif frequency (S5 Fig). Thus, for some genomes the CpG dinucleotide frequency appears to be an important constraint on codon selection which may also provide reduced susceptibility to APOBEC hotspot mutations. Thus, an apparently vulnerable genome may still adapt some level of resistance to APOBECs by adjusting its CpG usage accordingly, although we only observed this for two vulnerable viral genomes.
We demonstrate that a high GC content and the corresponding observed codon biases are a mechanism by which resistance is conferred (Fig 4 and S6 Fig). Not all hotspot resistance is explained by GC content, and we observe that varying the GC content of random sequences does not change susceptibility to the hAPOBEC3G hotspot CCC. Under our model of susceptibility, which is based upon the percentile of the fraction of nonsynonymous mutations, CCC may be robust to changes in GC content. For example, deamination at a CCC hotspot codon to CCT might be synonymous, whereas if the hotspot were on the opposite strand (and in frame) forming a GGG codon, then deamination to AGG will be nonsynonymous. Since the frequency of both CCC and GGG codons will presumably change linearly with GC content then the fraction of nonsynonymous mutations should remain approximately unchanged for different GC levels. Although this argument applies only when the hotspot is in frame, in practice we find that CCC hotspots are indeed robust to changes in GC content regardless (S3A Fig). Thus, there may be alternating roles for the different APOBEC hotspots, with human APOBECs that prefer NTC to be primed to target low GC content virus sequences and retrotransposons, and human APOBEC3G which may have evolved the CCC hotspot to be robust to the protection a high GC content may confer to other hotspots. Notably, human APOBEC3G is known to restrict HSV-1 in vitro [31], which is a high GC content virus [65]. Due to the differences in susceptibility to the hAPOBEC3G hotspot CCC and the hAPOBEC3B hotspot NTC, there may be an evolutionary advantage in there being multiple APOBEC hotspots since this increases the breadth of viruses (meaning variability in GC content) that the host can optimally restrict. This result is consistent with previous suggestions that the diversification of APOBEC3 genes has arisen due to an arms race against a wide variety of viral pathogens, given that many of the diversified APOBEC3 genes are under positive selection [2, 76, 77].
Our model and definition of susceptibility examines the impact of mutations on both strands, weighing both equally. This model has been sufficient to demonstrate overall trends and differences between a variety of mammalian and non-mammalian genes and viruses. However, in the particular case of retroviruses, APOBEC deamination is biased in that mutations are predominantly G-to-A on the plus strand, as has been observed for HIV and HTLV [78]. Thus, an alternative model that might be considered would examine susceptibility only on one strand. Under this scenario, differences in susceptibility between different retroviruses and among viruses from different clinical isolates might be used as additional evidence of an ongoing arms race between the deaminase and the retrovirus.
Although mutations are likely to be predominantly deleterious, viruses in particular need to maintain variation at the population level, and often within an individual host to escape host immunity. For HIV in particular, there is growing evidence that APOBEC-mediated mutagenesis may actually contribute to the generation of escape variants [27, 64]. Based on the idea that there is an optimal mutation level for viruses to propagate [79], it is conceivable that APOBEC susceptibility may benefit many viruses, particularly if we were to consider individual genes (that contain epitopes, for example) rather than entire genomes as we have done here. Many mammalian genes often showed great differences in susceptibilities to different hotspots as well. In the future, predicting which genes are particularly susceptible may illuminate the overall impact of AID/APOBEC mutagenesis on the mammalian genome as well. Our results suggest that the current AID/APOBEC hotspots are effective in mutating unadapted viruses and retrotransposons, while minimizing collateral damage. However, several of the viruses that we expected would be vulnerable are resistant and vice versa (such as vulnerable HIV, MHV4, and SFV, which are all mammalian viruses). These features may be explained by alternative APOBEC evasion mechanisms, such as the HIV gene vif. A similar mechanism was recently proposed to exist for MHV4 [63]. We expect future work, that uses more sophisticated models of sequence evolution and mutation and on more thorough strains of viruses, to elucidate these differences further.
We describe a model that calculates the susceptibility, UH, to motif H, from NNC to NNT, a wildtype coding sequence (denoted as swt), that contains L codons as follows. Given swt, we generate a vector:
CL= (c1,c2,…cL)
where ci is the ith codon in swt. We generate a set W, of N (in practice 1000) random sequences: W = (s1, s2, … sN) where each sequence sj, in W is a sequence generated from swt by randomizing codon usage as follows. For each sequence sj, for each codon ci in CL we replace the codon with another synonymous codon chosen either with uniform probability or a biased probability derived from a GC-bias parameter (see next paragraph). For each sequence sj, the k occurrences of (real or hypothetical) hotspot motif H are counted. Then for each sequence sj we associate with it a vector Hk = (h1, h2, … hk), where ht is the position of the tth occurrence of hotspot motif H on sequence sj, in both orientations. We define a function f(ht) which is 1 if changing the nucleotide at position ht from C to T (or G to A), causes an amino acid change, and is 0 otherwise. Then the nonsynonymous mutation fraction of sequence sj is:
M(sj) = ∑t=1kf(ht)k
The susceptibility, UH of wildtype sequence swt to the hotspot H, is calculated as:
UH = ∑l=1NR(l)N
Where R(l) = 1 if M(Sl) < M(swt), (M(swt) being the nonsynonymous mutation rate of the wild type sequence), and 0 otherwise.
Additionally, we may generate random sequences taking into account GC content, by assigning a weight to each synonymous codon. Given a GC content (fraction) of p, and a set of synonymous codons for a particular amino acid (e.g. GAG and GAA for Glu), a weight is derived for each codon based on the GC content of the codon. Within each codon, we assume that G and C nucleotides are chosen according to GC content with probability p, whereas A and T nucleotides are chosen with probability 1-p. The codon weight is the product of the three corresponding probabilities. Thus, for Glu, the codon GAG would be assigned a weight of p×(1-p)×p = (1-p)p2, whereas the codon GAA would be assigned the weight p(1-p)2. Codons are then selected with probability equal to a normalized vector of these weights. Note that if p = 0.5, each synonymous codon is selected with equal probability.
As described herein, susceptibility is impacted in many cases by GC content (Fig 4). We evaluated the larger gene sets on a gene-by-gene basis to assess the impact of GC content on susceptibility. These large gene sets include the human and mouse housekeeping gene sets and several virus genomes (S6 Fig). Although not every hotspot we examined demonstrated a strong relationship, the TTC motif had the strongest mean correlation among the 16 NNC motifs we examined, in the human and mouse housekeeping gene sets (Spearman correlation = -0.728 and -0.486 respectively, S6A and S6B Fig). For many of the viral genomes we analyzed, the GC content fell within a very narrow range and/or had very few genes, leading to weak correlations. We therefore grouped the genomes of EBV, HSV-1, Murid Herpesvirus 2, and the Ostreid herpesvirus, as well as many non-herpesviruses with 10 or fewer genes including HIV, Adenoviruses, and others to obtain a wider range of GC content across many genomes. By analyzing all the genes together we observed a trend similar to that for the housekeeping genes, where GC content of the gene correlates with the motif susceptibility (-0.493 for the TTC hotspot, S6C Fig). Because GC content influences susceptibility in some cases, but not others (Fig 4B), we chose in the first instance to keep our analysis neutral and we generated our null model without any constraints on GC, uniformly choosing each synonymous codon at random. We subsequently compared our susceptibility results to the GC content of the corresponding gene set to determine the extent to which GC content determines susceptibility (Fig 4B).
To control for CpG dinucleotide frequency we are interested in calculating susceptibility for each NNC motif conditioned on the CpG frequency in the original sequence. It is usually impractical to compute this conditional probability empirically. Controlling for this condition might be achieved by considering only those randomly generated sequences that have, by chance, produced the same number of CpG motifs as the original sequence. However, in practice the fraction of random sequences that exactly match the CpG motif count is often very small, particularly if the CpG content is already skewed, making it unlikely that such exact matches are produced by the random model. This may lead to very small numbers, making an approximation preferable. Thus, we approximated the joint distribution of CpG dinucleotide frequency and susceptibility, assuming the joint distribution is a bivariate normal distribution. The corrected susceptibility is simply the percentile of the approximated Gaussian conditional on the CpG frequency in the analyzed sequence. Correcting for CpG frequency made little difference to our results.
We obtained protein-coding sequences from a variety of biological sources, and broadly speaking our data sources are categorized into three groups: Known or suspected AID/APOBEC targets including viruses and repeat elements; non-targeted viruses; and putative unintentional or collateral targets of AID/APOBEC which include host genes. Known virus targets that we analyzed included HSV-1, AAV-2, HERV (specifically the HCML-ARV retrovirus) and SFV. Since papillomaviruses are highly diverse we selected a well- characterized strain with medical relevance, which was HPV16 due to its link to cervical cancer. We also analyzed EBV, due to its shared B-cell tropism with AID. MHV68, a close relative of the other herpesviruses we studied, is not suspected to be restricted by mAPOBEC3 although multiple human APOBECs were capable of restricting the virus [63], so to test the possibility that mAPOBEC3 hotspots may have impacted murid herpesvirus coding sequence vulnerability or vice versa we included both Murid herpesvirus 2 and MHV68. As controls, we included several viruses unlikely to be restriction targets, due to not being vertebrate or mammalian viruses, which consist of the Ostreid herpesvirus, as well as two plant viruses, the cherry and strawberry viruses (Table 2). These controls were evaluated by substituting other potential controls consistent with the same rationale, i.e. they are not restricted or exposed to AID/APOBEC. We also analyzed a set of potential mutational targets in the mouse genome, namely a data set that is enriched in potentially oncogenic AID off-target genes [56] and sets of housekeeping genes from the human and mouse genomes (see below). Altogether we examined a total of 22 distinct gene sets. We wanted to avoid including an overwhelming number of controls, but at the same time we checked whether those we did include were unusual. Therefore, additional control viruses including plant viruses were obtained and individually verified by including these viruses in the clustering analysis (Fig 3) and determining whether they clustered consistently with the viruses originally used as controls. In each case these clustered with our previous control viruses, demonstrating the validity of the original choices (S7 Fig). Additional controls used to validate our current control set included: Tobacco mosaic virus (NC_001367), Grapevine leafroll associated virus 1 and 3 (NC_016509, NC_004667), Citrus tristeza Virus (NC_001661), Carrot yellow leaf Virus (NC_013007), and Fly C virus (NC_001834). Housekeeping genes for the vertebrate zebrafish, which lacks APOBEC3, and the invertebrate C. elegans were obtained from a RT-PCR array (http://www.sabiosciences.com/rt_pcr_product/HTML/PAZF-000Z.html) and a study of C. elegans housekeeping genes respectively [2, 80]. The list of human housekeeping genes were identified in a previous study [81]. In cases where a gene had multiple splice variants, we used the longest sequence available (UCSC Genes track/ knownGenes table via the UCSC genome browser query tool). For the set of mouse housekeeping genes, we used a previously published list which assigned probabilities of being housekeeping to each gene [82]. We used a threshold probability of 0.75 in order to obtain a gene set of comparable size to our human housekeeping gene set. Again, the longest sequence was used in case one gene had several coding sequences. L1 sequences for both mouse and human were downloaded from the UCSC genome browser database under the RepeatMasker table from the mm10 and hg19 databases respectively. Sequences were filtered with a minimum of 5500 base pairs, to ensure only full length elements, which are canonically around 6000 bp long. ORFs were extracted using EMBOSS getorf, with a minimum length of 1000 nucleotides finding all sequences between start and stop codons. These ORFs were separated based on their family and all ORFs were subsequently separated by family. In total 14 families were identified, including members from the L1PA and L1PB families identified in [52]. For each family, sequences were clustered. Representative sequences were processed to reduce statistical redundancies between repetitive sequences using cd-hit with default parameters [83]. The final L1 set was obtained by merging the representative sequences from each cluster.
The data being clustered was a matrix of the average susceptibility scores, which range from 0 to 1, of all 16 NNC motifs in the columns, and gene sets (e.g. L1, Housekeeping genes, etc.) in the rows. Thus, both rows and columns were clustered using standard hierarchical agglomerative clustering and average linkage, using the Euclidean distance metric as the distances between each row and column. These clusters were computed and displayed using the heatmap.2 function in the R library gplots. We further validated these clusters using a k-means approach, showing that when we tested across a range of values for k, 2 clusters represents an optimal separation of the gene sets, indicated by the largest decline in intra-cluster sum of squares (S8A Fig). As an alternative approach, we used Principal Component Analysis (PCA). Here, one principal component accounts for a large amount of the overall trend in motif susceptibility (S8B Fig), and the clusters of gene sets are consistent for the different methods and separate primarily along the first principal component (S8C Fig).
Codon bias is likely correlated with the more general measure of trinucleotide (NNN) frequency. We confirmed this by calculating the trinucleotide frequency usage of all of our gene sets and clustering them similarly (S9 Fig). The occurrence of the trinucleotide AAA was skewed for the L1 elements for human and mouse, so it was excluded as an outlier. We observed the same clusters as before in Fig 3, except that Hepatitis B virus, and HERV, which under susceptibility are grouped with the vulnerable clusters that include L1 elements among others, now appears in the resistant cluster that includes housekeeping genes and mammalian viruses.
To assess statistically whether the rankings of certain groups of the 16 possible NNC (for example, the 4 human APOBEC3B NTC hotspots of ATC, CTC, GTC, and TTC) motifs were significantly high or low, we performed a pairwise comparison between each of the 10 resistant and the 12 vulnerable genomes of Fig 3. For each comparison, we ranked the motifs by the difference in average susceptibilities between the resistant and the vulnerable gene set (resistant-vulnerable), noting the ranks of the 4 motifs in each comparison. We assessed significance by sampling one resistant and one corresponding vulnerable gene set chosen at random, until one of each resistant gene set was sampled, and ensuring no gene set was sampled more than once. In each comparison we noted the ranks of the 4 motifs, summed them and compare the sum to a random null model selection of ranks, taken 4 at a time uniformly from the numbers 1 through 16 without replacement. For example, if examining the statistical significance of the motif NTC, the ranks of the hotspots ATC, CTC, GTC, and TTC, for a particular resistant vs vulnerable gene set, were 1, 3, 5, and 9, these would be compared to a random selection of numbers 1–16 which could be 3, 8, 9, 15. The sums of these ranks would be 18 and 35 respectively. We then summed the ranks of the 4 hotspots across all our comparisons and compared this to the sum of null model ranks. Repeated 10000 times, our bootstrapped P-value was the fraction of times the sum of observed ranks found to be greater than the sum of the random ranks. If this fraction (equivalent to a P-value) were greater than 0.05, this demonstrated that the observed ranks were not significant.
|
10.1371/journal.pgen.1000307 | A Simple Screen to Identify Promoters Conferring High Levels of Phenotypic Noise | Genetically identical populations of unicellular organisms often show marked variation in some phenotypic traits. To investigate the molecular causes and possible biological functions of this phenotypic noise, it would be useful to have a method to identify genes whose expression varies stochastically on a certain time scale. Here, we developed such a method and used it for identifying genes with high levels of phenotypic noise in Salmonella enterica ssp. I serovar Typhimurium (S. Typhimurium). We created a genomic plasmid library fused to a green fluorescent protein (GFP) reporter and subjected replicate populations harboring this library to fluctuating selection for GFP expression using fluorescent-activated cell sorting (FACS). After seven rounds of fluctuating selection, the populations were strongly enriched for promoters that showed a high amount of noise in gene expression. Our results indicate that the activity of some promoters of S. Typhimurium varies on such a short time scale that these promoters can absorb rapid fluctuations in the direction of selection, as imposed during our experiment. The genomic fragments that conferred the highest levels of phenotypic variation were promoters controlling the synthesis of flagella, which are associated with virulence and host–pathogen interactions. This confirms earlier reports that phenotypic noise may play a role in pathogenesis and indicates that these promoters have among the highest levels of noise in the S. Typhimurium genome. This approach can be applied to many other bacterial and eukaryotic systems as a simple method for identifying genes with noisy expression.
| According to the conventional view, the characteristics of an organism are determined by nature and nurture—by its genes and by the environment it lives in. Consequently, one would expect that two organisms that share the same genes and live in the same environment have identical characteristics. Recently it has become clear that this expectation is often not borne out; clonal families of simple organisms living under constant conditions often show variation in biological traits and sometimes even have markedly different properties and do different things. In order to investigate molecular causes and possible biological functions of such phenotypic noise, it would be very valuable to have a simple and fast method for identifying biological traits that are particularly noisy. Here, we developed such a method, and screened for noisy traits in the bacterial pathogen Salmonella typhimurium that vary at a time scale of one day. We found that traits involved in interaction with the host are particularly noisy, suggesting that phenotypic noise might be important in pathogenesis. This method can be readily adopted for other organisms and might contribute to elucidating the role of noise in biology.
| Clonal populations of unicellular organisms growing under constant conditions often show substantial variation in phenotypic traits. The rate at which some of these traits vary is so high that it cannot result from mutational change. Rather, this phenotypic noise has been shown to result from chance events in the cells, namely random fluctuation in the transcription and translation of genes [1]–[3]. Most of the research on phenotypic noise focuses on two questions. First, what are the molecular processes underlying this phenomenon? Second, are there cases in which phenotypic noise is beneficial? Can it provide a genotype with new biological functions and improve the chance that it will survive and reproduce?
To further our understanding of the biological significance of phenotypic noise, it would be helpful to have a simple method to identify genes whose expression varies stochastically at a given timescale and under specific environmental conditions. So far, most of the research on phenotypic noise was based on the detailed analysis of individual biological traits [4]–[6]. It is interesting to complement these studies with a global analysis, so that one can ask whether the traits studied so far are indeed particularly noisy, or whether a substantial fraction of all genes show such high levels of noise. One possibility for a global analysis of phenotypic noise is the exhaustive characterization of ordered libraries of strains marked with reporter proteins [7]. Here, we have established a simple alternative that allows identifying promoters whose activity varies on a specific time-scale; we used this method to identify promoters in the bacterial pathogen S. Typhimurium that switch between active and inactive over the course of a few generations.
The method is based on subjecting a promoter library to selection for high levels of random variation on a short time scale. The screen was initiated with a genomic library consisting of short genomic fragments upstream of a gene encoding green fluorescent protein (GFP). Cells carrying a fragment with an active promoter expressed GFP. In order to select for promoters with a high level of phenotypic noise, we used fluorescence-activated cell sorting to select cells based on the cellular concentration of GFP, and alternated between selecting for high levels of GFP, and selecting for low levels of GFP. There was no signal indicating the direction of selection during a given round of the selection experiment; one would thus expect that promoters that randomly switch between expressing and not expressing GFP would increase in frequency.
This screen led to a strong enrichment of promoters with high levels of noise. The promoters that showed the highest levels of noise were found to be flagellar promoters, which are involved in the interaction with the host. These promoters have previously been reported to be heterogeneously expressed in clonal populations of S. Typhimurium. Our screen demonstrates that these promoters stand out in terms of the level of noise, and that they vary on a very short timescale. This method thus offers a simple and powerful approach to identify genes with high levels of noise, and allows for easily modulating timescale and environmental conditions under which such phenotypic noise manifests.
We established a population of approximately 7×106 S. Typhimurium clones containing a library of genomic fragments ranging in size from 400 bp to 1200 bp linked to a GFP reporter (see Methods). In order to enrich for clones exhibiting increased levels of phenotypic noise in GFP concentration, we used a regime of alternating selection. Cells were grown into exponential phase, and subjected to selection on GFP concentration in a fluorescence-activated cell sorter (FACS). First, we selected only those clones having a level of GFP expression in the highest 5% of the population; these clones were saved and used to inoculate fresh cultures that were grown overnight. In the next step, the opposite selection regime was imposed, such that only those clones having a level of GFP expression in the lowest 5% of the population were saved and grown. It is also possible to first select cells expressing low levels of GFP, and then high, which hypothetically would result in noisy promoters with lower average expression.
This process of fluctuating selection was repeated, until a total of seven alternating selection events had occurred. The fluctuating selection regime was performed on five separate populations; five control populations were also exposed to the same regime of growth and FACS sorting, but no selection occurred for the level of GFP concentration (a random subset of cells covering the entire range of GFP fluorescence was saved and grown). After the seven rounds of selection, clones from all populations were plated onto agar plates. Twenty-four clones from each of the ten populations were randomly selected for future analyses.
Selection for increased phenotypic noise can only be successful if the level of variation is a stable property of a clone. We thus first asked whether the level of phenotypic noise in GFP expression was a stable and consistent trait in these clonal isolates. We used the 240 frozen clonal stocks described above to seed fresh cultures of cells, and analyzed GFP concentration for about 5×105 cells per clone (see Methods). We repeated the same procedure on a different day, and also gathered data on GFP expression for the same set of 240 clones. Phenotypic noise was quantified using the coefficient of variation in GFP expression from a subset of cells similar in size, shape, and cellular complexity (see Methods). We found that the level of phenotypic variation observed for a given clone on day 1 was highly correlated with the level observed on day 2 (r2 = 0.748, p<0.001). This shows that the level of phenotypic noise is a consistent property of a clone (presumably reflecting the noise of the promoter on the genomic fragment it contains), and that this property is stably maintained in clonal populations that are repeatedly grown from an individual cell.
Next, we asked whether fluctuating selection had led to an enrichment of clones exhibiting larger amounts of stochastic phenotypic variation. We compared the clones from the five selected populations to the clones from the five control populations. Among the clones from the selected populations, a sizable fraction showed high coefficients of variation in fluorescence, which is a measure of stochastic phenotypic variation. In contrast, the control population did not contain any clones with high coefficients of variation (Figure 1, Figure 2). An analysis of variance showed that the average coefficient of variation was higher in the selected than in the control populations (p-value = 0.016, by GLM univariate). This demonstrated that fluctuating selection on fluorescence enriched for strains with high levels of stochastic variation in this trait.
This simple selection scheme is thus a good tool for enriching for noisy promoters. Identifying the genes controlled by these promoters then gives a fairly unbiased look at genes whose expression is particularly variable, and might thereby provide new insights into the biological role of noise. In order to identify these genes, we sequenced the library inserts from the 240 frozen clonal stocks (24 from each experimental population). We found that the clones exhibiting the highest levels of variation were dominated by two promoter sequences that regulate genes involved in flagellar synthesis, namely fliC and to a lesser extent flgK (Figure 1; Table S1). On the other hand, none of the inserts sequenced from the control populations contained promoters associated with the expression of flagellar or related genes, suggesting that this result was not due simply to overrepresentation of flagellar promoters in the genomic library.
We focused on fliC for two tests of the robustness of our results. First, we tested whether the fliC promoter is also noisy in the native chromosomal context. To do so, we constructed a transcriptional fusion of gfp to the fliC promoter at its native location in the chromosome. Clones from this chromosomal construction showed very similar levels of phenotypic noise to the plasmid-based fliC promoter (Figure S1, Text S1). Second, we asked whether GFP expression from the plasmid is correlated with actual protein production of FliC. Clones containing the pfliC-GFP insert in the plasmid pM968 with high levels of variation in GFP expression were sorted into three fractions (expression of GFP, no expression of GFP, and cells expressing all levels of GFP). Western blot analysis with anti-FliC, anti-fljK antibodies on these three cell fractions confirmed that GFP expression is positively correlated with FliC protein production. (Figure S2 and Text S1). These two experiments indicate that the levels of noise we measured are, at least in the case of fliC, not an artifact of the plasmid-based reporter system, but do reflect actual differences in protein production between cells.
The variation in the expression of GFP under the control of flagellar promoters observed here is reminiscent of a genetic switch known as phase variation. S. Typhimurium express two distinct flagellin proteins, FliC and FljB [8], and switches between the two flagellar types using a site-specific recombination event in the chromosome. Can phase variation account for the phenotypic noise that we measured in the clones harboring the flagellar promoters? Site-specific recombination occurs at a rate of 10−3 to 10−5 per cell division [9]–[11]. In a clonal population grown from a cell in one phase, it thus takes many divisions until recombination-mediated phase variation has a reasonable likelihood of occurring. However, this is not what we observed in the clones with flagellar promoters: populations grown from single cells quickly attained substantial proportions of cells with both high and low expression of GFP (Figure 3, Movie S1). In contrast, clones isolated from the control populations maintained similar levels of GFP expression (Figure S3, Movie S2). This suggests that it is unlikely that the variation observed in these clones can be attributed to phase variation.
As a direct test of the effect of phase variation on stochastic phenotypic variation, we transformed the plasmid with the fliC promoter controlling GFP expression into a host strain that is incapable of phase variation [8] and into a wildtype strain. The resulting populations still showed strong variation in the amount of GFP between cells, and the coefficient of variation was not significantly different between the plasmid containing the fliC promoter in the wildtype background and the strain incapable of phase variation (t-test, p = 0.199, 95% Mean CV for wildtype background is 1.07, mean CV for fljAB promoter “locked” off background is 0.95, 95% Confidence interval for the difference is 0.068 and −0.299). This demonstrates that phase variation is not the main reason for the phenotypic noise observed here, and is most likely not involved.
Having identified promoters that are particularly variable, one can then ask whether variability in these promoters might serve a biological function. This question can be addressed by functional studies of the genes whose expression is particularly noisy. However, first insights can be gained from looking at the types of promoters that showed the highest levels of stochastic phenotypic variation.
By far the highest level of phenotypic noise observed in our experiment comes from flagellar promoters, most notably, fliC. This supports a previous report that the expression of FliC is heterogeneous in clonal populations of S. Typhimurium [12]. Bacterial flagella are required for colonization and tissue invasion [13],[14] and they interact with the host immune system in a myriad of ways, eliciting both innate and specific immune responses [15]–[18]. That variation in the expression of flagella might be advantageous is a well-established concept [19]; it usually refers to variation mediated by a site-specific recombination event, but has recently also been extended to variation that presumably does not involve changes in the DNA sequence [12],[20].
The advantage that is usually postulated is mediation between conflicting selection pressures on flagellar expression in the host. During initial stages of gut infection by S. Typhimurium, flagella are instrumental for swimming towards the host's epithelial mucus layer [14]. During later stages of infection, a switch towards not expressing flagellin might be of advantage for bacteria that have invaded epithelial tissue, as it avoids recognition by the innate immune system [TLR5, Naip/Nalp][21]. There is a second possible biological function of phenotypic noise in flagella and other factors involved in the interaction with the host. A recent study suggested that heterogeneous expression of these traits in clonal populations of S. Typhimurium promotes the division of labor between two phenotypically different subpopulations. One subpopulation invades the gut tissue and elicits an inflammation of the gut; the other subpopulation remains in the gut and benefits from the fact that the inflammation reduces competition from commensal bacteria [22].
Two main insights emerge from this study. The first insight is that the activity of some S. Typhimurium promoters varies on such a short time scale that these promoters can absorb rapid fluctuations in the direction of selection, as imposed during our experiment. This is an important experimental test of one of the main ideas for why phenotypic noise can be adaptive: variation in the phenotypes encoded by a single genotype can increase the long-term growth rate of this genotype in fluctuating environments [23],[24].
The second insight is methodological: fluctuating selection is a simple and fast tool to screen large pools of individuals in order to identify variable promoters in unicellular organisms, and thus complements exhaustive characterizations of individual genes [7]. Exhaustive characterizations require the construction of ordered libraries in which fluorescent markers are transcriptionally or translationally fused to every gene, as well as individual measurement of all resulting strains. In contrast, the method presented here only requires the relatively simple construction of a random genomic library, and sorting of the pooled library. It is thus also applicable to eukaryotic systems and organisms that are not genetic model systems, as long as they can be stably transformed. It should thus be feasible to identify noisy promoters in a diverse range of environmental, commensal, and pathogenic organisms, and to ask whether differences in the lifestyle lead to consistent differences in the types of genes that are variable.
One particular advantage of this tool is that the time-scale at which the direction of selection changes can be varied. By changing the direction of selection every few cell divisions, on can impose selection for promoters that switch at a very high rate; changing the direction of selection less frequently selects for promoters that switch at lower rates. It should thus be possible to identify promoters that vary at different time scales, and to investigate whether they might be associated with responses to environmental conditions that vary at different frequencies.
Once noisy promoters are identified, functional studies are needed to investigate the biological consequences of their variation. This might lead toward new answers to one of the fundamental and most challenging questions about the biology of noise – whether phenotypic noise is beneficial, and what its possible benefits might be.
Strains were grown at 37°C on LB agar plates or in 1 ml of liquid LB broth in 5ml polystyrene round bottom tubes (BD Falcon), with shaking at 200 rpm until mid-exponential phase. Ampicillin (Sigma) was used at a concentration of 100 µg/ml in strains containing plasmid pM968 or its derivatives.
A plasmid library (7×106 clones) was constructed by partially digesting S. Typhimurium SL1344 wildtype [25] chromosomal DNA with Bsp143I. Fragments within a size range of 400 bp to 1200 bp were ligated into BamHI digested pM968. This plasmid is low copy number promoter-less derivative of pBAD24 containing promoterless gfpmut2, described in [26]. Plasmids were transformed into E. coli Χ6060, re-isolated by standard methods and electrotransformed into S. Typhimurium M324 (Δ aroA invC::aphT ssaV::cat [26]). Colonies were selected by growth on LB agar plates containing Ampicillin, harvested, and pooled.
A 1∶1000 dilution of an overnight culture of the plasmid library was split into ten equal populations; five populations were assigned to “selected” and five to “control” groups. Cells were grown for 2 hours to reach exponential growth. Cultures were spun down at 3000× g for five minutes at 4°C. Growth media was removed and cultures were re-suspended in ice cold PBS. Cells were kept on ice until sorted or analyzed as described below.
We subjected the plasmid library to fluctuating selection on fluorescence intensity, where selection for bright cells alternated with selection for dim cells.
Cells were sorted using fluorescence-activated cell sorting (FACS) with FACS–Diva sorting software (Becton Dickinson, CA). Immediately prior to sorting, 5×105 cells from each of the ten populations were analyzed for GFP expression. Based on this analysis, on the first day, a gate was drawn for each population to include either the highest 5% of cells expressing GFP, or a gate that covered the entire range of GFP expression, for selected and control lines, respectively. From each gated area, 1×105 cells were collected into a sterile well of a 24-well plate. Cells were collected at a 2.0 flow rate and sorted on the basis of “single cell” and “purity”. After sorting, cells were spun at 3000 g for ten minutes and any FACS buffer was removed. Cells were re-suspended in 1ml LB media containing Ampicillin and grown overnight. The following day the process was repeated; however the gates for the selected populations included only the lowest 5% of cells expressing GFP. This process was repeated for a total of seven rounds of selection, with gates being drawn for selected populations in a fluctuating manner: selection on the highest 5% of GFP expression, then lowest 5%, and back again to the highest 5% of the total. After the 5th round of selection all populations were placed at 4°C for 48 hours. After this time, selection was resumed as normal. After all rounds of selection were completed, the populations were plated on LB agar plates containing Ampicillin, and 24 single colonies from each experimental population were randomly selected (240 clones in total). These were grown overnight in 1ml of LB containing Ampicillin and frozen at −80°C in 15% glycerol.
One day prior to analysis, the 240 frozen clonal stocks were used to inoculate 1ml of medium in 5ml polystyrene round bottom tubes (BD Falcon) and prepared in the same manner as described above (Growth for cytometry and cell sorting). For each clone, 5×104 cells were analyzed for GFP expression on the FACS Calibur (BD, CA).
Raw data was exported from FlowJo 4.6.1 software (TreeStar, Ashland, OR) into custom software. The software was used to exclude data deemed to be extraneous and for performing calculations relating to noise in fluorescence intensity.
The following conventions were applied to calculate variation in GFP expression and to limit the influence from cellular aggregates, cell detritus, and undefined values. Modified from Newman et al [7]:
When calculating the correlation between the coefficients of variation in fluorescence on two consecutive days, two data points were excluded from the analyses because they were more than 3 standard deviations away from expected values.
The following primers (F: 5′ GTCAGAGGTTTTCACCGTCATCAC 3′. R: 5′CAAGAATTGGGACAACTCCAGTG 3′) were used to PCR amplify the genomic segments inserted into plasmid pM968. Both primers anneal to regions on pM968 that flank the insert region. Inserts were sequenced using the reverse primer. The insert sequences were blasted against the genomic sequence of Salmonella typhimurium LT2 genomic and plasmid sequence (accession numbers NC_003197 and NC_003277), and the single best hit was retained as a hypothetical promoter. For each of these hypothetical promoters, the two nearest downstream genes were checked to see if either were oriented in the same direction as the hypothetical promoter. If either of these genes were oriented in the correct direction, the name and distance to the closest gene was noted. If neither of these genes were oriented in the correct direction, we concluded that it was unlikely that the insert sequence was actively driving transcription.
Cell tracking software was used to track cell lineages and analyze GFP expression in individual cells during microcolony growth as described in [27].
|
10.1371/journal.pcbi.1003642 | The Natural History of Biocatalytic Mechanisms | Phylogenomic analysis of the occurrence and abundance of protein domains in proteomes has recently showed that the α/β architecture is probably the oldest fold design. This holds important implications for the origins of biochemistry. Here we explore structure-function relationships addressing the use of chemical mechanisms by ancestral enzymes. We test the hypothesis that the oldest folds used the most mechanisms. We start by tracing biocatalytic mechanisms operating in metabolic enzymes along a phylogenetic timeline of the first appearance of homologous superfamilies of protein domain structures from CATH. A total of 335 enzyme reactions were retrieved from MACiE and were mapped over fold age. We define a mechanistic step type as one of the 51 mechanistic annotations given in MACiE, and each step of each of the 335 mechanisms was described using one or more of these annotations. We find that the first two folds, the P-loop containing nucleotide triphosphate hydrolase and the NAD(P)-binding Rossmann-like homologous superfamilies, were α/β architectures responsible for introducing 35% (18/51) of the known mechanistic step types. We find that these two oldest structures in the phylogenomic analysis of protein domains introduced many mechanistic step types that were later combinatorially spread in catalytic history. The most common mechanistic step types included fundamental building blocks of enzyme chemistry: “Proton transfer,” “Bimolecular nucleophilic addition,” “Bimolecular nucleophilic substitution,” and “Unimolecular elimination by the conjugate base.” They were associated with the most ancestral fold structure typical of P-loop containing nucleotide triphosphate hydrolases. Over half of the mechanistic step types were introduced in the evolutionary timeline before the appearance of structures specific to diversified organisms, during a period of architectural diversification. The other half unfolded gradually after organismal diversification and during a period that spanned ∼2 billion years of evolutionary history.
| Structural phylogenomics enables one to construct a historical timeline of the structural scaffolds known as protein folds and of the biocatalytic mechanisms that are embedded in them. This timeline defines a natural history of biocatalysis through its most granular components, the mechanistic steps. This history reveals an explosive diversity of catalytic mechanisms, which are used in a combinatorial manner in the different chemical reactions of the emergent metabolic networks. This evolutionary “big bang” of mechanistic innovation of protein reaction chemistries was based on mechanistic steps that were probably recruited from primordial chemistries that already existed on Earth, contributing uniquely and very early to life's nascent metabolic repertoire. This can benefit our understanding of protein structure–function relationships and of the origin of modern biochemistry.
| The three-dimensional (3D) atomic structures of contemporary proteins provide clues about how both structure and function unfolded in the course of billions of years of evolution [1]. The phylogenomic analysis of protein domain occurrence and abundance in modern proteomes [2], [3] enables retrodictive views of protein evolution that are unanticipated [4], [5] and can be used to study structural change and the relationship between protein structure and function [6]. Two recent studies of this kind showed congruently that the α/β architecture is probably the oldest type of fold design [2], [3].
An interesting observation [3], [7], regarding the Enzyme Commission (EC) [8] definition of the overall function of enzymes, is that the oldest fold structures were associated with the largest number of enzyme functions [3], [7], [9], [10]. The EC classification provides functional annotations that can be used to link a gene with the chemical reaction catalysed by its gene product. However, the EC classification does not explore the detailed chemical mechanism of the enzyme reaction. Indeed, the classification was designed before much information concerning enzyme structures [11] and mechanisms [12], [13] was available.
Understanding how enzymes adapt their chemical mechanisms under evolutionary pressure is still a challenging task in molecular biology. In this study, we explore the chemical mechanisms used in biochemical reactions catalysed by ancestral enzymes. We ask questions about the ways in which enzyme structure and chemical mechanism have evolved together, and about the evolutionary origination of new enzyme structures and new catalytic mechanisms. MACiE [12], [13] definitions of enzyme mechanisms and ages of domain structures (MANET) [14] derived from phylogenomic analyses of protein structure [3], [5], [15] dissected the evolutionary appearance of novel structures and functions. It has been suggested that the difficulty of evolving novel stepwise chemical reaction mechanisms could be the dominant factor limiting the divergent evolution of new catalytic functions in related enzymes [16]. We put this concept to the test with phylogenomic analysis of protein domain structure and careful annotations of reaction mechanisms. Our observations have important implications for the origins of modern biochemistry and for exploring structure-function relationships.
Biocatalytic mechanisms operating in metabolic enzymes were traced along an evolutionary timeline of appearance of domain structures defined at the homologous superfamily (H) level of structural abstraction of CATH [11]. Hereafter, we refer to these fold superfamilies as H-level structures. CATH unifies domain structures hierarchically from bottom to top into sequence families (SF), homologous superfamilies (H), topologies (T), architectures (A) and classes (C). H-level structures are considered evolutionary units. The timeline was built directly from a phylogenomic tree describing the evolution of 2,221 H-level structures [5], treating their phylogeny as monophyletic. The tree was reconstructed from a census of domains in 492 fully sequenced genomes (42 archaea, 360 bacteria and 90 eukarya). The census produced a data matrix of multistate characters coded alphanumerically with columns representing proteomes (phylogenetic characters) and rows representing H-level structures (phylogenetic taxa), which was used to build rooted phylogenomic trees in PAUP* version 4.0b10 [17]. Trees were reconstructed using the maximum parsimony (MP) method with 1,000 replicates of random taxon addition, tree bisection reconnection (TBR) branch swapping, and maxtrees unrestricted. Character states in the data matrix were polarized from ‘N’ to ‘0’ using the ANCSTATES command of PAUP*, where ‘N’ indicates the plesiomorphic (ancestral) state. The model of phylogenetic character transformation that was used assumes that domain age is in general proportional to domain abundance in proteomes. The biological basis for global increases in domain abundance is the existence of processes of gene duplication, amplification and rearrangement in genomes [18] that drive molecular innovation. Details and support for character argumentation have been presented previously [3], [15]. Since genomic abundance should be considered a natural evolving ‘heritable’ trait, trees are expected to be unbalanced. Indeed, trees of domain structures are highly unbalanced and follow a molecular clock of folds that links molecular evolution with the geological record [4]. Consequently, the relative age of a domain fold structure (nd value) was calculated directly from trees using a PERL script that counts the number of nodes from the ancestral structure at the root of the tree to each leaf and provides it on a relative zero-to-one scale. Using the molecular clock converts this relative evolutionary timeline into a truly temporal geological timeline expressed in billions of years. An nd value of 0 indicates the origin of proteins approximately 3.8 billion years ago and the oldest domain, and a value of 1 the present and the youngest domain structure.
Our phylogenetic methodology relates to definitions of structures that are modern, based upon a structural census in the proteomes of extant organisms. Consequently, retrodictions are derived from modern structural complexity and do not necessarily depict the actual structure of hypothetical ancestors, which will always remain unknown (molecules can be brought back from the past experimentally by resurrection but cannot be confirmed to be truly bona fide retrodictive constructs). However, if molecules become structurally canalized in evolution, then modern retrodictive statements truly approximate molecular history.
For enzyme function definitions we have retrieved data from the MACiE database, specifically the functional annotations describing the chemical nature of individual reaction steps; frequently observed examples are “Proton transfer” and “Bimolecular nucleophilic substitution” (adundances and definitions in Figures 1 and 2, respectively). These MACiE annotations relate specifically to the steps of the mechanisms by which the reactions occur, rather than to the overall chemical transformation; the EC number covers the latter. To test the hypothesis of the ancestral folds using the most mechanistic step types, we retrieved 335 enzyme reactions from MACiE [19] version 3.0, mapped over fold age [5] using data from MANET [14]. MACiE is designed to be as complete as possible at the 1st, 2nd and 3rd levels of EC, but only representative at the 4th level. Its coverage, relative to the numbers of nodes for which PDB structures exist, is 6/6 (1st level); 54/57 (2nd level); 165/194 (3rd level); 249/1547 (4th level), according to figures collated in 2010 [19]–[21]. In this study, we are using detailed mechanistic stepwise information extracted from the primary literature by the curators of MACiE.
Out of 335 MACiE enzyme reaction entries, 321 entries had unique overall functions at the 4th level of the EC classification. MACiE entries included catalytic domains which adopted 236 different structures, as indicated by CATH H-level structures, and received age assignments. We emphasise that we are specifically considering domains annotated in MACiE as catalytic. In many enzymes, not all domains were actually involved in catalysis. For example, MACiE enzyme reaction M0124 (EC 1.9.3.1, cytochrome c oxidase) was annotated with 16 domains, of which only one domain (CATH 1.20.210.10, cytochrome c oxidase chain A) was annotated in MACiE as a catalytic domain used to effect the reaction. So we included only one of the 16 CATH domains in this analysis, CATH 1.20.210.10. The catalytic domain distribution of the remaining enzyme structures was as follows: 240 enzyme entries with a single catalytic domain, 63 enzymes having two different catalytic domains, four enzymes with three catalytic domains and only one enzyme entry in MACiE (M0207, EC 2.7.9.1, pyruvate-phosphate dikinase) with four domains (CATH 3.30.1490.20, nd = 0.0539; CATH 3.30.470.20, nd = 0.058; CATH 3.20.20.60, nd = 0.112; CATH 3.50.30.10, nd = 0.377) that participate in catalysis; pyruvate-phosphate dikinase is a key enzyme participating in gluconeogenesis and photosynthesis. Thus, a total of 308 MACiE enzymes were considered for further analysis. Only these H-level structures were used further to explore the evolution of biocatalytic mechanisms.
Once the data were filtered, we associated H-level structures with the mechanistic step types, MACiE's annotations of the reaction steps catalysed by the corresponding enzymes. In this study, we used 51 mechanism annotation definitions from the MACiE database, which can be associated with the steps defined for the enzyme-catalyzed reactions. The data matrix was a presence and absence (PA) matrix where each column represents the occurrence of a “mechanistic annotation” and each row represents a fold with its corresponding fold age. For example, M0017 purine-nucleoside phosphorylase (CATH 3.40.50.1580, nd = 0.235) has only one domain and uses four reaction steps to complete its reaction. In order to effect the reaction, this enzyme goes through: step 1, “Proton transfer”; step 2, “Heterolysis”; step 3, “Bimolecular nucleophilic addition”; and lastly step 4, “Proton transfer”. In this analysis, “Proton transfer” was counted once for this enzyme. The glossary of the mechanistic step types can be found on the MACiE website (http://www.ebi.ac.uk/thornton-srv/databases/MACiE/glossary.html).
In cases where the enzyme had only one catalytic domain, we associated the mechanistic annotations of each step with the structure of the domain. In cases where enzymes used more than one domain to effect the reaction, we carefully selected the domain or domains participating in each step and issued the mechanistic annotation to the corresponding H-level structures. We assigned the mechanistic annotation only if at least one residue from the domain was catalytically involved in the corresponding reaction step in MACiE, either as a “Reactant” or as a “Spectator” [22]. The complete data culling process was done using an R script [23] for retrieving data from the MACiE database that filtered and mapped the 308 MACiE enzymes onto their relative fold ages.
In order to test the hypothesis that the most ancestral protein domains use the greatest number of biocatalytic mechanistic step types, we assume that extant protein domain structure is the best historical archive that is available to explore ancient enzyme functions. The assumption holds good ground. At high levels of structural complexity, evolutionary change occurs at an extraordinarily slow pace. A new fold superfamily may take hundreds of thousands to millions of years to materialize in sequence space while new sequences develop on Earth in less than microseconds [24]. In fact, a recent comparative analysis of aligned structures and sequences showed that structures were 3–10 times more conserved than sequences [25]. Here we use the ages of domain structures, derived from phylogenomic reconstruction and a recent census of CATH domain structure in hundreds of genomes [5], to study how chemical mechanisms developed in protein evolution. The use of molecular structure and abundance in phylogenomic analysis offers numerous advantages over traditional methods [26], eliminating phylogenetic problems such as alignment, phylogenetic inapplicables and taxon sampling. Their use does not violate character independence, a serious problem that has not been addressed in phylogenetic sequence analysis. To our knowledge, this is the first study to explore the evolution of biocatlytic mechanisms using a timeline of CATH homologous superfamily (H-level) domain structures and data analysis. However, there is another comprehensive database, FunTree [27], that brings together sequence, structure from CATH, chemical and mechanistic information from MACiE, and phylogenetics.
In order to explore the use and reuse of biocatalytic mechanisms in evolution, we mapped the mechanistic definitions of enzymatic functions to their respective CATH H-level structures, with structures ordered according to fold age (Figures 1, 3, 4). For this purpose we first created a presence and absence (PA) matrix, a heat map representing the distribution of the presence (red) and absence (yellow) of the mechanistic step types (rows, y-axis) in the fold (columns, x-axis) (Figure 1). The rows were ordered vertically according to the first appearance of the mechanism over fold age and were indexed with the numbers of: (i) MACiE enzyme entries (shades of grey and black), (ii) H-level structures (shades of grey and purple), and (iii) EC classes that appeared at each age. The complete data set is provided as Supporting Information, Dataset S1.
Remarkably, the most popular enzyme mechanistic step types were associated with the oldest H-level structures (Figure 1). This evolutionary trend suggests that the oldest enzymes already provided a sufficiently flexible scaffold to support many diverse mechanistic step types in order to complete their reactions. Within the early scaffolds, the mechanistic steps had more time to be adapted by the domain structures and to be further recruited in the course of evolution. The existence of late emerging structures with many mechanistic steps supports the presence of widespread recruitment processes in evolution. This trend seems to be explained in terms of the “preferential attachment principle” that guides the growth of scale-free network behavior, and implies that the more prevalent functions are typically the earliest, as previously shown in the exploratory analysis of the ancestral fold structures [28].
We observed that “Proton transfer”, “Bimolecular nucleophilic addition”, “Bimolecular nucleophilic substitution”, and “Unimolecular elimination by (or from) the conjugate base” (definitions are represented in Figure 2) are the most common mechanistic step types, in accordance with their distribution in MACiE enzyme reaction mechanisms (the prevalence of each step type is also given in Supporting Information, Table S1) [12], [29]. These types of mechanistic steps are recognisably fundamental building blocks of enzyme chemistry, which is carried out in aqueous solution usually at approximately neutral pH. Several of the canonical amino acids have pKa values close to neutral, with Holliday et al. having observed particularly strong propensities for His and Glu to facilitate proton transfer [12]. The chemistry of the amino acid side chains also means that several are negatively charged at roughly neutral pH, and hence it is no surprise that the enzyme far more often acts as a nucleophile, favoring mechanisms labelled as nucleophilic, rather than as an electrophile. Furthermore, it has been noted that enzyme active sites are well suited to stabilising the charged intermediates common in addition and elimination reactions, for instance by hydrogen bonding [22]. The ubiquity of aqueous environments in enzyme chemistry restricts the repertoire of reactions available. Indeed, most enzyme reactions are composed of steps that might seem unexciting to an organic chemist. The rare occurrence of more complicated organic chemistry, “Aldol addition”, “Amadori rearrangement”, “Claisen condensation”, “Claisen rearrangement”, “Pericyclic reaction” and “Sigmatropic rearrangement”, constitutes the exception rather than the rule, and enzymes sample the space of possible mechanisms notably differently from how an organic chemistry textbook would do so.
The rate of introducing new mechanistic step types at different fold ages is shown in Figure 3, which represents a cumulative plot where fold age is shown on the x-axis. The y-axis shows the proportion of the total number of defined step type annotations (N = 51) that have been uncovered up to that fold age on the x-axis. It is clear in this plot that the first four H-level structures (the first two increments of fold age, 0 to 0.0098 ) are responsible for introducing a third of the known mechanistic step types (18/51), and the first six structures (the first four increments of fold age, 0 to 0.049) are responsible for over half of them (27/51). However, the development of the other half was harder and required the unfolding of about ¾ of the evolutionary timeline, up to nd = 0.73, and about 2.5 billion years of evolution (inferred using a molecular clock of folds [4]). The detailed information regarding the introduction of mechanistic step types is provided in Table 1.
In order to look at the distribution of the mechanistic step types of an enzyme in evolutionary time, we counted the number of mechanistic step types associated with H-level structures (Figure 4). Figure 4 is a heat map representing the number of mechanism step types (y-axis) used by those structures having each different discrete value of fold age (x-axis). Each cell represents the number of H-level structures with a different color code; for example black represents 1 structure, yellow represents 2 structures and brown represents 3 structures sharing the same count of mechanistic step types. Moreover, each position indicates the number of H-level structures associated with a number of functions. For instance, black color at column 1 row 6 means that there is one structure that uses 6 different mechanistic step types to complete its reaction. The x-axis scale reflects the different nd values found in our dataset, arranged from the oldest on the left to the youngest on the right. Every unique nd value forms a separate column. The non-linear scale is defined by the number of unique nd values falling in each interval of nd. In a further section, we will discuss the patterns in detail.
The most ancient H-level structure that appears in the MACiE database is CATH 3.40.50.300, the P-loop containing nucleotide triphosphate hydrolase. This fold has been consistently identified as the most ancestral fold structure [2], [3], [5]. The P-loop hydrolase structure consists of the most ancient and abundant topology, the Rossmann fold (CATH 3.40.50), which has the 3-layer (αβα) sandwich (3.40) architecture. The CATH 3.40.50.300 superfamily contains enzymes with diverse molecular functions, including signal transduction, hydrolase and transferase enzymatic activities [30]. Wang et al. previously observed [15] diverse overall functions for this structure (the complete list of MACiE enzyme entries is given in Supporting Dataset S1). In the current analysis, there are only five MACiE enzyme entries that share this structure; these are associated with six mechanistic step types, “Proton transfer”, “Electron transfer”, “Bimolecular nucleophilic addition”, “Bimolecular nucleophilic substitution”, “Intramolecular nucleophilic addition” and “Unimolecular elimination by the conjugate base” (Table 1). MACiE enzymes associated with this oldest structure are dethiobiotin synthase (EC 6.3.3.3, M0074), estrone sulfotransferase (EC 2.8.2.4, M0154), H+-transporting two-sector ATPase (EC 3.6.3.14, M0178), nitrogenase (EC 1.18.6.1, M0212, multi-domain) and adenylate kinase (EC 2.7.4.3, M0290). Except for nitrogenase, the rest of these enzyme entries each have a single catalytic domain, hence, it is straightforward to annotate the function with this fold. Nitrogenase (M0212, PDB: 1n2c) [31] is a very important enzyme of nitrogen metabolism that fixes atmospheric nitrogen (N2) gas into the reduced forms that are usually assimilated by plants [32]. The enzyme has a complex 3D structure that is highly conserved across many different organisms and contains domains from three different homologous superfamilies. These H-level structures first evolved at different times. The ancient CATH 3.40.50.300 nitrogenase catalytic core was later accesorized with a domain from the CATH 3.40.50.1980 superfamily, which evolved at nd = 0.401 after the oxygenation of Earth's atmosphere [4], [33], [34], and a non-catalytic domain CATH 1.20.89.10, which appears to have been accreted last into the molecule (nd = 0.549). Residues from the ancient nitrogenase core with the oldest domain of the molecule are involved in the first two steps of the long 15-step reaction, which include the mechanistic step types “Bimolecular nucleophilic substitution”, “Electron transfer” and “Proton transfer”. The remaining 13 steps are carried out by catalytic residues from the CATH 3.40.50.1980 domain.
The three H-level structures at the second most ancient fold age include CATH 3.50.50.60, the T-level topology of which is 3-layer ββα; its H-level structure has no specific name assigned, but corresponds to the FAD/NAD(P)-binding domain FunFams definition in CATH and is found in 7 MACiE entries. Having the same fold age, we find CATH 3.40.50.720 (NAD(P)-binding Rossmann-like domain) in 12 MACiE enzymes, and CATH 3.40.50.150 (Vaccinia Virus protein VP39) in two MACiE entries. All three H-level structures appear at nd = 0.0098. These structures have 16, 15, and 4 catalytic mechanistic step types (Figure 4), respectively, of which a total of 11 are non-overlapping with those of the first P-loop hydrolase fold structure and were therefore newly introduced at this time (see Table 1). These newly evolved mechanistic step types include three involving aromatic groups, as well as the first involving radicals, and also “Bimolecular electrophilic addition”, “Bimolecular elimination”, “Redox”, “Colligation” and “Assisted keto-enol tautomerisation”. It was interesting to note that the “Bimolecular elimination” mechanism was shared by all three H-level structures of the same age. There are 9 different mechanisms shared by CATH 3.40.50.720 and CATH 3.50.50.60 (shown in Table 1). Studies by the Orengo group [35], [36] suggest there may be distant homology between these structures, based on their similarity in graph-based structure comparison and shared use of organic cofactors (NAD and FAD). The structures are functionally diverse due to the conformational change of the ligands, organic cofactors or structural plasticity of the proteins [37].
In MACiE, the ferredoxin-NADP+ reductase enzyme (M0142, EC: 1.18.1.2) combines the CATH 3.40.50.150 and CATH 3.50.50.720 H-level structures to complete its biochemical reaction. This enzyme plays a very important role in electron transfer from the flavoenzyme NADPH-adrenodoxin-reductase (AdR) to two P450 cytochromes; this process is involved in the production of steroid hormones. The two domains of this enzyme share the following functions: “Aromatic unimolecular elimination by the conjugate base”, “Aromatic bimolecular nucleophilic addition”, “Redox”, “Radical termination”, and “Radical formation”.
The next most ancient H-level structure (nd = 0.0147), CATH 3.40.50.620, the H-level Hups α/β layered fold, is responsible for 13 MACiE entries and introduces the novel “Intramolecular elimination” function. This structure supports central catalytic functions of the cell, including the aminoacylation reactions of aminoacyl-tRNA synthetase (aaRSs) catalytic domains that are crucially involved in the attachment of L-amino acids to cognate tRNA molecules and are responsible for the specificity of the genetic code. The structure includes the tyrosyl-tRNA ligase EC function (M0197; EC 6.1.1.1) of the tyrosyl-RS functional family, the oldest aaRSs delimiting the process of translation [38]. The enzyme activates a specific amino acid by condensation with ATP to form an aminoacyladenylate intermediate, which then esterifies the 2′ or 3′-hydroxyl group of the ribose at the 3′ end of the acceptor arm of tRNA. The aminoacylation site rejects larger amino acids and a proofreading site in an editing domain hydrolyzes small amino acids that were incorrectly activated through pre-transfer or post-transfer editing mechanisms.
Some H-level structures by nature use many diverse mechanistic step types to effect their catalytic activity. A member of the TIM barrel α/β structure that is highly popular in metabolism, the CATH 3.20.20.70 superfamily (aldolase class I, nd = 0.0196), which immediately follows the aaRS fold in the timeline, supports a diversity of chemistry that includes 20 different mechanistic step types. Five of these appeared for the first time with this fold (Table 1). It is not surprising that the fold has such diverse functions. Based on the Hierarchic Classification of Enzyme Catalytic Mechanisms (RLCP; where R: Basic Reaction, L: Ligand group involved in catalysis,C: Catalysis type and R: Residues/cofactors located on Proteins) classification [39] analysis of functional subclasses [40], Nagao et al. suggested that aldolase class I enzymes have various functional classifications. An interesting conserved property is that most of their ligands have at least one phosphate group. The mechanistic step types of aldolase class I (see Table 1) are rare in the MACiE database. Out of 335 MACiE enzyme entries, “Aldol addition”, “Aromatic bimolecular elimination”, “Assisted other tautomerisation”, “Heterolysis” and “Other tautomerisation”, respectively, appeared in 9, 6, 20, 25 and 9 MACiE enzyme entries in at least one stage of the reaction (the numbers of different MACiE entries containing each of the mechanistic step types are given in Table S1). This suggests that the aldolase class I superfamily contains a group of enzymes that possess very specific mechanistic step types.
Two additional H-level structures utilise 16 different mechanistic step types each, CATH 3.50.50.60 (nd = 0.0098) (which we have already mentioned) and CATH 3.40.50.970 (nd = 0.049), the second largest number of mechanistic step types associated with any structures in the timeline. These structures also belong to the most popular fold topology, the Rossmann fold. Following their appearance (nd = 0.049), most of the fundamental and common mechanistic step types had already been introduced. The CATH 3.40.50.970 structure introduces “Homolysis”, represented in only one MACiE entry (M0119; EC: 1.2.7.1; pyruvate: ferredoxin oxidoreducatse). We observed that two mechanistic step types, “Homolysis” and “Colligation”, were introduced at the same fold age but by different H-level structures. By definition, the “Homolysis” mechanistic annotation is the converse of the “Colligation” step that was introduced by CATH 3.50.50.60; “Homolysis” is the cleavage of a covalent bond where each atom retains one of the two bonding electrons, whereas “Colligation” is when two free radicals combine to form a covalent bond.
We were also interested to see what sets of mechanistic step types described the combinations of steps used by various enzymes to effect their reactions. To do so, we looked for the combination of the different mechanistic step types, irrespective of order, and at the various H-level structures sharing each combination of biochemical steps. Instances of reutilisation of particular mechanistic step types may shed light on evolutionary recruitment of common mechanistic steps by different structures. For this we first created “mechanistic annotation patterns”. These patterns reflect all the different combinations of the presence and absence of mechanistic step types. This kind of analysis illustrates that different H-level structures share common mechanistic annotation patterns. We found that there are 133 different mechanistic annotation patterns used by the enzymes in our dataset (the complete mechanistic annotation patterns are provided in the Supporting Information, Table S2 and Table S3). Pattern 4 is most popular mechanism combination, involving “Bimolecular nucleophilic substitution” and “Proton transfer” (see Figure 5, H-level structures are grouped together in the white box). There are 42 H-level structures in MACiE that use two mechanistic step types in order to complete their reactions. Out of these 42 structures, 30 use pattern 4 in order to complete their reactions. Patterns 4 and 15 suggest that there are few H-level structures (details of superfamilies and pattern association are represented in Table S3) that accommodate similar mechanistic step type combinations.
Pattern 15 is the second most popular pattern and includes “Bimolecular nucleophilic addition”, “Proton transfer” and “Unimolecular elimination by the conjugate base”. In MACiE, there are 46 different catalytic H-level structures that use three mechanistic step types in order to complete their reactions, out of which 22 structures use pattern 15 to effect their reactions. The enzymes of the CATH 3.20.20.70 (aldolase class I) structure use the maximum number of 20 different mechanistic step types to effect their overall reactions. These step types constitute pattern 133 (see Table 2), which is not shared by any other structure. These patterns suggest which mechanistic step types are compatible with one another or are preferentially combined together. There are 101 patterns unique to one structure (see Table S3).
To visualise the combinatorial patterns, we have plotted a heat map of similarity of the mechanistic step types between two H-level structures (Figure 5). We calculated the Jaccard similarity scores; where A and B are two sets and the Jaccard coefficient of similarity is defined as the size of the intersection divided by the size of the union between the two sets. To visualize computed similarity scores, we constructed a presence and absence (PA) matrix where columns represent the mechanistic annotation as an entity and rows represent the CATH H-level structures. The score ranged from 0 to 1, with 0 signifying that no similar mechanistic step types existed between two structures and 1 signifying that the two structures shared an identical combination of mechanistic step types in order to complete their reactions. The most popular mechanism combinations, pattern 4 (“Bimolecular nucleophilic substitution” and “Proton transfer”) and pattern 15 (“Bimolecular nucleophilic addition”, “Proton transfer” and “Unimolecular elimination by the conjugate base”), are labelled in the heat map of Figure 5 and are clearly distinguishable. As expected, these patterns include the most common and ancient mechanistic step types introduced with the CATH 3.40.50.300 structure.
The research goals of this paper are not to explore mappings of mechanistic step types along metabolic pathways, as this would require one to unfold a complex network structure with graph theoretical approaches. However, in order to make explicit the complex recruitment patterns that are expected we have mapped H-level structures in the nucleotide interconversion pathway of purine metabolism [41], the oldest of all metabolic subnetworks defined by the KEGG database [42]. Since nucleotide interconversion precedes purine biosynthesis in evolution [41], we compared mechanistic step types associated with this pathway (Table 3). In MACiE, we found only 8 H-level structures involved in purine metabolism, ranging in nd value from 0 to 0.411. Remarkably, and despite the absence of MACiE entries for the most ancient enzymes of energy interconversion (EC 2.6.1.3. and EC 3.6.4.1), the results reveal the very early rise of the highly abundant pattern 4 in evolution and complex patterns of recruitment of additional chemistries (Figure S1) which are ultimately associated with the combinatorics of mechanistic step types of Figure 5.
Contemporary protein structures consist of independently folding and compact domains that can be used as a fossil record of molecular evolution. We have utilised the available resources of enzyme mechanisms and the relative ages of CATH H-level domain structures to get a better insight into the natural history of biocatalytic mechanisms. Our analysis shows that the most designable structures (e.g., the α/β barrel and Rossmann fold) served as scaffolds to higher numbers of biochemical functions. The first two structures were responsible for introducing 35% (18/51) of the known mechanistic step types. Over half of these appeared in the evolutionary timeline of domains before structures specific to Archaea, Bacteria and/or Eukarya [5], during a period of architectural diversification (nd<0.39). The most common mechanistic step types were also the most ancient and included fundamental building blocks of enzyme chemistry, “Proton transfer”, “Bimolecular nucleophilic addition”, “Bimolecular nucleophilic substitution”, and “Unimolecular elimination by the conjugate base”. Later on in evolution, these mechanistic steps participated in a combinatorial interplay and were the highest represented in catalytic functions. The combination of “Bimolecular nucleophilic substitution” and “Proton transfer” was the most popular of all patterns of mechanistic step types. The other half of mechanistic step types appeared gradually after organismal diversification (0.67<nd<1) and during a period that spanned ∼2 billion years of evolutionary history.
Our phylogenomic approach is based on a census of protein domain structure in the proteomes of cellular organisms and the crucial axiom of polarization that claims that structural abundance increases in the course of evolution. This ‘process’ model of molecular accumulation in proteomes is based on Weston's generality criterion of homology and additive phylogenetic change [43] that in our case describes the slow and nested accumulation of homologous domain structures in the branches (proteome lineages) of the tree of life. A careful phylogenetic reconstruction analysis reveals that while both gains and losses of domain structures are frequent events, gains always overshadow losses in evolution [44]. This supports the general proportionality of domain abundance and evolutionary time of phylogenetic argumentation and the principle of continuity, the most important pillar of Darwinian evolution.
In these studies we trust the CATH classification scheme of domain structure, assignments of known structures to sequences, and current understanding of metabolic networks and associated chemical reactions. We note that it is highly likely that there is an ‘underground’ metabolism of weak catalytic specificities that is not annotated and involves a multiplicity of substrates and perhaps mechanistic step types. Our analysis is unable to capture this aspect of enzymatic function at this time. Similarly, our analysis does not explore biases in the distribution of annotations of molecular functions among structures and structures among functions nor the distribution of mechanisms across enzymatic reactions. Instead, it reveals patterns of accumulation of mechanistic step types in evolution.
The historical patterns we reveal uncover an explosive diversity of catalytic mechanisms embedded in the explosive discovery of EC functions [6], which are used in the different chemical reactions of the emergent metabolic networks. The evolutionary driver of mechanistic innovation of protein reaction chemistries was probably recruitment of strategies used in primordial metabolic chemistries that already existed on early Earth and their internalization into the emerging polypeptide scaffold. Support for this contention comes from a careful mapping of structures, functions and prebiotic chemical reactions in purine metabolism, the most ancestral metabolic subnetwork of metabolism [6]. This mapping revealed a gradual replacement of abiotic chemistries and the existence of concerted enzymatic recruitments driving the early evolution of pathways of nucleotide interconversion and the late appearance of pathways of biosynthesis, catabolism and salvage [41].
|
10.1371/journal.ppat.1007508 | The Vaccinia virion: Filling the gap between atomic and ultrastructure | We have investigated the molecular-level structure of the Vaccinia virion in situ by protein-protein chemical crosslinking, identifying 4609 unique-mass crosslink ions at an effective FDR of 0.33%, covering 2534 unique pairs of crosslinked protein positions, 625 of which were inter-protein. The data were statistically non-random and rational in the context of known structures, and showed biological rationality. Crosslink density strongly tracked the individual proteolytic maturation products of p4a and p4b, the two major virion structural proteins, and supported the prediction of transmembrane domains within membrane proteins. A clear sub-network of four virion structural proteins provided structural insights into the virion core wall, and proteins VP8 and A12 formed a strongly-detected crosslinked pair with an apparent structural role. A strongly-detected sub-network of membrane proteins A17, H3, A27 and A26 represented an apparent interface of the early-forming virion envelope with structures added later during virion morphogenesis. Protein H3 seemed to be the central hub not only for this sub-network but also for an ‘attachment protein’ sub-network comprising membrane proteins H3, ATI, CAHH(D8), A26, A27 and G9. Crosslinking data lent support to a number of known interactions and interactions within known complexes. Evidence is provided for the membrane targeting of genome telomeres. In covering several orders of magnitude in protein abundance, this study may have come close to the bottom of the protein-protein crosslinkome of an intact organism, namely a complex animal virus.
| Vaccinia is one of the most complex virions among the animal viruses, containing 70+ distinct gene products. Although virion ultrastructure has been apparent, at least in outline by electron microscopy since the year 1961 or earlier, its molecular architecture is largely unknown: Vaccinia is resistant to classical structural approaches requiring virus crystallization and moderately resistant to cryoEM. Molecular approaches requiring the maintenance of protein assemblies during virion deconstruction, reconstruction of protein complexes in heterologous or in vitro systems, or internalization of bulky reagents such as antibodies or gold particles may have been already pursued close to exhaustion. Here, protein interfaces within and around the intact virion were identified by virus incubation with bifunctional chemical crosslinkers in situ followed by proteolysis and peptide-level mass spectrometry. This minimally invasive approach revealed the molecular arrangements of structural and membrane protein complexes within the virus, confirming and extending several aspects of virus biology.
| The virion of Vaccinia, the prototypical poxvirus, is one of the largest among the animal viruses. While its ultrastructural characterization is the beneficiary of 60+ years of electron microscopic examination [1–3] and references therein, attempts to better understand its molecular and atomic architecture have fallen foul of various properties of the Vaccinia virion such as asymmetry, polymorphic character, tendency to aggregate, and the general incompatibility of enveloped viruses with X-ray crystallography.
Electron microscopy (EM) and atomic force microscopy (AFM) studies have established clear ultrastructural compartments of the mature virion (MV) [4] including a central, genome-containing ‘core’ that also houses a number of virus-encoded enzymes of mRNA transcription and modification, a proteinaceous wall surrounding the core, a pair of ‘lateral body’ structures flanking the core wall, a single lipid bilayer envelope, and an outer protein-rich coat that appears late during maturation. The virion contains between 58 and 73 distinct gene products [5]. Some of these have been localized at low resolution on the basis of immunogold EM [6–10], while the compartmental locale of others can be inferred from clearly identifiable transmembrane (TM) domains and other bioinformatics signatures, known function and/or the conditions required for the extraction from the virion. Proteins and visible structures localizing to outer compartments of the virion (outside of the core) have been identified via their fractionation in vivo during virus entry [10, 11] or under pseudo-entry conditions recreated by the gentle, controlled treatment of virions with nonionic detergent or nonionic detergent+disulfide reductant [9, 12–15]. A number of core enzymes, including the virus-encoded multisubunit DNA-dependent RNA polymerase (RPO), heterodimeric virion capping enzyme (CA), early transcription factor (ETF), poly(A) polymerase (PAP), two protein kinases, at least two proteases and two glutaredoxins have been released from the virion under more harsh conditions (0.2% ionic detergent (sarkosyl) and high salt [16]), retaining solubility, integrity and activity after detergent removal [4]. By contrast, a number of structural proteins of the virion core remain insoluble during virion extraction even in ionic detergent.
Aside from these compartmentalization approaches, little is known of the virion’s internal organization at the molecular level. Certainly, the heteromultimeric status of the above core enzymes has long been known [4], and the homomultimeric status of yet other virion proteins has been revealed by X-ray crystallography (eg. proteins H1 [17, 18] and A27 [19]). Some binary protein-protein interactions have been successfully recapitulated and identified in a yeast two-hybrid system [20]. Other proteins, and fragments thereof, have been co-immunoprecipitated from cell extracts, pulled-out as tagged complexes [2] or inferred by genetic and directed mutational studies.
However, larger macromolecular and ultrastructural assemblies clearly dissociate under the conditions required for full virion disruption. For example, the presence, within the virion core, of a ‘transcriptosome’ assembly was inferred in studies down-regulating the Vaccinia RNA polymerase subunit RAP94. Under non-permissive conditions, virions were morphologically mature but showed low infectivity [21]. Albeit the virus genome was packaged in normal amounts as were ETF and the structural proteins, low or undetectable amounts of RPO, CA, PAP large subunit, and proteins NTP1, RNA helicase and topoisomerase were packaged suggesting the coordinated packaging of the latter components. Such a ‘transcriptosome’ complex may correspond to the formation, within the core, of a genome-containing tubular ultrastructure [22] that can be resolved by EM under sample preparation conditions that include high pressure freezing [23]. However, no such ultrastructure or any subassembly thereof has been isolated biochemically: Capping enzyme can form a binary complex with RPO in vitro [24], but the soluble fraction from a sarkosyl virion core lysate, for example, even under gentle gradient sedimentation conditions, has yielded no higher order assemblies beyond the sedimentation of RPO as a discrete entity and the partial co-sedimentation of RPO with viral capping enzyme and NTP1 [25]. Other enzymes, including those apparently co-packaged with RAP94 (above) sedimented separately, towards the top of the gradient, suggesting an irreversible disruption of interactions within the transcriptosome upon core rupture. To our knowledge, no comprehensive transcriptosome, or other packaged superstructure has been (re)assembled biochemically as a positive correlate to the subtractive approaches of genetics.
Here, we have taken an approach to the molecular structure of the Vaccinia virion that is neither destructive, reconstructive nor exclusively applicable to binary complexes, namely protein-protein crosslinking mass spectrometry (XL-MS). We address the virion in its natural state in situ, with the potential to interrogate multivalent protein complexes. Technical challenges in this approach were not inconsiderable: At the outset of the current study, higher profile XL-MS studies in the literature had focused upon stoichiometric or near-stoichiometric isolated protein complexes, containing around ten or fewer polypeptides, with known crystal structures. Examples of these would include the 26S proteasome [26], multi-ringed TRiC/CCT chaperonin [27, 28], the RNA polymerase II pre-initiation complex [29–31], RNA polymerase I [32] and RNA polymerase III [33]. By contrast, the Vaccinia virion likely contains a variety of protein complexes covering an abundance dynamic range of ~5000 [34] or greater, only a minority of which have yielded X-ray crystallographic structures. Our XL-MS results with Vaccinia are described below.
Virions (intact or activated for mRNA transcription) were incubated with bifunctional chemical crosslinkers to impose inter-protein distance restraints. Crosslinked virus was then dissolved and trypsinized to peptides, followed by peptide-level nanoLC-MS/MS and bioinformatics to identify crosslinked peptides. For disuccinimidyl suberate (DSS), the crosslinker used in the majority of experiments, the restraint comprised a lysine Nζ-Nζ distance of ≤ 10–11.4 Å with corresponding Cα-Cα distances of ≤ 32 Å (give or take molecular dynamics considerations). Crosslinkable lysines thereby sweep a sphere of Cα-Cα distances up to ~6 nm, or ~2% of the diameter of a Vaccinia virion for proteins not forming extended, repeating arrays.
Due to the low intrinsic ionizability of crosslinked peptide pairs and the potential for low saturation crosslinking within/between low abundance proteins in the virion, a strategy of variation [5] (Table 1) was implemented to maximize opportunities for the detection of crosslink (XL) ions (Fig 1). This was combined with a total of six distinct XL search engines, used in parallel (Fig 1 and Materials & methods). After data thresholding and filtering, a unique meta-score (‘DFscore’, or detection frequency score) was introduced as a guide to the extent of internal confirmation within the dataset.
The resulting XL dataset yielded a total of 4609 confidently-identified unique-mass ions, each corresponding to a crosslinked peptide pair (S1A and S1B Table). Of these, 1486 (32.2%) had a DFscore > 1. The highest DFscore for any ion was 178, and the four top-scoring ions each corresponded to p4a intra-protein XL, of which the two highest scoring were light/heavy versions of the same ion and the third represented a small shift in XL position for one of the two crosslinked peptides (S1A and S1B Table)). 3725 of the 4609 unique-mass ions represented intra-protein XL while 884 were inter-protein, consistent with the known tendency for XL to fall within rather than between proteins. 273 of the 884 inter-protein XL ions had a DFscore > 1 among which the highest DFscore was 83 (p4a-position 876 crosslinked to p4b-position 563).
By merging (a) distinct charge states for a crosslinked peptide, (b) identical crosslinked accessions/positions detected within distinct peptide species, (c) light/heavy isotopic forms of the crosslinker and (d) crosslinked peptides with secondary modifications, the 4609 unique XL ion masses collapsed down to 2534 unique pairs of residues within the proteome. 625 of these were inter-protein and, of these, 157 (25.1%) had a DFscore > 1 with the highest DFscore for an inter-protein accession/position pair being 475 (for the p4a-876/p4b-563 XL mentioned above). This accession/position pair was represented by 43 distinct m/z crosslinked peptide ions. S1C Table shows all crosslinked protein pairs in the dataset. S1 Fig shows crosslinking partners among all proteins considered to be packaged in the virion [5] for which XL were detected, and Table A in S1 Text reconciles the proteome of S1 Fig with the contents of the XL search database.
Orthogonal approaches to the validation of in situ–detected protein-protein interactions all seemed less direct than XL-MS itself (involving virion disruption, recapitulation of interactions in vitro, and/or the expression of virus proteins in heterologous systems). We therefore sought to validate the XL dataset via inference criteria, asking four basic questions as follows:
All six XL search engines employed a target-decoy approach [35] (Table 2) and primary score thresholding comprised false discovery rate (FDR) or its surrogate, q-value (Materials & methods). For four of the six engines we took the unprecedented step of also applying a second threshold, via the score-type that is native to the engine itself (Table 2). A small fraction of the ions discarded solely on the basis of threshold 2 were then rescued according to the criteria described in Materials & Methods. With a primary threshold alone, namely 5% FDR, around 230 of our 4609 unique-mass ions would have arisen from our decoy database. Via our dual thresholding/rescue approach (see the “Data Assembly” section of “Materials & methods”), only 15 of the 4609 ions involved a decoy accession, representing an effective FDR of just 0.33%—an exceptionally low number. We regard our low effective FDR as a bona fide validation step, and an indication of low technical noise in the dataset. All 15 decoy hits had a DFscore of 1 with one exception, whose DFscore was 2.
Non-randomness was evaluated on the basis of several criteria:
Inter-protein vs. intra-protein XL: For a database of 86 proteins, random partner selection would result in a 1/86 (1.12%) chance of both tryptic peptides in a crosslinked pair arising from the same protein, assuming an equal number of tryptic peptides from each protein in the database. Experimentally, however, far more opportunities exist for efficient crosslinking within a protein than between proteins. Of the 1742 unique accession/position pairs in the dataset, 1294 (74.3%) were intra-protein, conforming to the experimental expectation rather than the random selection of peptides during bioinformatics.
Protein abundance: During MS data acquisition, ions were prioritized for sequencing on the basis of intensity (high-to-low) leading to an expectation of XL detection at a higher frequency for relatively abundant proteins. Consistent with this, the dataset was dominated by XL between the abundant virion structural proteins p4a and p4b (S1C Table). This provided a clear validation of data on the basis of known protein abundance.
Non-random lysine occupancy per protein: If search engines were picking lysine XL sites randomly, then the proportion of lysines occupied with XL would be expected to be fairly constant from protein-to-protein. However, lysine occupancy on a per protein basis covered a broad range, from 32.5% to 100% (Fig 2a). Search engines were therefore not simply picking sites from the database randomly. Some proteins were clearly more ‘detectably crosslinkable’ than others for reasons that presumably included protein abundance, solvent accessibility and lysine basicity for reaction with succinimide-based crosslinkers.
Non-random ‘hotspotting’ of lysine XL sites within a protein: Individual XL sites within a protein may vary in exposure, reactivity or flexibility or the number of reactive partners within crosslinking range, resulting in the appearance of crosslinking ‘hotspots’ [36]. The crosslinkability of some protein N-termini in particular (S1 Fig) likely arises from their exposure and flexibility, combined with a pKa [37] that promotes chemical reactivity. Consistent with this, individual lysines in our dataset showed substantial variation in predisposition towards XL ‘hotspotting’ (Fig 2b). F17 residue K74, for example, provided a particularly concentrated crosslinking hotspot, appearing in a total of 45 distinct accession/position pairs (Fig 2b) among 15 protein partners (S1 Fig). By contrast, many other positions in various accessions appeared just once (Fig 2b, S1 Fig).
Non-random coverage of inter-protein XL space: Our 86-protein search database provided a theoretical space of 3655 potential protein-protein pairs from which the XL dataset contained just 449. Despite the depth of analysis (4609 XL ions), this 12.3% coverage of theoretical inter-protein crosslinking space suggested a level of specificity.
At the time of writing, partial or complete X-ray crystallographic structures covered the crosslinked portions of 12 proteins in our XL dataset, with an additional two crystallographic structures from other orthopoxviruses (Table B in S1 Text). All possible lysine-lysine through-space (Euclidian) and solvent-accessible surface (SAS) distances within all of these structures [38] were binned, and the resulting two histograms were found to be centered at ~43 and ~54 Å, respectively (Fig 3). By contrast, the Euclidian/SAS distance histograms for all experimental XL found within the 14 proteins was centered at 14.9 and 13.5 Å respectively, with 103 or 114 (SAS/Euclidian) out of the 136 experimental XL distances being structurally rational (≤ 32 Å, Cα to Cα distance). Based on the Kolmogorov-Smirnov test, the probability that the “All lys-lys” and “experimental XL” distance histograms (Fig 3) were sampled from a single population was < 10−4, providing 99.99% statistical confidence that the crosslinking dataset was structurally rational.
Further assessment of the XL dataset was largely biological, namely, whether the identities of crosslinked protein pairs were consistent with known protein functions. For this analysis, accessions with strong functional annotations were collected into groups (Table 3). Interactions within any group were considered ‘biologically rational’, while the pairing of a membrane-group protein with a transcriptosome-group protein was designated ‘biologically non-rational’ since these two groups of proteins are considered, based on controlled degradation studies [9, 16], the most likely among the various groups to occupy distinct virion compartments—separated by the core wall. All other protein-protein pairings were disregarded for the purposes of biological validation as being relatively uninterpretable. Membrane-group proteins showed a moderate, yet unmistakable global positive predilection for other membrane-group proteins as crosslinking partners, and a mild antipathy, globally, for transcriptosome proteins (Fig 4a). Transcriptosome proteins, as a class, showed a mild but unmistakable predilection for other transcriptosome proteins as crosslinking partners and a mild antipathy for the membrane class (Fig 4b). While not absolute, the trends shown in Fig 4 were consistent with accepted compartmentalization models for virion proteins, with the likely location of the transcriptosome within the virion core enclosed by a core wall, and virion TM proteins likely occupying a two-dimensional membrane compartment surrounding the core wall. This provided a suggestion of biological rationality within the XL dataset. Among the top 28 crosslinked protein pairs by DFscore, 12 were ‘rational’ and only 2 were ‘non rational’ (S1C Table). The top 28 protein pairs contained 1205 of the 1849 total XL ions represented in S1C Table, and the top 12 “Y” protein pairs represent 92% of all XL ions in S1C Table associated with a “Y” (ie. that were biologically ‘rational’).
We have investigated the molecular structure of the Vaccinia virion, a highly non-stoichiometric protein assembly, via XL-MS. Analysis of protein-protein interactions in the virion in situ avoided the need for their preservation during virion extraction with reagents such as deoxycholate, an ionic detergent used for the release of virion core enzymes [16]. There was no requirement to rebuild virus protein complexes de novo, avoiding a need for the correct folding of challenging or insoluble structural proteins in vitro and/or in a heterologous system. Finally, multivalent/higher order complexes could be addressed that were not accessible via binary assays such as Y2H [20].
As in any XL-MS study, challenges included: The availability and appropriate spacing of crosslinkable sites at protein interfaces; good occupancy of crosslinking sites and robust reaction of both ends of the crosslinker; efficient laboratory digestion of crosslinked proteins (given the tendency of trypsin recognition sites, for example, to become derivatized); the detection of crosslinked peptide pairs against a large excess of non-crosslinked peptides in the same digest; rarity of inter-protein XL (the most informative kind) with respect to other kinds (intra-protein, intra-peptide, and single-ended XL); the tendency of large (more than double-size) crosslinked peptide pairs to ionize less efficiently during MS; inefficient fragmentation and combinatorial complexity of fragment ion mass spectra when simultaneously fragmenting peptide pairs, and the challenge of distinguishing true intra-molecular XL from those that may cross homomultimer interfaces. For Vaccinia as a target, the above issues were compounded by: Unknown permeability of the virion core to crosslinker; a protein abundance dynamic range in Vaccinia MV of 5000-fold [34] or more; a paucity of existing high resolution protein structures for validation, and the possibility of molecular heterogeneity arising from mixed viroforms in MV preparations and/or mixed proteoforms within a single particle.
Addressing the above challenges (most particularly the abundance range and sensitivity issues) we adopted a “strategy of experimental variation”, as explored initially in our analysis of the MV phosphoproteome [117]. For XL-MS this strategy involved a ‘multithreaded’ workflow (Fig 1) in which experimental steps were matrixed combinatorially (Table 1). In this way, individual XL were placed in a variety of ionic contexts for MS detection, and key interfaces were painted as clusters of alternative XL between closely spaced crosslinking sites. This was combined with the use of diverse XL search engines for the identification of crosslinked peptides, and the use of isotopically coded crosslinkers where available. Our 86-protein search database comprised the maximum set of viral proteins considered likely to be packaged [5]. For all but two of these proteins XL were detected, the exceptions being proteins A14 and I2. These two short proteins (90 aa, 73aa respectively) possess relatively few sites for crosslinking and trypsin cleavage (3 lys/2 arg; 4 lys/0 arg, respectively).
Due largely to the absence of strong corroborating data for our XL-MS dataset such as comprehensive atomic-resolution three dimensional structures, validation relied largely on statistics and trends. The effective FDR of 0.33% for the final dataset as a whole (“Results”), suggested a remarkably low level of bioinformatics noise. Consistent with this, non-target databases from uncorrelated proteomes, namely all human proteins or the non-packaged subset of Vaccinia proteins yielded very weak results in preliminary searches.
Alongside the detection of clear crosslinkome sub-networks (“Results”) were many single-detect inter-protein XL (DFscore = 1, S1 Fig). Notwithstanding the excellent bioinformatic signature for the dataset as a whole (above), it was difficult to ascertain to what extent the single-detect XL were real (from, for example, low abundance proteins, low abundance viroforms, inefficient XL, or poorly ionizing peptides), or represented biochemical noise (eg. virion dissociation pathways during virus preparation or specific experiments). On the one hand, evidence that single-detect XL were true positives included the tendency of single-detect crosslinking patterns within a protein sub-network to conform to patterns of XL with higher DFscore. For example, among the 22 inter-protein XL shown in Fig 5a, 18 were multi-detects vs. 8 single-detects, all contributing to the same overall crosslinking pattern. On the other hand, high DFscoring XL showed a higher ratio of biologically rational:non-rational XL than did single-detect XL, lending greater confidence to former. For example, among the 37 inter-protein XL in the dataset with DFscore > 5 (Table E in S1 Text), the number that were considered biologically rational exceeded the number designated non-rational by a factor of 9.5 while, among the single-detect XL from the same table, rational exceeded non-rational XL with a factor of only 1.5.
Transcriptosome proteins, albeit presumably packaged in relatively low abundance, nonetheless showed a number of strongly detected inter-protein XL (Tables F, G in S1 Text). Some of these, including some of the most strongly detected inter-protein XL in the dataset (S1C Table; Table F in S1 Text, orange), were between transcriptosome and membrane proteins (Table F in S1 Text), including ectodomains of the latter. These XL were considered biologically “non-rational” (above) since the transcriptosome is located within the virion core while the TM proteins surround it according to conventional models. They were strongly supported by their DFscores, were not filterable by raising score thresholds, and their DFscores did not drop when switching between singly- and dually-thresholded filtering (Materials & methods). We were therefore unable to falsify a hypothesis that contacts can occur between transcriptosome components and the ectodomains of membrane proteins, the significance of which is unclear. Possibilities for these resilient, yet ‘non-rational’ XL may include that: (a) the core wall is not a fundamental barrier to crosslinking (it is porous)—indeed the 7 nm inside-diameter pores that have been imaged in the core wall [9, 118] may be sufficiently large for the majority of Vaccinia polypeptides to pass through entirely if they are globular and approximately spherical [119], (b) TM and transcriptosome proteins are both implanted in the barrier (from opposite sites)–a situation, on the transcription side, observed in the cores of turreted Reoviruses [120, 121], (c) TM proteins are located in more than one compartment, (d) MV preparations contain developmental viroforms from a time prior to the full emergence of the core wall, (e) they are cryptically artefactual.
The dataset contained evidence for viroforms/proteoforms from proteolytic maturation. Peptides crossing known [2] sites of viral AG| specific proteolytic processing in proteins A17, VP8, G7, p4b and p4a (site2) can be found in tryptic digests of purified MV [5]. These peptides represent pre-cleaved proteoforms. Such peptides were also found in the current study, within crosslinked pairs, from proteins A17, VP8, A12 and G7 (Table H in S1 Text). XL connecting the N-terminal amino group of pre-cleaved A17 with the C-terminal region of the same protein (“Results”) may be an example of the same phenomenon. In some cases the crosslinker directly spanned an AG| processing site. Apparently, then, MV harvested from Hela cells late in infection followed by 2x sucrose gradient-purification were accompanied by immature viroforms that are detectable by highly sensitive MS. Among crosslinked peptides could be found no trace, however, of a characteristic and abundant marker of IV, namely the external scaffold protein D13 when using XL search databases that included this protein. Apparently, in MV, in which the external scaffold, along with fragment 2 of protein p4a (Fig 5b) are close to or below the detection limit, unprocessed forms of proteins A17, VP8 and A12 are still readily detectable. If, speculatively, MV preparations contain trace viroforms that appear morphologically mature (having already escaped the external scaffold and perhaps received tail-anchored and SFE proteins), but which still lack a fully formed interior and/or an impermeable core wall, then this may account for some of the more counter-intuitive XL detected here. Alternatively, some XL may represent structures that appear only transiently in the virion maturation pathway. Another possibility may be that MV particles, albeit fully mature, retain unprocessed proteoforms by design. Within an A17 homodimer, for example, one subunit might be processed and the other not.
Evidence for multiple viroforms/proteoforms also arose from interactions between p4a, p4b and TM proteins: p4a fragment 3 was found to be within crosslinking range of seven distinct membrane proteins (S1 Fig, Fig 6) and also within crosslinking range of the C-terminal region of p4b, while none of the seven membrane proteins were apparently within crosslinking range of p4b Fig 6, S2A and S2B Fig). Moreover, a crosslinking ‘hotspot’ in p4a fragment 3 (K736) interacted with three membrane proteins as well as p4b (S1 Fig), in the absence of any detectable crosslinking between the latter. While steric factors may allow p4a, p4b and membrane proteins to triangulate in a way that leaves all membrane proteins out of range of p4b, it seems also possible that membrane proteins and p4b may interact with alternate proteoforms of p4a. This could result from distinct and segregated p4a complexes within individual MV, or distinct viroforms in the virus preparation (e.g. the rearrangement of p4a fragment 3 during maturation).
In conclusion: Here, we have covered the crosslinkome of a relatively small whole organism in depth, detecting inter-protein XL for all but two of the 86 proteins that represent the maximal virion proteome. Strategies were developed to detect XL in a proteome covering a wide abundance dynamic range and with minimal pre-existing crystallographic information, allowing the reconstruction of several key virion protein complexes. The challenge of synthesizing the data into an extended understanding of the internal molecular architecture requires some knowledge of intra-particle protein stoichiometry.
Vaccinia virus was purified by sucrose or tartrate gradient as described [5] and protein quantitated using BCA (ThermoFisher Inc.), determining concentrations to be between 1 and 3.5 mg ml-1. DSS-H12, DSS-D12, DSG-H6, and DSG-D6 were obtained from Creative Molecules Inc. BS3-H4, BS3-D4, BS(PEG)5, BS(PEG)9, Zeba Spin Desalting Column (7K MWCO), and Lys-N were obtained from Thermo Scientific. DSS, bis(sulfosuccinimidyl)suberate (BS3), and disuccinimidyl glutarate (DSG) were used as 1:1 mixtures of DSS-H12/DSS-D12 (‘DSS-H12/D12’), BS3-H4/BS3-D4 (‘BS3-H4/D4’), and DSG-H6/DSG-D6 (‘DSG-H6/D6’) respectively. Trypsin, dimethyl sulfoxide (DMSO), Benzonase, iodoacetamide, n-LS, adipic acid dihydrazide (ADH), 4-(4,6-dimethoxy-1,3,5-triazin-2-yl)-4-methylmorpholinium chloride (DMTMM), 1-[bis(dimethylamino)methylene]-1H-1,2,3-triazolo[4,5-b]pyridinium 3-oxid hexafluorophosphate (HATU), and cyanogen bromide (CNBr) were from Sigma-Aldrich. GluC, AspN, LysC, LysN and ArgC were from Promega. AspN was from Roche Diagnostics. C18 and SCX filters were obtained from 3M. N,N-Diisopropylethylamine (DIPEA) was from Alpha Aesar. Centrifugal concentrators (Vivacon, 10kDa MWCO) were from Sartorius Stedim Biotech.
Prior to crosslinking, virus was washed 5x with phosphate buffered saline (PBS), pH 7.4, by centrifugation and resuspension. For some experiments, washed virus pellets were then resuspended in 10 μL of 0.1 M triethylammonium bicarbonate (TEAB, pH 8.5) and supplemented with an equal volume of 2x ‘pre-treatment’ buffer comprising either 0.1 M TEAB, 0.1% NP40 (pH 8.5), or 0.1 M TEAB, 0.1% NP40, 80 mM TCEP (pH 8.5), followed by 2 min incubation.
The method of ref. [122] (‘xQuest crosslink method’) was used with some modifications. Pre-treated virus suspension (above), or intact virus suspended in 0.1 M TEAB (pH 8.5), was supplemented with 1/10 volume of 10x crosslinking buffer (0.2 M 4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid (HEPES), KOH, pH 8.2). Crosslinker, dissolved freshly in DMSO, was then added at a final concentration of 7.5 mM. Following 30–60 min incubation at 37°C, samples were quenched by adding 1 M ammonium bicarbonate (AmBic) to a final concentration of 50 mM followed by 30 min incubation at 37°C.
ADH with either HATU/DIPEA or DMTMM: Pre-treated virus suspension (above) was supplemented with 10x ADH-XL buffer (0.2 M HEPES-NaOH, pH 7.2) to 1x ADH-XL buffer (final) then supplemented with ADH and HATU (dissolved separately in 1x ADH-XL buffer) to final concentrations of 6 mM and 9.2 mM respectively. 100% DIPEA was then added to a final concentration of 46 mM. After 120 min incubation at room temperature with continuous shaking, crosslinked virus was exchanged into 50 mM AmBic using a spin desalting column (Zeba, ThermoFisher, Inc.) following the manufacturer’s instructions. In some experiments, HATU/DIPEA were replaced with DMTMM, using concentrations of ADH and DMTMM described [123].
ADH/EDC/NHS or EDC/NHS alone: Pre-treated virus suspension (above) was supplemented with 10x ADH-XL buffer (0.2 M HEPES-NaOH, pH 7.2) to 1x ADH-XL buffer (final) then supplemented with ADH (dissolved separately in 1x ADH-XL buffer) to a final concentration of 6 mM. N-(3-dimethylaminopropyl)-N′-ethylcarbodiimide hydrochloride (EDC) and N-hydroxysuccinimide (NHS), dissolved separately in 1x XL buffer were then added at final concentrations of 8 mM and 10 mM, respectively. After 120 min incubation at room temperature, free crosslinker was removed by spin desalting into 50 mM AmBic (above).
ADH with EDC or EDC alone: Pre-treated virus suspension (above) was supplemented with 10x MES buffer (0.1 M 2-(N-morpholino)ethanesulfonic acid (MES), 20 mM NaCl, pH 4.7) to 1x MES buffer (final), then supplemented with ADH and EDC (dissolved separately in 1x MES buffer) to final concentrations of 6 mM and 2 mM, respectively. After incubation for 120 min at room temperature, free crosslinker was removed by spin desalting into 50 mM AmBic (above). For some experiments (EDC-alone crosslinking) ADH was omitted.
Crosslinked virus samples in 50 mM AmBic were disaggregated by supplementing with 0.5 M TCEP, 1 M TEAB and solid urea or guanidine, then diluting to achieve a final formulation of 8 M urea, 0.1 M TEAB, 10 mM TCEP, pH 8.5 (urea buffer) or 6 M GuHCl, 0.1 M TEAB, 10 mM TCEP, pH 8.5 (guanidine buffer). In some experiments, crosslinked virus suspension in 50 mM AmBic was instead supplemented with an equal volume of 2x detergent solution to achieve 0.5% sodium deoxycholate (SDOC), 12 mM n-laurosarcosine (n-LS), 5 mM TCEP, 50 mM TEAB, pH 8.5 (final). After 30 min incubation at 37°C, some samples were alkylated with iodoacetamide at either 5 mM (if supplemented with urea or GuHCl), followed by 30 min incubation in the dark) or 10 mM (if supplemented with detergents), followed by 15 min incubation in the dark. Some samples were then incubated with Benzonase (250 units) for 60 min. All samples were then diluted with 50 mM AmBic for cleavage, according to the manufacturer’s recommendation for tolerable denaturant (below).
Cleavage employed the following reagents/reagent combinations: Trypsin, ArgC, GluC, AspN, LysN, LysC, or Trypsin+GluC, Trypsin+AspN, ArgC+AspN, ArgC+GluC, AspN+GluC or CNBr+Trypsin.
For digestions containing GluC, samples were diluted to a final urea concentration of 0.5 M. For digestion with LysN, samples were diluted to a final urea concentration of either 1 M or 5 M. For all other proteases, samples were diluted to final denaturant concentrations of either 0.6 M GuHCl, 1 M urea, or 0.1% SDOC/2.4 mM n-LS/1 mM TCEP. With the exception of DigDeApr experiments (below), a protease:substrate ratio of 1:50 or 1:100 was used. With the exception of LysC, which was used for 72 hr at room temperature, all protease digestions were overnight at 37°C.
For CNBr+Trypsin digestion, quenched amine-amine crosslinking samples (above) were supplemented with 100% formic acid (FA) to 70% (final) followed by the addition of one crystal (~20–100 molar excess) of CNBr and overnight incubation at room temperature in the dark. After evaporation to dryness under vacuum, samples were redissolved in urea buffer (above), followed by 30 min incubation at 37°C in the dark. Samples were then diluted to 1 M urea with 50 mM TEAB (pH 8.5), and trypsin added to an estimated enzyme-to-substrate ratio of 1:100 followed by overnight incubation at 37°C. A fresh equivalent of trypsin (same amount) was then added, followed by a further 4 hrs digestion. Undigested material was precipitated by centrifugation at 14,000 g for 2 min followed by resuspension in 70% FA and re-digestion with CNBr and trypsin following the same method.
This was done following ref. [124] with modifications. Briefly, samples were digested with either trypsin alone (enzyme:substrate ratio of 1:2500) or Trypsin+AspN, Trypsin+GluC or AspN+GluC (1:1:2500). After overnight incubation at 37 °C, samples were centrifuged into a centrifugal concentrator (10kDa MWCO, Vivacon) at 2500 x g. After collection of flow through, the filter was washed by centrifugation at 2500 g with 8 M urea, 0.1 M TEAB pH 8.5 then with 2 M urea, 0.1 M TEAB pH 8.5 (Wash buffer) for 2 min at 2500 x g. Flow through and wash-throughs were combined. Using a new collection vial, urea buffer was added to the filter which was then inverted and spun for 2 min at 2500 x g. The process was repeated and the combined urea buffer washes were brought to 1 M urea with 0.1 M TEAB then treated again with the same protease combination at an enzyme:substrate ratio of 1:100 (for GluC digestion, samples were diluted to 0.5 M urea, 100 mM TEAB) overnight at 37 °C.
All cleaved samples were acidified with FA to 2% FA final then desalted as described [125] using stacked C18-SCX filters. After washing the filters, peptides were transluted from the C18 to the SCX phase using 80% CH3CN/0.1% FA (translution buffer). Peptides were eluted with 5% NH4OH/80% CH3CN (Buffer X) or with six steps of 20% CH3CN/0.5% FA containing ammonium acetate in the range 160–800 mM followed by a final step of Buffer X. Elutions were dried under vacuum then re-dissolved in 0.1% FA in water for MS.
nanoLC-MS/MS was performed using an LTQ Velos Pro Orbitrap mass spectrometer with Easy-nLC 1000 (ThermoFisher). 2 microL injections were followed by a segmented LC gradient (solvent A = 0.1% FA in water, solvent B = 0.1% FA in CH3CN), progressing from 0 to 10% B over 10 min then to 35% B over 230 min. Some runs used a straight gradient of 0–35% B over 135 min. Precursor spectra were acquired in FT mode at a resolution of 100,000 (centroid) in the range 200–2000 Th. For isotopic pairs with 12 Da mass split (DSS crosslinker), the top 3 most intense ions were selected for HCD activation (above a precursor signal threshold of 150) on the basis of isotopic pairs with m/z spacing of either 4.02524, 3.01893 or 2.41515 (representing +3 to +5 charge-states), and intensity ratio better than 2:1. For a 6 Da mass split (DSG crosslinker), m/z deltas for isotopic pair selection were 2.01456, 1.51092 or 1.20874. For a 4 Da mass split (BS3 crosslinker), m/z deltas were 1.34156 or 1.00616. Both isotopic partners were fragmented. HCD activation used a normalized collision energy (NCE) of 45, activation time of 0.1 mSec and an isolation width of 2 m/z. MS2 spectra were acquired in FT mode with a resolution of 7500 (centroid). The dynamic exclusion list size was 500, exclusion duration was 60 sec, repeat duration was 30 sec and the repeat count was 2, with early expiration enabled. Charge state screening was enabled, with rejection of 1+ and 2+ and unassigned charge states.
Data acquired for xQuest were activated in IT-CID mode instead of HCD. Here, NCE was 35, activation Q = 0.25 and activation time was 10 mSec. For non-isotopic crosslinkers, the 10 most intense ions in each precursor spectrum were subjected to HCD fragmentation, as above, if above a minimum signal threshold of 250 (or 2000 in some early experiments).
Protein names used throughout this report follow entry names in the UniProtKB Vaccinia WR reference proteome minus the species identifier suffix. They are comprehensively cross-referenced to other naming schemes in Table S1 of ref. [5]. Instrument raw files were converted to mgf, mzXML or mzML using MSConvert by ProteoWizard. Using the resulting data, XL were identified using the following XL search engines: Protein Prospector [126], pLINK [127], xQuest (in combination with xProphet) [122], Kojak [128] (in combination with ‘Percolator’ [129–131]), ECL [132] and ECL2 [133], as follows:
Protein Prospector: Instrument raw data files were converted to mgf format then uploaded to Protein Prospector via the UCSF online server using parameters outlined in Table I in S1 Text. Non-standard, PEGylated bis(sulfosuccinimidyl)suberatecrosslinkers (BSPEG5 and BSPEG9) were imputed as user defined parameters. The results file from each run was generated using the program’s Search Compare function. Results were sorted by ascending expectation value and “Report type” was set to “crosslinked peptides”. ‘SD-E’ = ScoreDiff–log10(Exp2) ([126], Robert Chalkley, Personal communication) where ScoreDiff is the difference in score between the top- and second-ranked peptide 1 in the search output for a crosslinked pair, Exp2 is the score for peptide 2.
pLink: pLink was downloaded from pFind Studio. A parameter file was configured for each experiment and a folder created, containing mgf files pertaining to that experiment along with the search DB. The ‘pLINK.ini’ configuration file was modified for each experiment to include the path to the mgf and search DB and search parameters (Table I in S1 Text). The enzyme.ini and xlink.ini files were modified for any non-standard cleavage specificities/combinations and crosslinkers, respectively. Results files for loop linked and mono linked peptides were generate using “non-interexport” and “drawpsm”, respectively. pLink was run through the flow.exe application.
xQuest: The xQuest VMware package was installed on a Windows PC. Directories were created following instructions provided with xQuest. Search parameters are given in Table I in S1 Text. Files “Xmm.def” and “xquest.def” were modified for the relevant crosslinker isotopic mass, shift and ion charge states. A text file was created containing the mzML file name and parameter files for xProphet. xQuest, then xProphet, were run from the command line. Results were viewed on the xQuest webserver then downloaded. Values reported by xProphet as "FDR" may be Percolator-derived q-values.
Kojak: Kojak and Percolator [134] were installed and run in Linux from the command line. Folders were created for mzML formatted data and search results. The program’s configuration file was modified to contain all relevant crosslinkers and the paths to individual data files. Parameters are outlined in Table I in S1 Text. Digestion specificity rules were based on the parameters provided.
ECL/ECL2: ECL and ECL2 were installed on a Java-capable Windows PC and run from the command line. The program’s parameter file was modified to contain search parameters given in Table I in S1 Text. Crosslinker masses were entered manually.
Percolator: For Kojak and ECL, FDR was converted to a q-value using the program ‘Percolator’ [134], run from the Linux command line. For Kojak, Percolator input comprised “inter”, “intra”, and “loop” search output files. q-value can be regarded as the expected proportion of false positives among all features as or more extreme than the observed one [132, 135] or, alternatively, the minimal FDR threshold at which a given peptide-spectral match is accepted [130, 131].
Data assembly: Using in-house code, XL search engine/Percolator/xProphet outputs corresponding to various nanoLC-MS/MS runs in various experiments were parsed in their native formats, accepting individual XL to a single unified dataset according to dual score thresholds for each program including in-house-calculated FDR for Protein Prospector (see above and Table 2). The resulting dataset was then sorted by ascending exp_Mr. Groups (blocks) of masses matching to within 25 ppm were annealed, then each block that contained multiple accession/PeptideSeq/ProteinPos was sorted and divided into distinct sub-blocks of ions that were tagged with a common ‘ambig code’ (representing sub-blocks having functionally isomeric mass but had been assigned, by XL search engines, distinct apparent identities). The resulting ‘mature blocks’ each represented a unique combination of exp_Mr and accession/PeptideSeq/ProteinPos.
This dataset was reformatted/collapsed into a matrix with one row per mature block, and one column for each nanoLC-MS/MS run in the project. The matrix was filled with XL search engine identifiers to indicate all engines identifying a specific mature block member in a specific nanoLC-MS/MS run and the number of times identified. Each row was assigned a DFscore as the sum of search engine identifiers/times identified by that engine that had been assigned to the row. Groups of mature blocks sharing a common ambig code were likelihood-scored against one another as follows: If they all represented intra-protein XL or all represented inter-protein XL, then the ambigscore assigned to those mature blocks was a simple proportion of its DFscore/∑(DFscores for all blocks sharing a common ambig code). If they were a mixture of intra-protein and inter-protein, then intra-protein mature-block(s) were scored 1.0 and inter-protein mature-block(s) 0 (assuming the intra-protein XL to be correct by default). If the ambiguity was simply in choice between multiple lysines within an otherwise identical peptide, both choices were scored 1.0 since both reflect the same approximate position within the same protein partners. Finally, for every specific position in a specific protein represented by multiple rows in the matrix: If the intra-protein XL were discovered by multiple engines and the inter-protein XL were discovered by one only, the latter were annotated as “filterable”. The resulting annotated matrix was written to an Excel worksheet then copied to a second sheet which was re-sorted by protein position then accession.
A ‘discard matrix’ (comparable in structure to the above, ‘passing’ matrix) was generated representing all XL ions in the above assembly that passed threshold1 but were rejected after failing threshold2. Each row of the ‘discard’ matrix was annotated with: (a) DFscore; (b) whether the XL (Accession1/Accession2/ProteinPos1/ProteinPos2) was also present in the passing matrix (above; this criterion being denoted ‘also’ in the following discussion) and (c) ‘biological rationality’ (‘BR’) based on six groups of functionally-related virion proteins (Table 3), annotating”Y”, if the two crosslinked proteins were in same BR group, and “N” if one was from the ‘membrane’ group and the other from the ‘transcription’ group.
Networks and sub-networks for individual accessions or groups of accessions were rendered using CrosslinkViewer [136]. Using in-house code, rows in the passing matrix (above) were picked provided either one or both crosslinked proteins did not match accessions within a user-definable excluded-accession group. ‘Filterable’ rows of the matrix (above) were excluded. The list of picked rows was supplemented with those from the ‘discard’ matrix (above) if DFscore > 1 or ‘also’ = “Y” or BR = “Y”. Rows with common Accession1/Accession2/ProteinPos1/ProteinPos2 were then collapsed summing DFscores, and the resulting dataset reformatted for input to CrosslinkViewer. The resulting DFscores were rendered. If 100% of matrix rows for a given Accession1/Accession2/ProteinPos1/ProteinPos2 had been flagged as ambig (above), then the XL was flagged to be rendered with a broken line. Protein monolinks were ignored in all data operations.
Domain prediction: TM regions were predicted using program TMHMM [68, 95]. Domain boundaries were predicted using DomPred [137, 138]. Output traces show endpoint density profiles for PSI-BLAST alignments generated between a query sequence and a database in which all sequence fragments had been removed.
ROC analysis of the global crosslinking dataset: Each of the 81 proteins in the dataset for which crosslinked partner proteins were found, was flagged according to membership of one of two ‘biological rationality’ groups in Table 3 (‘Membrane’ and ‘Transcription’), and total number of distinct crosslinking partner proteins was printed alongside. In each of two replicates of this listing was printed the # of partners belonging specifically to one of the two groups and the list was sorted (descending) according to proportion of total partners belonging to the specific group. After incrementing four number series at each row in the list that contained either: a membrane protein, not(a membrane protein), a transcription protein and not(a transcription protein), then proportionating each series to a scale from 0 to 1, ROC curves were drawn based on the proportionated values. In ROC space, points above and below the line of no-discrimination (diagonal) represent positive (better than random) and negative correlation, respectively such that a curve representing perfect positive correlation would ascend vertically from (x,y) = (0,0) to (0,1) then travel horizontally to (1,1). Perfect negative correlation would yield the converse curve: (0,0) to (1,0) to (1,1).
Inter-protein XL partitioning analysis: For each XL ion in the global XL dataset (each row of S1A Table) representing an inter-protein XL, DFscores from each experiment (column) in which the ion was detected were binned according to whether sample pre-treatment included or excluded NP40 or TCEP (-NP40, +NP40, -TCEP or +TCEP). The resulting tetra-bin DFscore values for individual ions were accumulated on a per-accession basis, according to the accession on each side of the crosslink. The accumulated four DFscore values for each accession were then converted to a proportion of the summed DFscore across the four bins (‘POSD’) for that accession, and the resulting POSD values were finally normalized to the mean POSD per accession for each of the four pre-treatment conditions. See legend to S3 Fig for further details.
Distance analysis of the form shown in ref. [29]) was generated using ‘TopoLink’ [38] installed on a computer cluster and run from the command line, in combination with 14 relevant pdb files (Table B in S1 Text). Before calculating Euclidean and SAS distances for each experimental XL, “inputfile.inp” was modified to include the crosslinker type, maximum linker distance and reactive residues. All lys-lys Euclidean and SAS distances were also calculated within the 14 structures, setting maximum linker distance to 100 Å. Each of the resulting four distance datasets, in spreadsheet format, was binned for display as a histogram. The mean and standard deviation (SD) from each “all lys-lys distances” histogram informed a normal (Gaussian) curve overlay. For each “experimental XL distances” histogram, Ln(mean) and ln(SD) informed a log-normal curve overlay. “Experimental XL distances” and “all lys-lys distances” datasets were compared to one another via the Kolmogorov Smirnov test, run using the Excel plugin XLSTAT (https://www.xlstat.com/en/).
|
10.1371/journal.pcbi.1000500 | Parallel Computational Subunits in Dentate Granule Cells Generate Multiple Place Fields | A fundamental question in understanding neuronal computations is how dendritic events influence the output of the neuron. Different forms of integration of neighbouring and distributed synaptic inputs, isolated dendritic spikes and local regulation of synaptic efficacy suggest that individual dendritic branches may function as independent computational subunits. In the present paper, we study how these local computations influence the output of the neuron. Using a simple cascade model, we demonstrate that triggering somatic firing by a relatively small dendritic branch requires the amplification of local events by dendritic spiking and synaptic plasticity. The moderately branching dendritic tree of granule cells seems optimal for this computation since larger dendritic trees favor local plasticity by isolating dendritic compartments, while reliable detection of individual dendritic spikes in the soma requires a low branch number. Finally, we demonstrate that these parallel dendritic computations could contribute to the generation of multiple independent place fields of hippocampal granule cells.
| Neurons were originally divided into three morphologically distinct compartments: the dendrites receive the synaptic input, the soma integrates it and communicates the output of the cell to other neurons via the axon. Although several lines of evidence challenged this oversimplified view, neurons are still considered to be the basic information processing units of the nervous system as their output reflects the computations performed by the entire dendritic tree. In the present study, the authors build a simplified computational model and calculate that, in certain neurons, relatively small dendritic branches are able to independently trigger somatic firing. Therefore, in these cells, an action potential mirrors the activity of a small dendritic subunit rather than the input arriving to the whole dendritic tree. These neurons can be regarded as a network of a few independent integrator units connected to a common output unit. The authors demonstrate that a moderately branched dendritic tree of hippocampal granule cells may be optimized for these parallel computations. Finally the authors show that these parallel dendritic computations could explain some aspects of the location dependent activity of hippocampal granule cells.
| Neurons possess highly branched, complex dendritic trees, but the relationship between the structure of the dendritic arbor and underlying neural function is poorly understood [1]. Recent studies suggest that dendritic branches form independent computational subunits: Individual branches function as single integrative compartments [2],[3], generate isolated dendritic spikes [4],[5] linking together neighbouring groups of synapses by local plasticity rules [6]–[8]. Coupling between dendritic branches and the soma is regulated in a branch-specific manner through local mechanisms [9], and the homeostatic scaling of the neurotransmitter release probability is also regulated by the local dendritic activation [10].
The computational power of active dendrites had already been demonstrated by several computational studies [11]–[16], but how local events influence the output of the neuron remained an open question. Using the cable equation [17] or compartmental modelling tools one can calculate the current or voltage attenuation between arbitrary points in a dendritic tree [14], which is in good agreement with in vitro recordings. However, cortical networks in vivo are believed to operate in a balanced state [18],[19], where the inhibitory drive is continuously adjusted such that the mean activity of the population is nearly constant [20],[21]. In this case, the firing of an individual neuron is determined, beyond its own input, by the activity distribution of the population. A simple cascade model [22] incorporating numerous dendritic compartments allowed us the statistical estimation of the activity distribution of neurons within the population. We used this model to study how localized dendritic computations influence the output of the neuron.
The present study focuses on hippocampal granule cells. Compared to pyramidal neurons granule cells have relatively simpler dendritic arborization: They lack the apical trunk and the basal dendrites, but are characterized by several, equivalent dendritic branches, extended into the molecular layer [23] (Figure 1A). Recordings from freely moving rats revealed that like pyramidal neurons, granule cells exhibit clear spatially selective discharge [24],[25]. However, granule cells had smaller place fields than pyramidal cells, and had multiple distinct subfields [24],[26]. It has also been recently shown that these subfields are independent, i.e., their distribution was irregular and the transformation of the environment resulted in incoherent rate change in the subfields [26]. The dendritic morphology of granule cells suggest that parallel dendritic computations could contribute to the generation of multiple, distinct subfields of these neurons.
In the present study we analyzed how synaptic input arriving to dendritic subunits influence the neuronal output. First, we introduce the model used in this study and we define statistical criteria to measure if a dendritic branch alone is able to trigger somatic spiking. We show, that generally neurons perform input strength encoding i.e., input to the whole dendritic tree but not activation of a single branch is encoded in the somatic firing. Next we demonstrate that if the local response is enhanced by active mechanisms (dendritic spiking and synaptic plasticity) then neurons switch to feature detection mode during which the firing of the neuron is usually triggered by the activation of a single dendritic branch. Furthermore we show that moderately branched dendritic tree of granule cells is optimal for this computation as large number of branches favor local plasticity by isolating dendritic compartments, while reliable detection of individual dendritic spikes in the soma requires low branch number. Dendritic branches of dentate granule cells could therefore learn different inputs; and the cell, activated through different dendritic branches, could selectively respond to distinct features (locations), participating in different memories. Finally using spatially organized input we illustrate that our model explains the multiple independent place fields of granule cells and these dendritic computations increase the pattern separation capacity of the dentate gyrus.
We set up a cascade model [22] to study the somato-dendritic interactions in neurons, that is simple enough for mathematical analysis but can be adequately fitted to experimental data. The long, parallel branches of dentate granule cells are represented by distinct compartments connected to the somatic compartment of the model (Figure 1B). The activation of the somatic and dendritic compartments are described by the following equations:(1)(2)where Cm is the membrane capacitance, and are the total dendritic and somatic membrane resistances, respectively, and is the axial resistance between the dendritic and the somatic compartments. Each of the N dendritic branches are contacted by M presynaptic axons, uj is the firing rate of axon j, and wij is the synaptic strength between the dendritic branch i and presynaptic axon j (see Methods for parameters specific to hippocampal granule cells). f(U) is the dendritic integration function that specifies the form of the local integration of synaptic inputs, and Ui = Σjwijuj is the total synaptic input to a given branch. Because the firing rate of the presynaptic entorhinal neurons depend mostly on the location of the animal [27] we assume, that input varies slowly compared to the membrane's time constant in dentate granule cells (τm≈37 ms, [28]). Therefore, we rewrite Equations 1–2 to their steady-state form:(3)(4)where and is the proportion of the axial and the membrane resistivity. In granule cells the area and the electrical resistance of the somatic membrane is similar to the membrane area and resistance of a single dendritic branch [28] (see Methods). Therefore, in the following calculations we use R = Rs = Rd to denote the electrical isolation between somatic and dendritic compartments. Three different functions were used in this study to approximate the local integration of synaptic inputs within the dendritic branches of hippocampal granule cells (Text S1, Figure 1C): a linear (FL(U) = 0.26U) and a quadratic (FQ(U) = 0.13U2) function were used in the analytical calculations; and the results were also tested with a sigmoid () function in the Supporting Information (Text S2). We also performed some of the analytical calculations by decreasing the degree of nonlinearity, where we used FC = 0.07U2+0.12U or FC = 0.02U2+0.22U. Note, that the action potential generation is not incorporated in the model, and all active properties of the dendrites are modeled by the integration function F(U).
Supposing that firing rates of presynaptic neurons (uj) are independent and identically distributed we assume that the total input of the dendritic branches Ui = Σjwijuj is drawn randomly from a Gaussian distribution with mean μ and variance σ2:(5)where p[U] indicates a probability distribution over U (Figure 1C; see Eq. 17 in Methods for parameters specific to hippocampal granule cells). More specifically, indicates the distribution of the magnitude of possible total inputs to a single dendrite over many different instances. Based on the distribution of the total input, we can compute the distribution of the somatic activation and determine the firing threshold (β) according to the proportion of simultaneously active cells (the sparseness of the representation, spDG) in the DG [24]. First, we rearrange Eq. 3 using the input distribution to express the distribution of :(6)where indicates that the inputs of the dendritic branches are randomly sampled from a Gaussian distribution. We substitute Eq. 6 into Eq. 4, and we get(7)We can assume again, that the inputs (Ui) of the dendritic branches are independent and identically distributed variables. (Note, that while the activations are not independent because of the back-propagation of currents from the soma, the inputs are.) If N is high enough, we can approximate the sum in Eq. 7 with a Gaussian distribution, and rewrite the equation:(8)where indicates a probability distribution over , while μF and are the expected value and the variance of the dendritic integration function F(U) given the input distribution :(9)(10)We calculated the integrals 9–10 with two different forms of dendritic integration of synaptic inputs: a linear and a quadratic function (Figure 1C). The details of these calculations are in the Supporting Information (Text S4).
In this paper we do not model inhibitory neurons in the dentate gyrus, however, we assume, that they play a substantial role in continuously adjusting the firing threshold of principal neurons and regulating the activity of the network [20],[21]. As a result of this regulation always the most depolarized neurons are able to fire, and the proportion of simultaneously active neurons is characteristic for different hippocampal areas [24],[29]. Given that all neurons share a common input statistics and have similar internal dynamics, equation 8 also describes the distribution of across the granule cell population at a given time. If only the most depolarized 1–5% of the population are able to fire [29], this also means that only those neurons exceed their firing threshold whose activation is within the uppermost 1–5% of the distribution described by Eq. 8. Therefore, the proportion of simultaneously active neurons within the dentate gyrus spDG [24],[29] also determine the firing threshold β for granule cells.
We approach the dendritic independence by focusing on the statistical distributions of the input to dendritic branches, as these branches form the basic computational subunits in our model. We ask whether the input of a single branch could be sufficiently large to significantly depolarize not only the given branch but also the soma of the neuron. We defined two conditions to study whether the spiking of the neuron is caused by the activation of a single dendritic branch or by the simultaneous depolarization of multiple branches.
First, the conditional probability is the probability of firing given that any branch k has total input Uk = Σjwkjuj, while inputs to all other branches are random and independent samples from the distribution of (Figure 2A). At those Uk values where this probability is close to 1 the cell tends to fire when any of the dendritic branches gets that input. Second, the conditional distribution is the distribution of the synaptic input of the most active branch at the time the depolarization of the soma exceeds the firing threshold (β), where U* is the total synaptic input arriving to the most active branch (Figure 2A). K(U*) can be regarded as the marginal distribution of above the firing threshold (Figure 2B). The probability mass of this function shows the typical maximal input (U*) values when the neuron fires.
These two conditions together determine whether a single branch can be sufficiently depolarized to trigger somatic spike or not. If the probability of firing is high (H(U)≈1) at typical input values (K(U*)) then the firing of the cell is caused by a single branch. With the definition of Gasparini and Magee [30] we call this form of information processing as independent feature detection. On the other hand, if the firing probability is low (H(U)≪1) even if one of the branches receive extremely large input (U* is high) then the cell mostly fires when the overall dendritic activation is high, and even the most depolarized branch usually fails to make the neuron fire. We use the expression input strength encoding [30] to denote this second type of computation. The calculation of the two functions H(U) and K(U*) is described in the Methods section.
First we chose unstructured synaptic input, i.e., the firing of entorhinal neurons were independent and the strength of all synapses were equal. In this case we approximated the total synaptic input U to a branch with a Gaussian distribution (Eq. 5, Figure 1C). Given the input distribution we asked whether the excitation of single branches can be sufficiently large to cause significant depolarization in the soma.
The typical largest input values, indicated by the probability mass of K(U*) (Figure 2C–D) are unable to sufficiently depolarize the soma and determine the neuronal output (indicated by the low H(U) values) in the case of both the linear (Figure 2C) and the quadratic (Figure 2D) integration functions. Wherever K(U*) has high values, H(U) is low in both cases, which indicate, that these branches are not able to independently influence the output of the neuron. Only coactivation of several branches could make the neuron fire in this case, and the output of the neuron encodes the strength of all dendritic inputs. As H(U) converges to 1 for high input values extremely high inputs to a single dendrite could reliably trigger somatic firing. In the next sections, however, we study how synaptic plasticity selectively modifies individual synapses and contributes to the sparse occurrence of extraordinarily high input values.
During Hebbian learning synapses contributing to postsynaptic activation are potentiated while other synapses may experience compensatory depression [31],[32]. We simulated the learning process by showing a finite number of uncorrelated samples from the input distribution (see Methods) to the model neuron initiated with uniform synaptic weights. The synaptic weights of those dendritic branches where the activation exceeded a threshold, βd were modified according to the following Hebbian plasticity rule [33] that incorporates heterosynaptic depression [31]:(11)where is the local dendritic activation, uj is the presynaptic firing rate and wij is the synaptic strength. is the Heaviside function and γ<1 is a constant learning parameter. Note, that the learning rule is local to the dendritic branches: the synaptic change depends on the local activation but not on the somatic firing.
Next, we calculated the total input to the branches Ui = Σjwijuj after modification of synapses (Figure 3A), and recalculated the two functions H(U) and K(U*) defined previously with the new input distribution (Eq. 18). As shown on Figure 3A the total synaptic input in response to a learned pattern increases significantly after learning (compare blue and grey curves on Figure 3A), while untrained patterns generate smaller synaptic inputs (compare grey and black curves on Figure 3A). The main consequence of synaptic plasticity is that the trained patterns generate much larger local response than untrained patterns, which raise the possibility of their detection in the soma. Note, that an unspecific increase of synaptic weights would result in an upward shift of both the input distribution (Eq. 5) and the firing threshold, but would not affect the somatic detection of individual dendritic events.
The neuron is able to selectively respond to the dendritically learned patterns if a single branch, when facing with its preferred input, is able to induce significantly more depolarization at the site of the action potential initiation compared with the case when all of the branches get random, not learned input. Figure 3B–E shows the dendritic input and the activation of the soma after learning. If the maximal input U* is small (left bumps on Figure 3B,D) and none of the branches got its preferred input then the somatic activation is usually small. If U* is high (Figure 3B,D; right bumps), which means that one of the branches receives its preferred input pattern, then the somatic activation is increased. The increase of the somatic activation with learned input is only moderate in the linear case (Figure 3B,C) resulting in an incomplete separation of learned and not learned inputs by the somatic firing threshold. However, if synaptic inputs are supra-linearly (quadratically) integrated within the dendritic branches, efficient separation is possible: the probability that the presentation of a learned pattern elicits subthreshold somatic response, called dendritic spike detection probability was over 95% (Figure 3D,E). In this case the output of the neuron encodes whether or not one of the stored features was present in the neuron's input and not simply the strength of the total input arriving to the whole dendritic tree. In other words, if dendritic nonlinearity enhance the response of a given branch to its preferred input, then this branch alone is able to trigger somatic spiking. In the following sections we use the term dendritic spiking to refer to these supra-linear dendritic events. Although there is no data available on the synaptic induction of local dendritic spiking in hippocampal granule cells, voltage dependent Ca2+ currents are present in the membrane of granule cells [34],[35] and whole-cell recordings from these neurons suggest that T-type Ca2+ channels can generate dendritic action potentials at least in young neurons [36] or under hyper-excitable conditions [34],[37].
Next, we explored how the independent feature detection ability of the model depends on the resistance between the somatic and dendritic compartments with nonlinear dendritic integration. In the passive cable model of dendritic trees the space constant of the membrane λm≈(Rm/Ri)1/2 plays a substantial role in determining the voltage attenuation among two sites. Consequently, an increase in the intracellular resistivity Ri or a similar decrease in the membrane resistance Rm will contribute to the separation of dendritic subunits by decreasing the membrane's space constant λm. In the present study we used the inverse of the space constant R≈Ri/Rm to characterize the degree of electrical resistivity between the somatic and dendritic compartments. Indeed, an increased resistivity (R) between the compartments (smaller space constant) induced larger degree of electrical isolation as the somatic response to the same amount of dendritically applied current decreased (compare Figure 4A left and right panels). However, this isolation did not modify the dendritic spike detection probability in the soma: Large dendritic spikes localized to a single compartment could be reliably separated from subthreshold events with a somatic firing threshold at a large range of resistances R (Figure 4A–B). This was also true for the selective alternation of the somatic or the dendritic membrane resistance (Figure 4B).
On the other hand, the resistance parameter had a substantial impact on the isolation of different dendritic compartments which might be necessary for the independence of synaptic plasticity. To measure the isolation of the dendritic subunits we calculated the influence of other compartments on the activation of a given branch (external influence) quantified by the standard deviation of . Figure 4C shows the activation of a dendritic branch in the function of its input at different R values. If the resistance is small (, Figure 4C, left), then the local activation depends only slightly on the local input and the external influence is high (Figure 4D). In this case the local input spread out to the entire dendritic tree and activates similarly all branches. On the other hand, if the resistance is high (R = 1, Figure 4C, right) then the external influence is small, and the depolarization of a dendritic branch depends mostly on the local input. Interestingly, decreasing the resistance of the perisomatic membrane () alone was more efficient in separating the dendritic subunits than decreasing the resistance of the dendritic membrane or both (Figure 4D). The extensive GABAergic [38],[39] and glutamatergic [40] innervation of the proximal dendritic and perisomatic region of granule cells may therefore contribute significantly to the isolation of the dendritic compartments.
The impact of a single branch on the somatic activation, and also the coupling between dendritic branches may depend highly on the structure of the dendritic tree. Therefore we varied the number of dendritic subunits, N, and calculated the probability of detecting dendritic spikes in the soma and the external influence on the dendritic subunits (Figure 5). The probability of detecting a dendritic spike in the soma decreased gradually after a few (N≈30) number of branches from 1 to 0.3 (N≈1000, Figure 5A–B). If the number of branches was low, then the effect of a single branch on the soma was relatively high, and the somatic detection of single dendritic events was reliable. Conversely, one out of hundreds of branches had relatively low impact on the neuron's output even if the local depolarization was significant.
The electrical coupling between the dendritic subunits characterized by the external influence on the local activation also decreased with the number of branches, (Figure 5C–D). In the model the branches are connected through the somatic compartment, and because the variance of the somatic activation decreases if N increases (Eq. 8), the external influence will also decrease. However, in a complex dendritic tree containing higher number of subunits the branches are electronically more isolated which is required for local plasticity. To keep the probability of dendritic spike detection high and the dendritic coupling low at the same time, the number of branches should therefore be as high as possible, but not higher than N≈60.
As we showed on Figure 4, the dendritic coupling depends on the resistance R, as high resistance separates better the subunits. Therefore we conclude, that a medium number of branches with relatively high resistance is ideal for parallel dendritic computations. The optimal number of dendritic subunits, however, depends on the size of the dendritic event determined by the local integration of the synaptic inputs (Figure 5B). Appropriate detection of dendritic responses to learned patterns with linear integration is possible only in very small dendritic trees, whereas supra-linear integration allows the detection of individual dendritic events also in a larger dendritic arbor. Nonlinear integration by dendritic spiking therefore permits the neuron to selectively respond to a larger number of distinct input pattern.
During the calculation above we assumed, that the activity of the presynaptic neurons are independent and that the samples from the distribution are uncorrelated. It is known, however, that the firing of entorhinal neurons are not independent: At least half of layer II cells in the medial entorhinal cortex (EC) are grid cells, whose firing depend mostly on the position of the animal [27]. Moreover, in reality animals do not face with discrete uncorrelated samples, but they experience the continuous change of their environment which is mirrored by the activity of the entorhinal neurons. In order to test our model under more realistic conditions, we simulated the activity of the rodent's EC during exploratory behavior as input to our modeled granule cell. The EC consisted of two neuron population: A population of grid cells (1000 neurons, 5 spacing, 5 orientations) representing a path integrator system [41] and a population of visual cells (1200 units), representing highly processed sensory information available in the EC [42]. In these simulations we used the Webots mobile robot simulator [43].
The firing statistics of the entorhinal neurons was the same as used in the analytical calculation except that the activity of the neurons was location dependent. Moreover, as we simulated the trajectory of the rat during continuous foraging for randomly tossed food pellets [26] the subsequent input patterns were highly correlated. We simulated a single granule cell with N = 20 dendritic branches each of them receiving a total number of M = 100 synaptic contacts from entorhinal neurons. The resistance was R = 1, we used the quadratic integration function and the neuron was tested in 5 different environments. During the 5 min. learning period (while 2000 spatial locations was sampled with an average running speed of 0.22 m/s) 0–8 branches learned usually at different spatial locations in each of the 5 environments. In most of the time synaptic plasticity in different branches occurred at different places, therefore the subunits were able to learn independently. Moreover, learning occurred only in naive branches, i.e., each branch learned only in one environment at a specific location and synapses of trained branches did not engage in learning at a different location. After the training period the synaptic weights of those branches that were subthreshold for synaptic plasticity (βd = 1.11) in all environments were scaled down manually.
Next we studied the spatial activity pattern of the somatic and dendritic compartments while the robot was moving on a different track in the same environments. The dendritic branches responded with high activation (“dendritic spikes”) to subsequent visit of places close to their preferred locations leading to the formation of dendritic place fields (Figure 6). Moreover, since the activation of the soma was substantially increased in each of these dendritic place fields, the neuron had a multi-peaked activity map in several environments (Figure 6).
Finally we explored the effect of the size of the dendritic tree on the spatial firing pattern of the neuron (Figure 7). If there were only a few functional dendritic subunit than the neuron obviously had a small number of dendritic place fields (Figure 7A), but the individual branches had strong influence on the somatic activity. Therefore the correlation between the somatic activation as and the maximal dendritic input U* was high (Figure 7B,C), as predicted by the analytical calculations. On the other hand, in neurons with large number of dendritic subunits there were more dendritic place fields (Figure 7A), but a single branch had only a little impact on the activity of the neuron (Figure 7D). Accordingly, the correlation between the maximal dendritic input and somatic activation was reduced (Figure 7B). In these cases the cell fired when the overall excitation was high or when more than one branch were simultaneously excited. Therefore, the moderately branching dendritic tree of granule cells seems optimal for parallel dendritic computations since extensive branching inhibits the detection of individual dendritic events.
We conclude, that clustered plasticity together with dendritic spiking may be an adequate cellular mechanism to explain the generation of multiple place fields in the DG [24],[26].
In the present paper we set up a statistical criteria to determine the effect of single dendritic events on the output of the neuron. Using this criteria we have shown that by supra-linear dendritic integration, given that branches have learned different input patterns, individual dendritic branches are able to trigger somatic firing. Next we have shown that high resistivity and large number of branches supports the segregation of dendritic subunits required for local plasticity. On the other hand, a single branch has a substantial effect on the output of the neuron only if the number of branches is sufficiently low. Finally using spatially organized input we have demonstrated that parallel computational subunits explain multiple, independent place fields of hippocampal granule cells.
Dendritically generated spikes mediated by voltage-gated Na+ [3] and/or Ca2+ channels [44] as well as glutamate-activated N-methyl-D-aspartate (NMDA) channels [45] have been described in a variety of neurons (for a review see [46] or [47]) including hippocampal granule cells [34]–[37]. We used a quadratic integration function in order to analytically model supra-linear dendritic integration [15] which differs from the sigmoid form of nonlinearity realized by dendritic spiking (Text S1, [3],[4],[45]). We believe, however, that at this level of abstraction the exact form of nonlinearity does not affect our results: As that is the difference between the dendritic responses to learned and not learned patterns that influence the somatic detection of dendritic events, a sigmoid integration function give qualitatively similar results (Text S2). Moreover, we studied only passive interactions between individual dendritic events as the effect of voltage and calcium dependent currents (including A-type and Ca2+-dependent potassium [48] and the H-current [49]) regulating the propagation of dendritic spikes were not included in the model. Future studies using a compartmental model equipped with dendritic spiking could support our results and clarify further details.
Our analysis has revealed that a moderately branched dendritic tree is optimal for the independent branches model, and we have shown that this mechanism could contribute to the spatial firing properties of granule cells in the DG. The dendritic tree of cerebellar Purkinje cells as well as the apical dendrites of hippocampal and neocortical pyramidal cells is typically larger, and more ramifying [50]. Their morphology is suitable for local plasticity within single branches [6],[8], and although it seems that individual branches may function as single integrative compartments [3],[4],[51],[52], dendritic spikes localized to these compartments fail to propagate to the soma and directly influence the neuron's output [53]. Larger dendritic events, active spread of dendritic spikes towards the soma or interactions among dendritic subunits could contribute to the generation of somatic action potentials in this case. The dendritic tree of pyramidal neurons is, however, far more complex than that of granule cells: it has several morphological and functional subregions with different afferent inputs and membrane excitability [50]. Understanding how their spatial firing characteristics arise from their cellular properties would require at least a different model structure and is beyond the scope of this paper.
Whether individual dendritic events influence the output of the neuron depends - beyond the structure of the dendritic tree - on the size and the frequency of the large dendritic events and the output sparsity. The size of the events depends on the exact form of the dendritic integration function and the plasticity rule while the input statistics determine the frequency of such events. We have shown that given the sparseness of the output, sufficiently large, localized dendritic events arriving with appropriate frequency are able to separately determine the output of the neuron. Whether a local event is sufficiently large depends on the geometry of the dendritic tree: A smaller event may be sufficient if there are only a few subunits, or if the events actively propagate to a large part of the entire dendritic tree (e.g, the apical tuft in pyramidal neurons, [54]). Conversely, in neurons such as cerebellar Purkinje cells with large, ramifying dendritic tree, where individual events are localized to small branches, very large dendritic spikes would be required to influence the output. Indeed, detailed compartmental modelling of dendritic morphology revealed that the forward propagation of the action potential initiated in the apical trunk of pyramidal neurons was very effective, while in Purkinje cells dendritic action potentials were rapidly attenuated [53].
Clustered plasticity allows the neuron to simultaneously learn several different patterns but requires the electrical and/or biochemical isolation of the dendritic compartments [47],[55]. However, the intracellular resistance () in dentate granule cells is relatively low and granule cells are usually regarded as electrically compact neurons [28]. Indeed, signal propagation from somata into dendrites in vitro is more efficient in granule cells compared with CA1 pyramidal cells and distal synaptic inputs from entorhinal fibers can efficiently depolarize the somatic membrane of granule cells [28]. However, in vitro studies do not take into account that neurons are embedded in a network of spontaneously active cells. As thousands of synapses bombard the dendritic tree in vivo, the dendritic membrane becomes “leakier” and, consequently, the membrane's space constant decreases significantly [56]. Moreover perisomatic inhibition [57] and feed-back excitation (via hilar mossy cells [40]) further decrease the resistance of the proximal membrane contributing to the separation of the somatic and dendritic compartments [54],[58]. More specifically, we predict, that the membrane resistance of granule cells is considerably smaller at the perisomatic region than in the distal dendrites. Indeed, computational studies predict a 7–30 fold increase in the somatic leak conductance due to the synaptic background activity [59]. On the other hand, large space constant at long terminal branches facilitate interactions among synapses distributed on the same branch. Therefore the long dendritic branches of dentate granule cells may act as single integrative computational subunits, separated from each other by the perisomatic region of the cell. Furthermore, in the present paper we used steady-state approximations and we neglected temporal characteristics of the input and the integration. For rapidly varying inputs the coupling between dendritic sites and the soma is much smaller than for slowly varying currents since the distributed capacitance throughout the tree will absorb the charge before it reaches the soma [14]. Therefore dendritic compartments in a passive tree are more isolated for transient events such as dendritic spikes than for steady-state current. Finally, biochemical compartmentalization is likely to play a substantial role in the cooperative induction of LTP in both hippocampal [60] and neocortical neurons [7].
If, on the other hand, dendritic branches are not isolated during the learning process and synapses across the whole dendritic tree are modified simultaneously then different dendritic branches will be sensitive for different component (modalities) of the same episode. A new episode with partial overlap with the previously learned one may trigger dendritic spiking in the corresponding dendritic branch. As the somatic detection probability of dendritic spikes does not depend on the degree of electrical isolation (Figure 4), individual branches trigger somatic spiking, and, in this way the dentate gyrus contributes to the associative recall of the previously encoded episode in the hippocampus.
Since the first description of LTP at perforant path - granule cell synapses [61] synaptic plasticity has become widely accepted as the physiological basis of memory [62]. As Hebbian plasticity is intrinsically unstable, simply because it is a positive feed-back mechanism multiple stability-promoting mechanisms have been proposed, including heterosynaptic depression [31],[63]. Indeed, in the present model synaptic plasticity results in an average decrease of synaptic strengths (Figure 3A), which have several functional consequences: First, as the dendritic response to untrained patterns and likewise the baseline activation of the cell decreases during training, the somatic detection of individual, large dendritic events becomes easier. Consequently, feature detection is less efficient in semi-trained neurons where synaptic weights at only a part of the dendritic tree has already been modified due to the learning precess. Therefore, in this model, appropriate training of each dendritic branch is required for proper functioning. Second, increased excitability stimulates learning in naive branches, while decreased responsiveness of previously trained branches prevents overlearning. Indeed, newly generated granule cells are more excitable than the neighboring old neurons [36], and they are preferentially incorporated into functional networks in the dentate gyrus during acquisition of new memories [64].
One of the most interesting prediction of the present model is how the number of presynaptic spikes required for the postsynaptic induction of dendritic spiking changes during the course of learning. We can calculate this by dividing the total input U needed for dendritic spiking with the mean synaptic weight parameter (μw) before and after learning. Our model predicts, that while in young neurons the simultaneous occurrence of ≈70–80 presynaptic spikes (randomly distributed across the presynaptic neurons) would trigger a postsynaptic dendritic spike, after learning (i.e., in matured neurons)≈130–160 would be required.
A recent study showed that the homeostatic regulation of the neurotransmitter release probability at neighbouring synapses depends on the local dendritic activity [10]: Increased dendritic depolarization elicits a local homeostatic decrease in the release probability and vice versa. This mechanism may also prevent overlearning in trained branches where dendritic spikes has sufficiently high rate by reducing the excitability of that branch. On the other hand the same mechanism may stimulate learning new patterns in naive or disused branches where dendritic spikes are not present. One of the key elements of our model was the local nature of the synaptic plasticity, i.e., the change of the synaptic weights was controlled by the local dendritic but not the somatic activity [6]–[8]. Specifically, in hippocampal granule cells the induction of LTP was shown to be independent of the discharge of the neurons during the high-frequency stimulation [65]. Our model predicts that, if the postsynaptic signal for synaptic plasticity is localized to individual dendritic branches than, due to the associative nature of the LTP, the synapses from entorhinal cells with overlapping firing become potentiated. If LTP is accompanied by structural remodeling, than the entorhinal neurons with overlapping place fields project to the same dendritic branches of granule cells as also proposed by Hayman and Jeffery (2008) [66].
The variation in the strength of perforant path-granule cell synapses was found to be critical in the generation of multiple place fields in a recent modelling study [67]. This heterogeneity caused a greater average synaptic excitation in a fraction of granule cells. This extra excitation therefore selects the subpopulation of neurons active within a given environment similar to the proposed role of contextual inputs in the model of Si and Treves [68]. One possible source of synaptic heterogeneity is synaptic plasticity [69] which was also crucial in the present model to amplify the local responses to learned patterns.
Hippocampal granule cells receive afferent fibers from the medial and the lateral portion of the entorhinal cortex, and these two pathways differ both in their pattern of termination [70],[71] and information content [72]. Fibers originating in the lateral EC display weak spatial selectivity and terminate on the most distal branches of granule cells, while medial entorhinal neurons innervate the middle third of their dendritic tree and show strong spatial selectivity [72],[73]. It has been recently suggested by modelling studies [66],[68] that inputs originating from the lateral EC conveys contextual information to granule cells. In these models the contextual input select a subpopulation of neurons (or dendritic branches in [66]) that can be activated within the given context (environment) while medial entorhinal fibers determine the exact location of the place fields. The selection of a subpopulation by contextual inputs can also contribute to the multiple firing fields of granule cells by reducing the number of available neurons within the given environment [67],[68]. However, the spatial distribution of the individual place fields become regular (grid-like) if the multiple firing peaks are the consequence of an incomplete competition between neurons, especially if the input grid cells are organized into a finite number of ensembles [74],[75].
In the present paper we have shown that synapses, irrespective of their origin, arriving at different branches of hippocampal granule cells can be modified at different spatial locations. We have also shown, that in granule cells each dendritic branch is able to activate the neuron, therefore each subfield on the cell's multi-peaked activity map corresponds to a dendritic place field. The segregation of contextual and positional information could explain the sensitivity of the subfields to contextual manipulations [26],[66] and is consistent with the role of DG in context discrimination [76].
Along with the laminar organization of excitatory input, different interneurons innervate different dendritic domains of granule cells [39],[77]. It appears, that distinct types of interneurons have evolved to selectively and locally modulate the computations performed by the postsynaptic membrane [57],[78]. According to our model, basket and axo-axonic cells may continually adjust the inhibitory drive such that the mean activity of the population remains nearly constant; HICAP cells, targeting the proximal dendritic domain of granule cells together with the excitatory mossy cells [40] may increase electrical isolation of distal dendritic regions by raising the conductance of the proximal membrane; whereas MOPP and HIPP cells associated with the entorhinal afferents may contribute to the de-inactivation of calcium channels required to dendritic spiking by providing rhythmic hyperpolarization to distal dendritic branches.
Hippocampal interneurons have also a substantial role in shaping the temporal dynamics of the network [78]. The firing of neurons in the hippocampal formation is strongly modulated by the theta rhythm [25],[79],[80] which is a prominent, large amplitude field potential oscillation in the rodent hippocampus during exploratory behavior [81]. The relative synchronization of presynaptic spikes by the theta rhythm allows the temporal integration of their postsynaptic potentials despite the relatively small time constant of granule cells' membrane [28]. Moreover, the synchronization of synaptic inputs can also influence the form of dendritic integration by switching from linear to nonlinear integration [30]. Extending the present model with temporal dynamics could be an exciting direction for future research.
What is the additional computational power gained from the present model? We argue, that smaller and uncorrelated place fields may help pattern separation in the dentate gyrus. Theoretical considerations suggest that the DG helps the hippocampal storage of new episodes by producing sparse representations via competitive learning [82],[83]. It was demonstrated by modelling studies that competitive learning on spatially organized input results in the formation of place fields [68],[74],[84],[85] that is a sparse and orthogonal representation of the input space. In the present paper we proposed that parallel dendritic computations explain the formation of multiple, independent place fields of hippocampal granule cells even within a relatively small environment [24],[26].
Pattern separation by the DG can be more efficient if granule cells have multiple, irregularly placed fields and the individual fields are smaller. The neural representation of neighbouring locations is more similar if neurons have one, larger field than if they have several but smaller fields (Text S3). In our model the place fields of a dendritic branches are analogous to the to the single place field of an electrically compact neuron. The multi-peaked somatic firing of the granule cells mirrors the several dendritic fields of the same neuron. We argue, that if the size of the somatic firing fields is limited by competition between simultaneously active neurons [86], then the place fields of granule cells could be smaller than the corresponding dendritic fields. If the individual place fields of granule cells become smaller, than the neural representation of adjacent places becomes less correlated which further increase the pattern separation ability of the DG. Therefore independent dendritic subunits increase the computational power of the DG while keeping the number of cells and their sparsity constant. Moreover, clustering of different inputs into different dendritic domains could explain the remapping of hippocampal place cells under several experimental conditions [26],[66].
The impact of both dendritic nonlinearity and clustered plasticity on the computational power of neurons was rarely addressed by modeling studies. Poirazi and Mel [16] predicted, that nonlinear dendritic integration with local (structural) plasticity rule increase the representational capacity of neural tissue. They showed on binary input, that the number of attainable input-output functions (representational capacity) is maximal if the neuron has many, relatively short branches, and the performance of the model in a linear classification task correlates remarkably well with the logarithm of representational capacity. However, in order to approach the combinatorial bound of the representational capacity in a neural tissue and to amplify slight differences in the input extremely large subunit nonlinearity was required (they used F(U) = U10). In the present study we showed that a moderate increase in the memory-capacity can be achieved with local, Hebbian learning rule and slightly supra-linear dendritic integration. We emphasized that under certain conditions a single branch is able to evoke somatic output. However, if the amplitude of the individual events is smaller, a larger spatial extent involving the depolarization of additional branches will be required to trigger output spiking. This mechanism could induce a combinatorial increase in the representational capacity as shown by [16].
According to our model hippocampal granule cells can be regarded as a two layer neural network of abstract integrate and fire elements: In the first layer corresponding to the terminal branches the units integrate separately their inputs and they innervate a common output unit (second layer, the somatic compartment) that implements a logical OR computation. The idea that a dendritic tree may perform logical computations was originally proposed by [87] to explain directional selectivity of retinal ganglion cells. Shepherd and Brayton [88] further elaborated this approach but instead of branches they used dendritic spines as basic computational subunits. Our approach is more similar to how Poirazi et al. [89] describe hippocampal pyramidal cells, however, in that model the output unit performs (nonlinear) summation prior to final thresholding. Another similar model was proposed by Gasparini and Magee [30], in a paper where they showed that the apical trunk of hippocampal pyramidal neurons integrate spatially clustered and synchronously arriving synaptic inputs nonlinearly, whereas distributed or asynchronous inputs are linearly integrated. They suggest that processing in the nonlinear mode could functionally separate the dendritic arbor into a large number of independent nonlinear computational units, each sending its own output to the soma. In the present paper, we showed that a single computational units is powerful enough to determine the output of the neuron only if there are not too much similar units (N<100) and if the local integration is sufficiently nonlinear.
A similar picture emerged form a recent series of in vitro experiments performed on the basal dendrites of neocortical pyramidal neurons: These branches behave as independent computational subunits as nearby inputs on the same branch summed sigmoidally due to the presence of local NMDA spikes [2],[45],[90] and synaptic plasticity required the pairing of local NMDA spikes with biochemical signals [7]. Moreover, an NMDA-spike localized to a single basal dendrite could efficiently induce somatic UP-state like depolarization accompanied by bursts of action potentials [5]. These results suggest that our model describes remarkably well the neuronal computations performed by the basal dendritic tree of pyramidal neurons.
Although we tried to fit our model to the available experimental data we had to make some assumptions regarding the integration of neighbouring inputs in dentate granule cells. Moreover, based on the model described in the present paper we make some explicit predictions. Both the assumptions and the predictions of our model should be tested experimentally.
We used the data from [28] to estimate the passive membrane parameters of the granule cells the DG (Table 1). First we computed the membrane area of a single branch (Adend) falling into the perforant path termination zone (the outer two third of the dendritic tree):(12)where ld is the total length of the dendritic tree, db = 1.1 µm is the average diameter of a single branch, N is the number of branches and α = 1.9 is a correction factor for the membrane area of dendritic spines.
Similarly, the area of the somatic compartment (Asoma), assuming a sphere with diameter ds:(13)The area of the cross section of a single branch is , and the length of the proximal third of the branches, that do not receive input from the entorhinal cortex is lds = 50 µm. Finally, we estimate the parameters in Eqs. 1–2:(14)(15)(16)where Rm and Ri are the membrane resistance and the intracellular resistivity, respectively. As the somatic and the dendritic membrane area (and hence the resistance) were similar, we used that . The parameter R used in our calculations was R = Ra/Rm0.01 for a passive granule cell in the DG. Note that due to the synaptic conductances activated in vivo the membrane resistances of functioning granule cells are certainly lower than its in vitro estimates [59].
A single dentate granule cell receive synaptic input from nEC–DG≈2500–4000 entorhinal layer II cells distributed on N≈25–40 branches, whereas a single branch receives M≈100 synapses in the rat's hippocampus [28]. According to Amaral and Lavenex [71], there are nDG≈1.2 · 106 granule cells in the rat's DG, and nEC≈0.11 · 106 projection cells in the layer II of the entorhinal cortex. It is known, that a given location in the hippocampus may receive inputs from more than 25% of the dorsomedial-to-ventrolateral axis of the medial entorhinal cortex [94],[95]. Therefore, while a single dendritic branch get its M≈100 synaptic inputs randomly from nearly 25000 entorhinal cortical neuron, we assume that each synapse on a dendritic branch comes from different entorhinal neurons.
By electrical recordings from different hippocampal regions one can estimate the proportion of simultaneously active cells within a reasonable time window. We call this number the sparseness of the representation in the given area. Specifically, 1–5% of the granule cells are active simultaneously in the DG [24],[29], therefore we used spDG = 0.05. The sparseness of the entorhinal input is somewhat larger, spEC = 0.2 [27],[80],[96].
Experimental data provide a good estimate for the mean firing rate of these neurons, however, they give the variance of the mean across neurons, but not the variance in the firing rate of individual cells. To estimate the variance in the firing rate of an individual cell, we generated random spike trains based on the ISI histogram on Figure 5 of [80]. The expected value and the variance of the number of spikes in a 100 ms time bin (corresponding to one period of the hippocampal theta rhythm) was μEC′ = 0.32 and and there was at most 4 spikes during 100 ms in the case of an entorhinal excitatory cell. We scaled these values relative to the maximal firing rate, so we had μEC = μEC′/4 = 0.08 and characterizing the distribution of the presynaptic firing uj. Possible differences in firing statistics across different (medial-lateral or dorsal-ventral) regions in the EC and across individual neurons are neglected here.
Next, we start with originally equal synaptic weights, wij = w = 3 · μEC. In this case, if we assume that the firing of entorhinal neurons are independent and identically distributed, we can approximate the total input to a branch with a Gaussian distribution:(17)where and . The distribution of the total input U is shown on Figure 1C.
Learning alters the distribution of the total input Ui = Σjwijuj of dendritic branches (Eq. 17) by modifying synaptic weights. From Eq. 11 used to describe synaptic plasticity, we can see that synaptic weights converge to a fixed point wij = uj whenever the activity of the postsynaptic branch i is above threshold βd. In the stationary state, the weight vector reflects a presynaptic firing pattern . In other words, the learned presynaptic firing pattern is stored in the corresponding synaptic weights.
In order to stimulate initial plasticity in naive branches and prevent learning in those branches that have already learned a pattern, we initialized the synaptic weights to wij = w = 3μEC, which is higher than their expected value at the fixed point (μEC). This initialization ensured that the response (U) to unlearned inputs decrease during the process of learning, and prevented interference in branches that already have learned a specific pattern. Indeed, synaptic plasticity is enhanced in newly generated granule cells of the hippocampus compared with mature neurons already integrated into functional circuits [36],[64],[97].
After learning we can approximate the distribution of the total synaptic input U to a branch by the sum of two Gaussians representing the total input in the case of learned () and not learned patterns (), respectively:(18)where pl (pn) is the probability that one of the branches receive a learned (not learned) input, and μl and (μn and ) are the mean and the variance of the response to learned (not learned) inputs. If we have a finite number (NS) of different inputs, and each branch learns one of them, then(19)Distribution of the total input to the dendritic branches before and after learning is shown on Figure 3A. Parameters μn = 1, σn = 0.39, μl = 5.6 and σl = 0.58 were estimated numerically based on the reconstructed firing characteristics of entorhinal neurons. We assumed that each branches learned one of the samples and the probability that one of the branches receive its learned input (N/NS) was the sparseness in the DG (spDG≈0.05. Note, that the distribution of is the theoretical distribution of the responses to learned inputs, from which each branch draw only a few (perhaps one) sample because learning is very sparse.
We recalculated the two functions H(U) and K(U*) with the new input distributions by replacing μ and σ with μn and σn in Equations 9–10 and 27–28, and by changing the distribution of U in Eq. 29 from Eq. 17 to Eq. 18. In these calculations, we neglected the possibility that two (or more) branches may both get their learned input at the same time. Finally, we determined the firing threshold by solving the following integral to β (see Eq. 24–25):(20)
In the case of continuous variables we can write that H(Ui) = H(U). The function H(U) has the form:(21)The conditional probability has a form similar to Eq. 8, except that we have only N−1 random variables from the Gaussian distribution of U (Eq. 17) with parameters μF and , therefore we can write that:(22)We can compute the second function K(U*) as follows:(23)(24)(25)where is the conditional distribution of the somatic activation as and the maximal dendritic input U*. The distribution is similar to the distribution of in Eq. 8 with two important differences: First, we have only N−1 random variables. Second, we know that U<U*, therefore the distribution of the inputs to other branches is different from the Gaussian in Eq. 17. Hence we can write, that(26)where and are the conditional expectation and variance of the distribution p[F(U)|U<U*]. We calculate and by integrating Equations 9–10 from −∞ to U*:(27)(28)where is a normalization factor. Finally we calculate the last term of Eq. 24, the distribution of U* as follows:(29)where P(U) is the cumulative distribution function (CDF) of U and [X]′ marks derivation. The intuition behind Equation 29 is that: First, P(U) is the probability that a given input is smaller than U. Second, P(U)N is the probability that all inputs are smaller than U, also the (CDF) of U*. Third, its derivative [P(U)N]′ gives us the probability density function (PDF) of U*. The PDF of U is a Gaussian function, its CDF can be expressed with the Gauss error function (erf{}).
To calculate the dependence of the dendritic activation on the inputs, we first repeat Eq. 6:(30)Next, we substitute in Equation 30 with Eq. 7:(31)The two terms of the sum in Eq. 31 are independent, because Ui is independent from Ujs, therefore we can calculate the distribution of by the convolution of two distributions (corresponding to the two terms in the sum). The second term in Eq. 31 is the sum of independent random variables and we approximate it with a Gaussian (similarly as we did it for previously, Eq. 8). The distribution of Ui is a Gaussian (Eq. 17), that we can transform into the first term of Eq. 31 by a Jacobian factor [98]:(32)where V = F(U). We get the distribution by substituting the first term of Eq. 31 by a Dirac delta distribution. Similarly, we can calculate by first computing a conditional sum in the second term (Uj+Σk≠{i,j} Uk) as described by Eq. 22 and then performing the convolution.
The R software environment [99] was used to analyze the data and to prepare the figures.
|
10.1371/journal.pbio.1001525 | White-Opaque Switching in Natural MTLa/α Isolates of Candida albicans: Evolutionary Implications for Roles in Host Adaptation, Pathogenesis, and Sex | Phenotypic transitions play critical roles in host adaptation, virulence, and sexual reproduction in pathogenic fungi. A minority of natural isolates of Candida albicans, which are homozygous at the mating type locus (MTL, a/a or α/α), are known to be able to switch between two distinct cell types: white and opaque. It is puzzling that white-opaque switching has never been observed in the majority of natural C. albicans strains that have heterozygous MTL genotypes (a/α), given that they contain all of the opaque-specific genes essential for switching. Here we report the discovery of white-opaque switching in a number of natural a/α strains of C. albicans under a condition mimicking aspects of the host environment. The optimal condition for white-to-opaque switching in a/α strains of C. albicans is to use N-acetylglucosamine (GlcNAc) as the sole carbon source and to incubate the cells in 5% CO2. Although the induction of white-to-opaque switching in a/α strains of C. albicans is not as robust as in MTL homozygotes in response to GlcNAc and CO2, opaque cells of a/α strains exhibit similar features of cellular and colony morphology to their MTL homozygous counterparts. Like MTL homozygotes, white and opaque cells of a/α strains differ in their behavior in different mouse infection models. We have further demonstrated that the transcriptional regulators Rfg1, Brg1, and Efg1 are involved in the regulation of white-to-opaque switching in a/α strains. We propose that the integration of multiple environmental cues and the activation and inactivation of a set of transcriptional regulators controls the expression of the master switching regulator WOR1, which determines the final fate of the cell type in C. albicans. Our discovery of white-opaque switching in the majority of natural a/α strains of C. albicans emphasizes its widespread nature and importance in host adaptation, pathogenesis, and parasexual reproduction.
| Phenotypic transitions enable fungal pathogens to better adapt to their ever-changing environments. Approximately 10% of natural Candida albicans strains, which are homozygous at the mating type locus (MTL, a/a and α/α), can switch between two distinguishable morphological forms: white and opaque. The two cell types differ in a number of biological aspects including virulence, susceptibility to host immune attacks, and mating competency. Here, we demonstrate that white-opaque switching competency is not restricted to the MTL homozygous strains, but is a general characteristic of all MTL strain types of C. albicans (a/a, α/α, and a/α). Two host environmental cues, N-acetylglucosamine and CO2, promote white-to-opaque switching and stabilize the opaque phenotype. Thus, although switching is normally blocked in a/α cells, this block can be overcome through specific environmental changes. We further show that three transcriptional regulators (Rfg1, Brg1, and Efg1) help to regulate white-opaque switching in MTL heterozygotes of C. albicans. This study generalizes white-opaque switching to strains with all mating-type configurations and emphasizes its importance in host adaptation, pathogenesis, and parasexual reproduction.
| Phenotypic plasticity is critical for microorganisms to survive under fluctuating environments. For fungal pathogens, phenotypic switching is a common strategy to rapidly adapt to different host niches and facilitate colonization and infection [1]. A specific phenotype can also confer the fungus a growth advantage over competing microorganisms in a specific environment or host niche. Candida albicans, the major causative agent of fungal infections in humans, can switch between two different visible cell types: white and opaque [2]. The two cell types differ in a number of biological aspects including morphology, virulence, and mating competence [3]–[5]. White cells are small and round and form “white,” dome-shaped colonies on solid media, while opaque cells are large and elongated and form darker and flatter colonies [6]. White cells are more virulent than opaque cells in systemic infections, whereas opaque cells appear more suited to cutaneous colonization [7],[8]. Opaque cells possess pimples on the cell wall and exhibit unique antigenicity, which may help the pathogen in evading the host immune system [3]–[5]. Moreover, opaque cells are significantly less susceptible to phagocytosis by cells of the fly and mouse innate immune systems than white cells [9]. Perhaps the best studied feature of opaque cells is their mating competency. Opaque cells mate ∼106 times more efficiently than white cells [10]. It has recently been shown that Candida tropicalis, another important human fungal pathogen, can also undergo white-opaque switching and parasexual mating [11],[12].
Despite the importance of white-opaque switching in host adaptation, pathogenesis, and parasexual reproduction in C. albicans, only a minority (<10%) of natural strains have been reported to undergo white-opaque switching in vitro [13]. It has been shown that the mating-type locus homeodomain proteins (MTLa1/α2) inhibit white-opaque switching via controlling the expression of the master regulator WOR1 [14]–[16]. Consistent with this, the minority of natural isolates capable of white-opaque switching in vitro are homozygous at the MTL locus; this relieves the block of the mating locus proteins. The majority (>90%) of C. albicans MTLa/α isolates in nature were thought to be incapable of white-opaque switching unless they underwent homozygosis of the mating type locus [13]. These ideas raised a fundamental question. Given the importance of white-opaque switching in host adaptation and pathogenesis, why do the majority of a/α natural isolates of C. albicans not undergo white-opaque switching unless they undergo a genetic rearrangement?
In this study, we provide reasonable answers to this basic question. We show that naturally occurring a/α isolates of C. albicans can indeed undergo white-opaque switching under a specialized set of environmental conditions. Previous studies typically used glucose as the sole carbon source and grew cells in ambient CO2. We show here that a/α strains can undergo white-opaque switching in 5% CO2 when N-acetylglucosamine (GlcNAc) is used as the sole carbon source. GlcNAc and CO2, primarily produced by bacterial commensals, are abundant in the gut and have synergistic effects on the induction of the opaque cell phenotype in C. albicans [17]. Therefore, this culture condition likely mimics certain aspects of host niches, such as those in the gut. Opaque cells of a/α strains exhibit similar phenotypes of typical MTL homozygous opaque cells, except they lack the ability to mate. Further experiments demonstrate that three transcription factors Rfg1, Brg1, and Efg1 are involved in the regulation of white-opaque switching in a/α C. albicans strains. This study indicates that there is an alternative gene circuit, which can bypass the a1/α2 block to switching, and promote white-opaque switching in C. albicans under certain environmental conditions that are reminiscent of niches of the host. We propose that white-opaque switching is not limited to the minority of MTL homozygotes, but rather is a general characteristic of natural C. albicans strains.
There are three MTL types of natural C. albicans isolates (a/a, α/α, and a/α). Under normal conditions, MTL heterozygotes (a/α) are blocked for switching and “locked” in the white phase in vitro [10],[13]. Since a/α strains are more competitive than their a/a or α/α derivatives (at least in some in vivo assays) and carry the entire set of opaque-specific genes essential for switching [18], we suspected that the a/α isolates of C. albicans could also undergo white-opaque switching in their natural niches. We also reasoned that routine laboratory media and culture conditions were totally different from conditions in natural niches and might not be conducive for the transition in a/α strains of C. albicans. To test our hypothesis, we took advantage of the synergistic effects of two host environmental cues, GlcNAc and CO2, on the induction of the opaque cell phenotype [17]. We grew 94 natural isolates of C. albicans on Lee's GlcNAc medium in 5% CO2. We found that 34 strains (36%) formed opaque colonies under this condition. We then examined the MTL genotype of all 94 tested strains. Of them, 92 were a/α, one was a/a, and one was α/α. The two MTL homozygotes (one a/a and one α/α) were identified as switching to opaque, along with the 32 a/α strains in the switchable strain list (Table S1).
An example of an a/α clinical strain that could undergo white-opaque switching (SZ306) is shown in Figure 1. We noticed that SZ306 could also form opaque colonies on rich medium (YPD) when cultured for an extended time period (Figure 1A); some other a/α strains also exhibited this behavior. The white and opaque cells of a/α strains were similar to their counterparts of MTL homozygotes in the size and shape of cells (Figure 1B): white cells of a/α strains were small and round with no pimples on their cell wall surface, while opaque cells were elongated and possessed obvious opaque-specific pimples (Figure 1C). Northern blot analysis demonstrated that two opaque-enriched genes, OP4 and the master regulator WOR1, were expressed in opaque cells of a/α strains but not in white cells (Figure 1D). Conversely, the expression levels of the white-enriched genes WH11, EFG1, and RFG1 were significantly higher in white cells than in opaque cells of a/α strains. These results suggest that opaque cells of a/α strains exhibit similar characteristics of colony and cellular morphology and gene expression profile to the opaque cells of MTL homozygotes.
To exclude the possibility of homozygosis of a/α cells during growth, we re-plated several opaque colonies of each switchable a/α strain onto Lee's GlcNAc medium and incubated them in ambient CO2 for 5 days. Three single opaque colonies of each re-plated culture were examined for the MTL configuration, and we verified that all remained heterozygous at the MTL locus. An example of this analysis is given in Figure 1E. These results demonstrate that C. albicans a/α isolates can indeed undergo white-opaque switching.
Additional examples of white-opaque switching in natural a/α strains of C. albicans are shown in Figure S1 and Table S1. The white colonies of different a/α strains showed variability in their abilities to filament on Lee's GlcNAc medium in 5% CO2 at 25°C, indicating that the white-opaque switchable strains are genetically diverse and probably not derived from a single strain with a specific genetic background. To characterize the genetic background of these natural strains, we sequenced their CAI microsatellite loci by using a reported assay [19]. As shown in Table S1 (Column D), these strains exhibited several distinct patterns of the CAI genotype, demonstrating their genetic diversity.
The strains listed in Table S1 were all isolated in China. To exclude the possibility of geographical specificity, we tested the white-to-opaque switching ability in 29 clinical strains of C. albicans isolated from different countries. These strains, which were demonstrated incapable of switching on glucose-containing media, were all originally heterozygous at the MTL locus (a/α) and belonged to five different genetic clades [13]. We found that 15 of them (52%) underwent the white-to-opaque transition on Lee's GlcNAc medium in 5% CO2 at 25°C (Table S2). Two opaque colonies of each switchable strain were examined for the MTL configurations. Twelve of the 15 strains were a/α heterozygotes, two (P75010, P22095) α/α, and one a/a (P78042, perhaps due to spontaneous loss of the MTLα locus, Pujol and Soll, unpublished data). These results further indicate that the white-opaque switchable a/α strains of C. albicans are genetically and geographically diverse.
White-opaque switching and mating are two coupled biological processes that are both controlled by the MTLa1/α2 complex in C. albicans [10]. One possibility that could explain how a/α isolates could undergo white-opaque switching is that the a1/α2 complex might not function properly; thus, cells could behave as though they were a or α cells. Although our DNA sequencing analysis showed that the MTL locus of the switchable a/α isolates were normal and with no obvious defects, the expression of MTLa1 or MTLα2 could, in principle, be defective. To exclude this possibility, we performed a mating experiment with opaque cells from three independent a/α strains. As shown in Figure 2A, these cells showed no mating response, whereas a/a and α/α opaque cell controls mated normally (Figure 2Ba, b, c, e, f, and g). Quantitative mating assay demonstrated that the mating efficiencies of the MTLa/α x MTLa or α cells crosses were undetectable (<1×10−7). The mating efficiency of the MTL a/Δ x α/α cross-control was (2.3±0.8)×10−2, at least 1×105 times higher than that of the MTLa/α crosses (Figure 2C). These results demonstrate that a/α opaque cells cannot mate with either a/a or α/α opaque cells, suggesting that the white-opaque switchable a/α strains are mating-incompetent. However, once the opaque cells of these a/α strains were converted to a/Δ or Δ/α strains by deletion of one allele of the MTL locus, they acquired mating competence and mated as efficiently as the WT a/a or α/α controls (Figure 2Ah). We conclude from these experiments that the a1/α2 complex is functional in the regulation of mating, and that the white-opaque switching in these strains is not due to the inactivation of a1 or α2 proteins.
As described in the introduction, GlcNAc and CO2 are two potent inducers of white-to-opaque switching and are believed to be characteristic of host niches such as the gastrointestinal (GI) tract [17]. As shown in Figure 3, the frequencies of white-to-opaque switching in the a/α strain SZ306 was extremely low on Lee's glucose (<0.6%) or GlcNAc (0.5%) medium in ambient CO2. CO2 alone also had little effect on the induction of opaque phenotype on Lee's glucose medium in this a/α strain (switching frequency<0.4%). However, the switching frequency of white-to-opaque in SZ306 was increased to 7.5±3.1% when cultured on Lee's GlcNAc medium in 5% CO2, indicating that GlcNAc and CO2 had a synergistic effect on the induction of the opaque cell phenotype. To compare the switching features of a/α strains and MTL homozygous “a” or “α” strains, we converted SZ306 (a/α) to an MTLa/Δ strain, namely SZ306a, and RVVC10 (a/α) to an MTLΔ/α strain, namely RVVC10α, by deletion of one allele of the MTL locus. As shown in Figure 3, although the frequency of white-to-opaque switching in SZ306a was only 0.4% on Lee's glucose medium in ambient CO2, GlcNAc, or 5% CO2 alone increased the switching frequencies to 3.0±2.7% and 34.4±0.9%, respectively. Notably, SZ306a underwent a mass conversion (switching frequency = 100%) on Lee's GlcNAc medium in 5% CO2, consistent with our previous study of the synergistic effect of GlcNAc and CO2 on white-to-opaque switching in MTL homozygotes [17]. As in SZ306 and SZ306a, GlcNAc and CO2 had a similar effect on the induction of the opaque cell phenotype in RVVC10 and its derivative, RVVC10α (unpublished data). These results indicate that a/α strains are less sensitive than their “a/Δ” or “Δ/α” derivatives to GlcNAc and CO2, but that white-to-opaque switching is stimulated by GlcNAc and CO2 in all three MTL configurations.
GlcNAc and CO2 can also stabilize the opaque phenotype in MTL homozygotes of C. albicans. We next tested whether this was also the case in heterozygous a/α strains. As shown in Figure S2 and Table S3, the opaque phenotype of a/α strains was extremely unstable in Lee's glucose medium when cultured in ambient CO2 at 25°C (switching frequency to white was 100%). The switching frequencies were 38.6±7.7, 34.6±5.3, and 18.7±7.1 on Lee's GlcNAc medium in ambient CO2, on Lee's glucose in 5% CO2, and on Lee's GlcNAc in 5% CO2, respectively (Figure S2 and Table S3). These results suggest that GlcNAc and CO2 stabilize the opaque phenotype of a/α strains. For the “a/Δ” and “Δ/α” strains, SZ306a and RVVC10α, the opaque phenotype was very stable on both Lee's glucose and GlcNAc, irrespective of whether the cells were cultured in air or 5% CO2. Under the four conditions tested (Lee's glucose in air, Lee's GlcNAc in air, Lee's glucose in 5% CO2, and Lee's GlcNAc in 5% CO2), the opaque-to-white switching frequencies of SZ306a and RVVC10α were all less than 1% (Figure S2 and Table S3). We sequenced the WOR1 promoter of several switchable a/α strains and found what is believed to be the major a1/α2 cis-regulatory sequence site was intact. These results indicate that although the a1/α2 complex does not provide an absolute block to white-to-opaque switching in these a/α strains, it reduces switching to favor white cells, likely by turning down (but not off) the expression of WOR1.
Since the physiological temperature of human hosts is 37°C, we therefore examined whether MTLa/α strains can undergo white-to-opaque switching under this temperature. White cells of CY110 and RVVC10 (two MTLa/α strains) were plated onto Lee's glucose and Lee's GlcNAc medium plates and cultured at 37°C for 3 to 4 days. The cells of both strains were locked in white phase on Lee's glucose medium in air or in 5% CO2, whereas they formed opaque, opaque-sectored, or mixed colonies on Lee's GlcNAc medium (Figure 4A and 4B). The switching frequencies of CY110 and RVVC10 on Lee's GlcNAc medium in 5% CO2 were as high as 60.6±10.3% and 100% (mass conversion), respectively. The cellular morphologies demonstrated that opaque or mixed colonies contained typical opaque cells (Figure 4A and 4B). WOR1 is an opaque phase-specific gene, while WH11 and EFG1 are white phase-specific genes [5]. To further verify their cell identities, we constructed WOR1, WH11, and EFG1 promoters-controlled GFP reporter strains in the MTLa/α strain CY110. As shown in Figure 4C, GFP fluorescence was only observed in opaque cells of the WOR1/WOR1::WOR1p-GFP strain, but not in opaque cells of the EFG1/EFG1::EFG1p-GFP and WH11/WH11::WH11p-GFP strains. As expected, GFP fluorescence was observed in white cells of the EFG1/EFG1::EFG1p-GFP and WH11/WH11::WH11p-GFP strains. These results indicated that the opaque cells formed at 37°C were genetically opaque.
We next tested the stability of opaque cells of MTLa/α strains under host physiological temperature. Opaque cells of three a/α strains (SZ306, RVVC10, and CY110) were plated onto Lee's glucose and Lee's GlcNAc medium and incubated in air at 37°C for 3 days. On Lee's glucose medium, opaque cells underwent a mass conversion to the white cell phase (switching frequency = 100%). On Lee's GlcNAc medium, most colonies (>95%) remained in the opaque phase. The cellular morphology of representative colonies is shown in Figure S3, indicating that GlcNAc can stabilize the opaque phenotype of a/α strains at 37°C.
In MTL homozygous strains of C. albicans, white and opaque cells show differences in their behaviors in systemic and skin infection models [7],[8]. White cells are more virulent in systemic mouse model than opaque cells, while opaque cells are better at cutaneous infections. We then tested whether white and opaque cells of C. albicans MTLa/α strains also differed in virulence in different infection models. As shown in Figure 5A, in a systemic mouse infection system, burdens of opaque cells in the liver were notably less than those of white cells of RVVC10 and SZ306 (Student's t test p value<0.05), suggesting opaque cells of C. albicans a/α strains proliferated or colonized less well than their white cell counterparts. This was also the case for colonization of the kidney for RVVC10, although the difference of fungal cell burden between white and opaque cells of SZ306 was not significant. This result is consistent with previous studies; the fungal burdens of opaque cells of the MTL homozygous reference strain WO-1 in both the kidney and liver were less than those of white cells of WO-1 [7], and the fungal burdens of opaque cells of the MTL homozygous reference strain WO-1 in both the kidney and liver were less than those of white cells of WO-1. To test whether opaque cells of a/α strains were better at cutaneous infections, newborn mice were used and the fungal colonization of the skin was assessed by scanning electron microscopy as described previously [8]. Compared to white cells, opaque cells of both SZ306 and the reference strain WO-1 showed increased colonization in a cutaneous mouse model (Figure 5B). The number of opaque cells that colonized the skin was significantly higher than that of white cells (Student's t test p value<0.002) (Figure 5B). These results indicate that the different behaviors documented for white and opaque cells in the systemic and cutaneous mouse models also apply to white and opaque cells of the a/α strains described here.
To characterize the genome-wide transcriptional profiles of white and opaque cells of the MTLa/α strains, we performed RNA-Seq analysis of CY110, a clinical isolate of MTLa/α genotype. As shown in Table S4 (Sheet 1), the expression levels of 1,631 genes demonstrated a greater than twofold change in white and opaque cells. As expected, previously characterized white cell–enriched genes, such as WH11 and EFG1, were up-regulated in white cells, while opaque cell–enriched genes, such as WOR1 and OP4, were strongly up-regulated in opaque cells. A total of 838 genes demonstrated a greater than 3-fold change in our RNA-Seq analysis. Of them, 459 were previously reported as white (205) or opaque (254) cell–enriched genes [20],[21], and 379 were only found in our analysis, which could be MTL genotype-dependent phase-specific genes. As shown in Table 1, of the highly differentially expressed genes, the ratio of potential MTL genotype-dependent genes remarkably decreased, suggesting that highly differentially expressed genes are less MTL genotype-dependent. Interestingly, many cell wall protein and biofilm-induced genes were among the MTLa/α-specific genes (Table S4, sheet 3). Of note, the MTLa/α-specific genes may contain a proportion of genes specific to the strain background, especially for those with lower fold-change of expression levels.
Similar to the MTL homozygous strains, opaque and white cells of the MTLa/α strain CY110 specialized in their metabolic pathways (Table S4, sheet 2). Fermentative metabolism–associated genes were highly expressed in white cells of CY110 (e.g., glucose transporter genes HGT6, HGT7, and HGT8), while oxidative metabolism–associated genes were highly expressed in opaque cells (e.g., isocitrate dehydrogenase IDP2, malate synthase MLS1, acyl-CoA oxidase POX1, and 3-hydroxyacyl-CoA epimerase genes FOX2 and FOX3). Moreover, the differentially expressed genes in white and opaque cells of CY110, which were also found in their MTL homozygous counterparts, included genes associated with the metabolism of other nutrients (such as nitrogen and phosphate), cell wall components, stress response, and transcription factors (Table S4, sheet 2).
In MTL homozygous strains, only the opaque cell type is mating-competent [10]. Consistently, it has been demonstrated that mating-related genes MFα (α-pheromone) and STE2 (a-pheromone receptor) are highly enriched in opaque cells of WO-1, an MTLα/α isolate of C. albicans [20]. However, the expression levels of either MFα or MFa were not detectable in white and opaque cells of CY110. The expression levels of their receptors STE2 and STE3 in opaque cells of CY110 were very low and similar to that of white cells (Table S4, sheet 4). Additionally, the transcriptional expression of the four genes at the MTL loci (a1, a2, α1 and α2) was all detected. These results served to validate the a/α cell identity of CY110 and its mating incompetence.
WOR1 is the master regulator of white-opaque switching in MTL homozygotes of C. albicans and is extensively up-regulated in opaque cells of both MTL homozygotes and heterozygotes (Figure 1D and Table S4) [14]–[16]. Deletion of WOR1 in an MTLa/α strain SZ306u, a derivative of SZ306, blocked white to opaque switching on all media tested including Lee's GlcNAc medium in 5% CO2 (switching frequency<0.03%) (Figure S4). Under this culture condition, the white-to-opaque switching frequency of the wild-type SZ306u (WOR1/WOR1) and the single copy mutant (WOR1/wor1) were 4.3±1.0% and 0.5±0.3%, respectively, suggesting that the copy number of WOR1 could affect its own expression and the white-to-opaque switching frequency. Therefore, Wor1 is also essential for the induction of opaque phenotype in MTLa/α strains.
The a1/α2 complex inhibits the expression of WOR1 and thus controls white-to-opaque switching in SC5314 background strains [14]–[16]. The promoter region of WOR1 is extremely long (>10 kb), indicating the regulation of WOR1 expression could be very complex. Two facts imply that the a1/α2 complex does not work alone to control the expression of WOR1. First, even in the MTL homozygous strains (that therefore lack the a1/α2 complex), the default cell type is the white form, at least in typical laboratory media, indicating some other regulators must repress the expression of WOR1. Second, there appears to be only a single binding site of the a1/α2 complex on the long promoter region of WOR1. To find the regulators coordinately working with the a1/α2 complex in repressing WOR1 expression, we screened a library of ∼160 transcription factor null mutants (of the MTLa/α genotype) of SC5314 background [22]. The library was suitable for the screening because SC5314 and its derivatives (a/α) used for making the mutants are nonswitchable on the Lee's GlcNAc medium. We predicted that inactivating the transcription factors involved in inhibiting WOR1 expression would lead to the opaque phenotype. And we found three a/α mutants (rfg1/rfg1, brg1/brg1, and efg1/efg1) could undergo white-to-opaque switching on the Lee's GlcNAc medium in 5% CO2 at 25°C, suggesting the transcription factors Rfg1, Brg1, and Efg1 are involved in the regulation of white-opaque transition in MTLa/α strains of C. albicans. PCR analysis was conducted to confirm that the MTL genotype of the rfg1/rfg1, brg1/brg1, and efg1/efg1mutants were a/α (Figure 6A). Rfg1 is a member of the HMG domain family of sequence-specific DNA-binding proteins that has been shown to be a regulator of filamentous growth and virulence in C. albicans [23],[24]. We observed that the rfg1/rfg1 mutant (a/α) could also form opaque colonies or sectors in Lee's glucose and YPD media when cultured at 25°C for an extended time period (unpublished data). Consistent with the phenotype of rfg1/rfg1 mutant in white-opaque switching, Northern blots showed that the expression of RFG1 was enriched in white cells, relative to opaque cells in C. albicans MTLa/α strains (Figure 1D). Brg1, a GATA-type zinc finger transcription factor, has been characterized as a regulator of filamentous growth, biofilm formation, and virulence [22],[25]. Efg1 is a bHLH domain containing transcription factor required for maintaining the white cell phenotype of C. albicans MTL homozygotes [26]. The efg1/efg1 null mutants of MTLa/α strains could not switch to opaque in glucose containing medium [27]. However, both brg1/brg1 and efg1/efg1 mutants of MTLa/α strains could indeed undergo white-to-opaque switching on Lee's GlcNAc medium (Figure 6B). Our findings indicate that numerous environmental signals converge on Wor1 and regulate the ability of C. albicans cells to undergo white-opaque switching.
For decades, white-opaque switching was observed in only a minority (<10%) of natural C. albicans isolates: those that were homozygous at the mating-type locus [2],[4]. How does this species maintain such a complex switching system if the majority of strains (which are a/α) do not do it? One possibility is that white-opaque switching in C. albicans has been maintained as a means to attain mating competency [10]. However, C. albicans populations in the host are primarily clonal, indicating that, if parasexual mating actually occurs in nature, its role may not be to generate genetic diversity [28]. In this study, we have generated evidence for a different explanation for the widespread maintenance of white-opaque switching in C. albicans clinical isolates. We show that many naturally occurring MTLa/α strains of C. albicans can indeed undergo white-opaque switching, with the opaque phenotype of MTLa/α strains of C. albicans being largely similar to that of MTL homozygotes, except that they do not mate. Although such switching of a/α strains does not readily occur under typical laboratory conditions, we show that the combination of GlcNAc and CO2 are strong inducers of switching in a/α strains. Importantly, some a/α strains can undergo white-to-opaque switching at 37°C, the physiological temperature of the human host. These conditions are believed to be present in host niches such as the gut, where glucose is limiting and the carbon sources are largely from GI mucus and cell debris of microbes [29]. Together with our recent discovery of white-opaque switching in MTLa/α heterozygotes of C. tropicalis [12], our findings thus generalize white-opaque switching to strains with all mating-type configurations and suggest that the ability to switch is conserved in C. albicans and C. tropicalis.
We have shown that opaque cells of MTLa/α isolates of C. albicans share many features with opaque cells of MTL homozygotes. However, there are some important differences. For example, opaque cells of MTLa/α isolates undergo mass conversion to white cells on glucose containing media, while opaque cells of MTL homozygotes are very stable. Thus, a/α opaque cells are not as stable as opaque a or α cells and require the continuous presence of the environmental signals. Secondly, opaque cells of MTLa/α isolates are mating-incompetent.
The MTLa1/α2 complex inhibits the expression of the master regulator WOR1, thereby blocking white-opaque switching in the laboratory strain SC5314 [14]–[16], which is an a/α strain. However, in the a/α strains described here, white-opaque switching is permitted; the a1/α2 complex “turns it down” but does not completely block white-opaque switching. The long upstream region of WOR1 implies that multiple environmental signals and transcriptional regulators feed into it, and thus it is easy to imagine that strains could vary in the precise response of Wor1 to environmental signals. We have demonstrated that more than one third of natural isolates of MTLa/α C. albicans strains tested in this study formed opaque or opaque-sectored colonies on Lee's GlcNAc plates in 5% CO2. We propose that the white-opaque phenotypic transition itself is a general feature of C. albicans, but the quantitative response of the switch to features of the environment and to the mating type configuration differs from strain to strain.
The regulation of white-opaque switching in MTL homozygotes involves an interlocking transcriptional circuit, in which Wor1 occupies the central position [27]. We propose that Wor1 also acts as a master regulator in the process of white-opaque switching in MTL heterozygotes of C. albicans. Ectopic expression of WOR1 in the “non-switchable” MTL a/α strain CAI4, a derivative of SC5314, induces white-to-opaque switching on glucose-containing laboratory media, suggesting that Wor1, if ectopically expressed, can override the repressing effect of the a1/α2 complex on the white-to-opaque transition [14],[16]. By screening a deletion mutant library of C. albicans, we have identified three transcription factors, Rfg1, Brg1, and Efg1, involved in the regulation of white-opaque switching in MTLa/α strains. These three transcription factors inhibit opaque cell formation in MTLa/α strains since their null mutants are capable of switching between white and opaque cell types. Consistent with the phenotype of their null mutants, the transcriptional expression of RFG1 and EFG1 was enriched in white cells of MTLa/α strains (Figure 1). Nobile et al. have recently demonstrated that Brg1 and Efg1 bind to nearly the entire 10 kb intergenic region between WOR1 and its adjacent divergent gene ORF19.4883 [25]. However, the location of the peaks of Brg1 and Efg1 binding were distant from the putative a1/α2 cis-regulatory sequence. These results not only provide direct evidence of Brg1 and Efg1 binding to the promoter of WOR1, but also indicate that they may work together with the a1/α2 complex to reduce the expression of Wor1 in white cells and prevent switching to opaque cells. The transcriptional repressor Rfg1 may work in a similar manner as Efg1 and Brg1. We propose that inactivation of any of these three regulators would lead to increased expression of WOR1, which then initiates a self-positive feedback loop to induce the opaque cell phenotype (Figure 7A). Together with Wor1, additional transcriptional regulators, such as the positive regulators Wor2 and Czf1, coordinately regulate the expression of WOR1 by binding directly to the WOR1 upstream intergenic region (A.D.H., C.J.N., and A.D.J. unpublished data), and maintain the cells in the opaque phase (Figure 7B). Consistent with the model in Figure 7B, deletion of WOR2 or CZF1 results in increased opaque-to-white switching frequencies [27].
In summary, we have shown that, in many naturally occurring C. albicans strains, the a1/α2 repressor is not an absolute block to white-opaque switching as it is in the standard laboratory strain SC5314. Rather, the a1/α2 activity reduces switching frequency (and renders the opaque form less stable) in these newly described strains, but this reduction can be partially overcome by the addition of GlcNAc and CO2 to the growth medium. We propose that the a1/α2 repressor and other regulators (including Efg1, Brg1, and Rfg1) as well as these environmental signals all impinge on the long regulatory region of Wor1, the master regulator of white-opaque switching. This information is somehow integrated by the Wor1 regulatory region, and the level of Wor1 transcription is set accordingly. Because Wor1 appears to be the major determinant of the white-opaque switch frequency [30], the model can account for nearly all the observations in this article.
The most important implication of the work is that all strains of C. albicans (not just strains that are homozygous at the mating type locus, as previously believed) can undergo white-opaque switching if the appropriate signals are present in the growth medium. Thus, we propose that multiple environmental inputs combined with internal transcriptional regulators can activate white-opaque switching in virtually all C. albicans strains. White-opaque switching, in essence, produces two radically different types of cells from the same genome, thereby explaining the ability of C. albicans to occupy different niches in the host. We believe that the discovery of white-opaque switching in naturally occurring a/α strains accounts for the widespread conservation of the white-opaque switching machinery.
The strains used in this study are listed in Table S5. YPD (20 g/L glucose, 20 g/L peptone. 10 g/L yeast extract) was used for routine growth. Lee's + glucose and Lee's + GlcNAc media were used for mating and white-opaque switching assays [17].
The plasmid pSFS2A-URA3 was generated by inserting two DNA fragments containing sequences homologous to the 5′- and 3-terminals of C. albicans URA3 gene into the ApaI/XhoI and SacII/SacI sites of pSFS2A [31]. The auxotrophic strain SZ306u for uridine was constructed by disruption of one copy of URA3 with the linearized plasmid pSFS2A-URA3 and then grown on 5-fluoroorotic acid (5-FOA) containing medium. The white-opaque switching-competence of SZ306u was then confirmed. SZ306 and RVVC10 were converted to SZ306a and RVVC10α by deletion of one MTL allele with the plasmid T2A-MTL (Srikantha and Soll, unpublished). The first copy of WOR1 was deleted with the PCR product of pGEM-URA3 with the primers of WOR1-5DR and WOR1-3DR in SZ306u [32]. The second copy of WOR1 was then deleted with the linearized plasmid T2A-WOR1 [33]. A couple of primer sets were used to confirm the correct disruption of WOR1 in SZ306u.
To construct the WOR1/WOR1::WOR1p-GFP, EFG1/EFG1::EFG1p-GFP, and WH11/WH11::WH11p-GFP strains, CY110 was transformed with PCR products of the GFP-caSAT1 fragment (amplified from the template plasmid pNIM1 with GFP reporter primers, Table S6) [34]. The forward primers contained 60 bp of hanging homology to the promoter region of WOR1, WH11, or EFG1, while the reverse primers contained 60 bp of hanging homology to the 3′-UTR of WOR1, WH11, or EFG1. Correct integration of the transformations was verified by genomic DNA PCR with checking primers. All primers used in this study are listed in Table S6.
The CAI genotypes of C. albicans isolates were determined as described by Sampaio et al. (2003) [35]. Briefly, the microsatellite locus CAI was amplified by PCR using a pair of primers (forward, 5′- ATG CCA TTG AGT GGA ATT GG -3′; reverse, 5′- AGT GGC TTG TGT TGG GTT TT -3′). The forward primer was 5′ fluorescently labeled with 6-carboxyfluorescein. The sizes of the amplicons were determined by GeneScan analysis using a DNA sequencer, and the number of trinucleotide repeat units in each fragment was calculated. Because of the diploid nature of C. albicans, the CAI genotype of a strain is determined by the repeat number in both alleles of the locus. For example, a strain with a genotype CAI 17–21 means that one allele of the locus contains 17 trinucleotide repeats and the other 21.
White-opaque switching and mating assays were performed as previously described [36]. The cells were incubated in air or in 5% CO2 for 4 to 10 days as indicated in the main text. We examined 350 to 500 colonies for each strain. More were tested for nonswitchable strains or on nonconducive media. To verify the colony phenotype, several randomly selected colonies were examined for the cellular morphology. The dye phloxine B, which exclusively stains opaque colonies red, was added to the media. Scanning electron microscopy (SEM) assay was described as we described previously [22]. To observe the mating response, 106 cells of each of the two mating strains indicated in the text were mixed and spotted onto Lee's GlcNAc agar and incubated at 25°C for 4 days. At least 1×107 cells of each mating patch were examined with a light microscopy. Quantitative mating assay was performed as previously described with slight modifications [10]. Briefly, the mating experiments were performed on Lee's GlcNAc medium at 25°C. The experimental opaque cell samples were collected from Lee's GlcNAc medium plates. To test the mating ability of the MTLa/α strain (SZ306u), 1×106 of MTLa/a (or MTLα/α) cells and 1×106 of MTLa/α cells were mixed and cultured on Lee's GlcNAc medium plates for 48 hours. The mating mixtures were resuspended, diluted, and plated onto three types of selectable plates (without uridine, or arginine, or both) for prototrophic growth. Mating efficiencies were calculated as previously described [14].
The library of transcription factor mutants contains the TF mutants generated by the Johnson lab [37] and strains collected from Candida community [36]. Cells of each mutant were plated onto Lee's GlcNAc plates and incubated at 25°C for 7 to 10 days. Opaque colonies were replated and tested for the MTL genotype with PCR.
White and opaque cells were cultured on Lee's glucose and Lee's GlcNAc plates at 25°C for 4 days, and then inoculated in Lee's glucose and Lee's GlcNAc liquid media, respectively. Cells were collected from the cultures in exponential phase for RNA extraction. Purified PCR products of WH11, OP4, EFG1, WOR1, and RFG1 genes were used to make probes for Northern blot hybridization. Primers used for the PCR reactions are listed in Table S6. RNA-Seq analysis was performed by the company BGI-Shenzhen.
Systemic infection of mice was performed elsewhere [22]. Male ICR mice (18–22 g) were used in this study. Each male mouse was intravenously injected with 200 µl 1× PBS containing 2×106 cells via the tail vein. Three to four mice per strain were used for the injections. Mice were sacrificed on the 3rd day postinfection. The fungal burdens of kidneys and livers were tested. The systemic infection experiments were performed for four independent times. Newborn ICR mice (2 to 4 days) were used for cutaneous infection. The experiments were performed according the protocol reported by Kvaal et al. [8]. The skin colonization by C. albicans cells was assessed by scanning electron microscopy. The skin infection experiments were performed for three independent trials. All animal experiments were performed according to the guidelines approved by the Animal Care and Use Committee of the Institute of Microbiology, Chinese Academy of Sciences. The present study was approved by the Committee.
The RNA-Seq data have been deposited into the NCBI Gene Expression Omnibus (GEO) portal under the accession number GSE43938.
|
10.1371/journal.pbio.2006571 | Organic cation transporter 3 (Oct3) is a distinct catecholamines clearance route in adipocytes mediating the beiging of white adipose tissue | Beiging of white adipose tissue (WAT) is a particularly appealing target for therapeutics in the treatment of metabolic diseases through norepinephrine (NE)-mediated signaling pathways. Although previous studies report NE clearance mechanisms via SLC6A2 on sympathetic neurons or proinflammatory macrophages in adipose tissues (ATs), the low catecholamine clearance capacity of SLC6A2 may limit the cleaning efficiency. Here, we report that mouse organic cation transporter 3 (Oct3; Slc22a3) is highly expressed in WAT and displays the greatest uptake rate of NE as a selective non-neural route of NE clearance in white adipocytes, which differs from other known routes such as adjacent neurons or macrophages. We further show that adipocytes express high levels of NE degradation enzymes Maoa, Maob, and Comt, providing the molecular basis on NE clearance by adipocytes together with its reuptake transporter Oct3. Under NE administration, ablation of Oct3 induces higher body temperature, thermogenesis, and lipolysis compared with littermate controls. After prolonged cold challenge, inguinal WAT (ingWAT) in adipose-specific Oct3-deficient mice shows much stronger browning characteristics and significantly elevated expression of thermogenic and mitochondrial biogenesis genes than in littermate controls, and this response involves enhanced β-adrenergic receptor (β-AR)/protein kinase A (PKA)/cyclic adenosine monophosphate (cAMP)-responsive element binding protein (Creb) pathway activation. Glycolytic genes are reprogrammed to significantly higher levels to compensate for the loss of ATP production in adipose-specific Oct3 knockout (KO) mice, indicating the fundamental role of glucose metabolism during beiging. Inhibition of β-AR largely abolishes the higher lipolytic and thermogenic activities in Oct3-deficient ingWAT, indicating the NE overload in the vicinity of adipocytes in Oct3 KO adipocytes. Of note, reduced functional alleles in human OCT3 are also identified to be associated with increased basal metabolic rate (BMR). Collectively, our results demonstrate that Oct3 governs β-AR activity as a NE recycling transporter in white adipocytes, offering potential therapeutic applications for metabolic disorders.
| Adipose tissues (ATs) can be divided into three distinct types: white fat (or white AT [WAT]), brown fat, and beige fat. Growing evidence suggests that the development of beige fat cells in WAT, also called browning or beiging of WAT, might protect against obesity and improve systemic metabolism. Norepinephrine (NE)-induced β-adrenergic signaling is a major regulator of adaptive thermogenesis (a process of generating heat under conditions of physical activity) and leads to activation of protein kinase A (PKA) and phosphorylation of cyclic adenosine monophosphate (cAMP)-responsive element binding protein (Creb), thereby controlling the expression of thermogenic genes. In this study, we found that the catecholamine transporter, mouse organic cation transporter 3 (Oct3), is highly expressed in WAT, where it mediates NE uptake in the white fat cells in vivo and in vitro. Removing Oct3 in the fat cells leads to enhanced lipid breakdown, increased thermogenesis, and browning of WAT when stimulated by NE or cold exposure via activation of the β-adrenergic receptor (β-AR)/PKA/Creb pathway. In humans, reduced functional alleles of OCT3 are also associated with increased basal metabolic rate (BMR). Our results indicate that Oct3 is an essential regulator of NE recycling and the beiging of WAT.
| Obesity, a disease characterized by excess body fat, is a major risk factor for many human diseases, including type 2 diabetes, cardiovascular disease, and hepatic steatosis [1]. In mammals, fat is stored primarily in adipose tissue (AT), and three distinct types of ATs have been characterized: white AT (WAT), brown AT (BAT), and beige AT [2]. The morphology, function, cellular origins, and molecular features of the three types of ATs are quite distinct [3]. WAT stores nutrients as triglycerides in unilocular adipocytes, which can be used to generate free fatty acids (FFAs) by lipolysis [4]. BAT is the main tissue responsible for thermogenesis, especially when stimulated by cold [5]. After prolonged thermogenic induction, brown-like adipocytes can also be found in WAT, thus named beige or brite adipocytes [3]. The heat produced by BAT or beige AT is indispensable for survival during cold acclimatization. Also, the adaptive thermogenic beige AT may have the potential to counteract obesity and its related disorders [6, 7].
Beiging of WAT can be induced by various stimuli, including cold challenge, bile acids [8, 9], pharmacological agents, and hormones such as norepinephrine (NE) [3, 10]. NE activates β-adrenergic receptor (β-AR), which is coupled to G-proteins, resulting in increases in cyclic adenosine monophosphate (cAMP) concentration in adipocytes that enhance the activity of cAMP-dependent protein kinase A (PKA) [11]. This signal not only leads to lipolysis by phosphorylating hormone sensitive lipase (Hsl) but also by phosphorylating cAMP-responsive element binding protein (CREB), subsequently increasing the expression levels of thermogenic genes (UCP1, PGC1α, and DIO2) to promote nonshivering thermogenesis [11, 12]. The tightly regulated dichotomy of NE release and clearance maintains the balance between NE-induced thermogenesis and excessive NE in the vicinity of adipocytes. Thus, it is critical to elucidate the mechanism of NE clearance in AT. Although previous studies reported NE clearance mechanisms via SLC6A2 on sympathetic neurons [13] or proinflammatory macrophages in AT [14, 15], the low catecholamine clearance capacity of SLC6A2 may limit its cleaning efficiency.
Catecholamines, including NE, epinephrine, serotonin, and dopamine, act as neuromodulators in the central nervous system and as hormones in ATs and blood circulation [16]. Catecholamines are actively cleared in extracellular environment and are translocated into cells by specialized transporters that belong to two distinct transport mechanisms: neuronal transport (Uptake1, which is mediated by SLC6A2, SLC6A3, and SLC6A4) and extraneuronal transport (Uptake2, which is mediated by organic cation transporters [OCTs]) [17]. Studies have suggested a role of AT in the clearance and metabolism of catecholamines. Indeed, it has been shown that plasma concentrations of epinephrine and NE decrease after passing through AT [18, 19]. In addition, human adipocytes exhibit high expression levels and enzymatic activity of the catecholamine-degrading enzyme monoamine oxidase (MAO) [20]. Transport of catecholamines by adipocytes displays similar characteristics as the Uptake2 system and can be inhibited by the OCT3 inhibitor disprocynium 24 [20]. Ayala-Lopez et al. identified that perivascular ATs in male Sprague-Dawley rats had a NE uptake mechanism and could be reduced by norepinephrine transporter (NET) or/and OCT3 inhibitors ex vivo [21]. We reason that extraneuronal monoamine transporter OCT3 is an important route in NE clearance and adaptive thermogenesis in adipocytes complementary to sympathetic neurons or macrophages. In addition, the SLC22A3 (OCT3)-LPAL2-LPA gene cluster has been identified as a risk locus for coronary artery disease, implying its potential relationship to lipolysis [22].
In this study, we identified Oct3 as a novel non-neural NE clearance route in adipocytes and demonstrated its role in governing β-AR activity to mediate the beiging of WAT. Adipose-specific Oct3 knockout (KO) mice under cold acclimatization or NE administration exhibited significantly higher expression levels of genes related to thermogenesis, mitochondria biogenesis, and glycolysis in inguinal WAT (ingWAT) compared with Oct3fl/fl (Ctrl) mice. Of note, through mining of the Genome-Wide Association Studies (GWAS) Catalog, the database of Genotypes and Phenotypes (dbGAP), and UK BioBank databases, we also identified genetic variants of human OCT3 associated with higher basal metabolic rate (BMR). These findings suggest that Oct3 regulates catecholamines levels in the vicinity of the β-AR in adipocytes and therefore plays essential roles in the browning of WAT during adaptive thermogenesis. Inhibition of OCT3 may provide distinct therapeutic application by activating energy expenditure pathways.
Oct3 mRNA levels were examined in several C57BL/6J mouse tissues by real-time PCR. The highest expression levels were found in ingWAT, followed by gonadal WAT (gonWAT) (Fig 1A). In human tissues, OCT3 had the highest transcript levels in skeletal muscle and liver, followed by ATs (S1A Fig). Compared with transcript levels of other catecholamine transporters Slc6a2, Slc6a3, and Slc6a4 in murine WAT, mRNA levels of Oct3 were much higher (Fig 1B), suggesting that Oct3 is the predominant transporter for catecholamines in WAT. We further showed that adipocytes expressed high levels of NE degradation enzymes Maoa, Maob, and Comt, which was as high as the adipocytes abundant gene Adiponectin in ingWAT (Fig 1C), providing the molecular basis on NE clearance by adipocytes itself together with its transporter Oct3. Fractionation of ingWAT showed that Oct3 was enriched in mature adipocytes rather than in stromal vascular cells (SVCs) (Fig 1D), with Perilipin used as a marker for mature adipocytes, as its expression was limited to mature adipocytes and was not found in SVC (S1B Fig). Similar expression patterns of Oct3 in mature adipocytes and SVC were also observed in gonWAT (S1C and S1D Fig). Immunofluorescence of WAT localized Oct3 to cell membranes of adipocytes (Fig 1E).
To compare the Oct3 uptake activity against its main substrates (NE, serotonin, 1-methyl-4-phenylpyridinium (MPP+), epinephrine, dopamine, and tyramine), substrate uptake assays were performed in HEK-293 Flp-In cells stably expressing Oct3. NE had the highest uptake activity when cells were incubated with equivalent concentrations of catecholamines (Fig 1F). Human OCT3 and mouse Oct3-overexpressing HEK-293 cells both showed strong NE transport activities compared with empty vector (EV) in vitro, with Km and Vmax of 0.182 ± 0.0275 mM, 3.57 ± 0.174 nmol/mg protein/min (human) and 0.336 ± 0.0726 mM, 5.77 ± 0.448 nmol/mg protein/min (mouse), respectively (Fig 1G). Both Km and Vmax of OCT3 were more than 600 times greater than previously reported values of the Uptake1 transporter, SLC6A2 (0.28 ± 0.03 μM; 5.83 ± 0.49 pmol/mg protein/min) [23], consistent with OCT3 being a high-capacity, low-affinity Uptake2 transporter of NE. These data suggested an important role of OCT3 in regulating NE concentrations in peripheral tissues including AT, especially when a high concentration of NE existed, such as under NE stimulation or cold challenge.
To further examine the interactions between catecholamines and Oct3, we built a three-dimensional Oct3 homology model based on human glucose transporter 3 (GLUT3) template. Docking of structurally divergent catecholamines (NE, epinephrine, histamine, dopamine, serotonin, and tyramine) indicated that epinephrine (E = −45.7) and NE (E = −43.6) showed the best docking energy scores, resulting from favorable interactions formed between these two bioamines and polar residues in the primary substrate binding site of Oct3 including K215, Q242, E385, and E446 (S1E Fig).
To address the connection between Oct3, catecholamines, and adipocyte metabolism in vivo, we crossed Oct3fl/fl mice with Adiponectin-Cre mice to conditionally delete Oct3 in adipocytes (S2A Fig). Disruption of Oct3 was confirmed by immunohistochemistry (Fig 1H) and western blot analysis in ATs (S2B Fig). In addition, no compensational changes were observed in expression levels of the other catecholamine transporters Slc6a2, Slc6a3, and Slc6a4 in Oct3-deficient ATs compared with controls (S2C Fig). When fed a normal chow diet, adipose-specific Oct3 KO mice were healthy and viable, with no significant differences in body weight (S2D Fig), metabolic parameters (S2E–S2H Fig), or adipose morphology (S2I Fig) compared with Ctrl mice under room temperature (RT).
To determine NE uptake activity in different adipocytes, we performed [3H]-NE uptake assays in primary white and brown adipocytes. When treated with NE at a physiologically relevant concentration (0.20 μM), NE accumulation was significantly reduced (80% reduction) in Oct3 KO white adipocytes compared with controls, but not in brown adipocytes, suggesting that Oct3 is required for NE uptake in white adipocytes (Fig 1I). Moreover, in vivo NE uptake assay was performed after chemical sympathectomy, which eliminated neural NE uptake in ingWAT (S3A and S3B Fig). Oct3-null WAT showed significantly reduced total radioactivity, which reflected the accumulation of NE and its metabolites in adipocytes (73% and 65% reduction in ingWAT and gonWAT, respectively) (Fig 1J). Both the in vitro and in vivo data supported the role of Oct3 as the NE transporter in WAT.
Given that fact that NE was a preferential Oct3 substrate, we sought to investigate whether genetic deletion of Oct3 would influence NE clearance and NE-induced thermogenesis in vivo. After subcutaneous injection of low-dose NE (0.3 mg/kg), adipose-specific Oct3 KO mice showed significantly higher body temperature (Fig 2A and 2D) and higher maximum body temperature (Fig 2B and 2C) than that of Ctrl mice, as well as prolonged retention of increased body temperature (Fig 2A).
Adipose-specific Oct3 KO and Ctrl mice that received NE injections both demonstrated marked elevations in both whole-body O2 consumption (Fig 2E and 2F) and heat production (Fig 2G and 2H), while these parameters remained consistent between both genotypes under nonintervention conditions (S2E–S2H Fig). The magnitude of NE-induced O2 consumption (Fig 2E and 2F) and heat production (Fig 2G and 2H) in adipose-specific Oct3 KO mice was significantly higher than in the Ctrl littermates. Among the thermogenic markers tested, adipose-specific Oct3 KO mice treated with NE displayed a nearly 82-fold increase in Ucp1, a 5-fold increase in Ppargc1α (Pgc1α), and a 3-fold increase in Dio2, as well as an increasing trend in Hsl and Atgl mRNA expression in ingWAT (Fig 2I). Ucp1, Pgc1α, and Dio2 were also increased in gonWAT of adipose-specific Oct3 KO mice (Fig 2J). However, the relative expression of Ucp1 in gonWAT was far lower than that in ingWAT, suggesting that ingWAT might be more responsible for whole-body thermogenesis. No significant alterations in thermogenic genes were detected in BAT of both genotypes after NE injection (S4A Fig). Collectively, these results suggested that Oct3 deficiency improved thermogenic capacity in WAT in the presence of adrenergic stimuli that resulted in an increase in core body temperature, O2 consumption, and whole-body energy expenditure.
Activation of β-AR signaling has been found to promote lipolysis via the PKA-mediated phosphorylation of Hsl. To investigate the role of Oct3 in catecholamine signaling in lipolysis, we tested in vivo lipolysis by measuring serum concentration of FFAs. There was no difference in basal lipolysis, while NE injection enhanced lipolysis more significantly in adipose-specific Oct3 KO mice than in Ctrl (Fig 3A). Epinephrine had a similar effect on lipolysis in vivo (S4B Fig). Increased levels of Hsl phosphorylation at two PKA-responsive serine residues (pHsl-S563 and pHsl-S660) were observed in ingWAT (Fig 3B) and gonWAT (Fig 3C) of adipose-specific Oct3 KO mice 4 hours after NE stimulation, which is necessary for Hsl translocation into lipid droplets and NE-induced lipolysis in rodents [24]. No noticeable difference was observed in Hsl phosphorylation 4 hours after injection in BAT between the two genotypes (S4C Fig). One possible explanation for these results was that deletion of Oct3 led to impaired NE uptake into adipocytes and extracellular NE accumulation. Therefore, excess NE would activate β3-AR, increase cAMP concentration, and phosphorylate PKA, which would promote lipolytic activity via phosphorylating Hsl at S563 and S660 [25].
Next, we compared basal and NE-stimulated glycerol release from WAT explants derived from adipose-specific Oct3 KO mice and Ctrl mice. The rates of glycerol release were similar under basal conditions in both genotypes, but were enhanced more in Oct3 KO AT when treated with NE (Fig 3D). Ablation of Oct3 in primary adipocytes led to an elevated response to adrenergic stimulation, as demonstrated by an increased lipolytic response to NE (Fig 3E) and intracellular signaling molecule activation (pHsl-S563 and pHsl-S660) (Fig 3F).
To compare NE uptake capacity in Ctrl and Oct3 KO adipocytes, we assayed intracellular NE levels using immunofluorescence. Primary adipocytes were treated with catecholamine-degradation enzyme inhibitors decynium-22 in the presence or absence of NE and were then incubated with anti-NE antibodies. Staining patterns were similar between unstimulated adipocytes; however, Oct3 KO adipocytes had lower signals compared with Ctrl under NE stimulation, indicating that ablation of Oct3 in adipocytes impaired NE uptake (Fig 3G).
We next investigated the role of Oct3 on lipolysis in vitro by using a gain-of-function approach. Preadipocytes were transfected to produce Oct3-overexpressing 3T3-L1 cells (Oct3-OE). Although Oct3 overexpression did not significantly affect adipocyte differentiation based on Oil Red O staining (S4D Fig) or measurement of basal lipolysis (Fig 3H), Oct3-OE cells displayed attenuated NE-stimulated lipolysis compared with EV cells (Fig 3H), likely a result of enhanced NE transport to adipocytes accompanied by reduction of extracellular NE. In addition, Oct3-OE and EV cells were treated with the adenylyl cyclase agonist forskolin (10 μM), and similar lipolytic capacities were observed (S4E Fig), indicating that cAMP signaling (blocked by forskolin) and downstream lipolytic capacity were not primary causes of the difference in NE-stimulated lipolysis in our model.
In light of the enhanced lipolysis in Oct3 KO AT after NE stimulation, we examined the effect of adipose-specific KO of Oct3 on the liver. Fatty liver was not observed (S4F Fig), and there was no significant difference in liver triglyceride levels between genotypes 4 hours after NE injection (S4G Fig). One possible reason was that single NE injection represented only a short-term effect on lipolysis and FFA flux in blood, and another was that the 4-hour duration time might be insufficient to generate significant differences in fat accumulation in the liver.
To study the role of Oct3 in physiologically stimulated thermogenesis, adipose-specific Oct3 KO and Ctrl mice were exposed to a 4°C ambient environment for 1 month. Both groups had similar rectal temperatures under RT (Fig 4A). However, adipose-specific Oct3 KO mice showed significantly higher rectal temperatures (by 1.7°C) than those of Ctrl littermates after prolonged cold acclimation (Fig 4A and 4B). As seen in S5A Fig, body weight of adipose-specific Oct3 KO mice had a slight but significant decrease compared with that of Ctrl mice after sustained cold exposure. We also observed a more reddish appearance of ingWAT in adipose-specific Oct3 KO mice compared with Ctrl (Fig 4C). Notably, adipose-specific Oct3 KO ingWAT contained more clusters of uncoupling protein 1 (Ucp1)-positive multilocular adipocytes than Ctrl after cold exposure (Fig 4D). Whole-body energy expenditure of mice after cold challenge was measured by Comprehensive Lab Animal Monitoring System (CLAMS). Adipose-specific Oct3 KO mice showed markedly higher O2 consumption rates (OCRs) (Fig 4E), CO2 production rates (Fig 4F), and heat production (Fig 4G) than Ctrl, indicating a higher BMR. In addition, higher NE levels were observed in Oct3 KO ingWAT (S5B Fig), probably due to accumulation of extracellular NE derived from sympathetic nerve endings during cold challenge. Meanwhile, NE concentration in serum was also slightly increased, probably due to a leakage of extracellular NE in WAT into blood circulation (S5C Fig). These results showed the role of Oct3 in altering NE distribution between adipocytes and blood circulation.
In order to examine global gene expression changes in ingWAT, RNA from ingWAT of both genotypes were subjected to RNA sequencing (RNA-seq) analysis. Ablation of Oct3 led to significantly differential expression of 2,521 genes in ingWAT (adjusted p < 0.05) (S5D Fig and S1 Table). Widespread up-regulation of BAT-selective and beige genes were identified in Oct3 KO ingWAT (Fig 4H), while WAT-selective genes showed a down-regulation tendency (Fig 4I). Furthermore, expression of genes involved in fatty acid oxidation (FAO) also showed substantial up-regulation (Fig 4J). Based on the RNA-seq data analysis, we conducted real-time PCR analysis on genes involved in thermogenesis and adipose physiology. Oct3 KO ingWAT from cold-exposed mice had significantly higher mRNA expression levels of thermogenic genes Ucp1, Dio2, Pgc1α, Cidea, Pparα, Pgc1β, Tmem26, and Slc27a2—as well as genes that showed an up-regulation trend, including Tbx1 (Fig 4H)—whereas WAT-selective genes Gsn, Sncg, and Rarres2 were significantly decreased (Fig 4I). Several key genes of FAO were significantly up-regulated in Oct3 KO ingWAT, including Cpt1a, Hadha, Hadhb, Acdvl1, and Acsl5 (Fig 4J). Moreover, a negative regulator of FAO, Twist1, was significantly reduced in Oct3 KO ingWAT, by 80% (Fig 4J). No significant changes in morphology of BAT and gonWAT from both genotypes were observed (S5E Fig), consistent with unchanged mRNA levels of Pgc1α, Dio2, and other thermogenesis genes, as well as representative FAO genes (S5F and S5G Fig).
Given the close relationships between Oct3, NE, and thermogenesis, we next focused on the NE/β-AR/cAMP pathway. Western blot analysis showed that phospho-PKA (pPKA) substrate (Fig 4K), phospho-Creb (pCreb), and downstream Ucp1 and Pgc1α levels (Fig 4L) were significantly elevated in ingWAT from adipose-specific Oct3 KO mice compared with Ctrl. Protein levels of pCreb, Ucp1, and Pgc1α remained unchanged in BAT (S5H Fig) and gonWAT (S5I Fig). Moreover, genes encoding components of the cAMP/PKA pathway (Adcy3 and Adcy10) were also up-regulated in ingWAT after cold exposure (S5J Fig). Based on the data above, we hypothesized that the enhanced browning effect in adipose-specific Oct3 KO mice during cold challenge may be due to higher metabolic rate and elevated thermogenic gene expression.
We next characterized the effect of Oct3 deletion on mitochondria in ingWAT after cold exposure. Electron microscopy showed more accumulated mitochondria and enlarged mitochondrial area (Fig 5A), a larger number of mitochondria per cell area (Fig 5B), and increased mitochondrial DNA content (Fig 5C) in Oct3 KO adipocytes from ingWAT compared with Ctrl. Larger surface area of mitochondria was also observed in Oct3 KO ingWAT stained with Mitotracker (Fig 5D and 5E). Consistently, the gene ontology (GO) analysis of RNA-seq data revealed that differential genes associated with mitochondrial component were enriched (Fig 5F and S1 Table). Moreover, Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis highlighted specific changes in genes associated with oxidative phosphorylation, glycolysis and fatty acid metabolism, and citric acid cycle (Fig 5G and S1 Table).
In ingWAT of adipose-specific Oct3 KO mice, the increase in cold-induced thermogenesis was coupled with broad up-regulation of electron transport chain (ETC) genes (Fig 5H). Real-time PCR analysis showed that expression of mitochondrial respiratory chain complexes genes (Cox7b, Uqcrb, Uqcrh, and Cytochrome c) were remarkably increased in Oct3 KO ingWAT compared with Ctrl under cold stimulation (Fig 5I). In line with these changes, western blot analysis showed that protein levels of cytochrome c oxidase subunit 7b (Cox7b), Uqcrb, Uqcrh, and Cytochrome c were significantly elevated in Oct3 KO ingWAT from mice under cold exposure (Fig 5J). We found no difference in the expression of ETC genes in BAT (S5K Fig) and gonWAT (S5L Fig) between genotypes.
The profound differences of glycolytic genes in Oct3 KO ingWAT suggested that Oct3 deletion resulted in a metabolic rewiring during cold exposure (Fig 5K). Real-time PCR analysis of ingWAT showed up-regulation of glycolytic genes (Hk2, Pdk4, Pfkl, Eno1, Gapdh, Pfkm, and Pkm1) in adipose-specific Oct3 KO versus Ctrl mice (Fig 5L). The protein levels of a key enzyme of glycolysis, Hk2, and the amount of lactate (the product of glycolysis) were up-regulated in ingWAT from adipose-specific Oct3 KO mice (Fig 5M and 5N). These alterations were selective in ingWAT, as no significant difference between groups was observed in BAT (S5K and S5M Fig) or gonWAT (S5L and S5M Fig).
Collectively, our results suggested that the elevated browning effect in adipose-specific Oct3 KO mice may be owing to increased energy expenditure, mitochondrial biogenesis, and enhanced glycolysis in ingWAT.
To further investigate the relationship between Oct3-depletion–induced NE-stimulated thermogenesis and β-AR signaling, we firstly cotreated mice with NE and propranolol, a pharmacological β-AR inhibitor. An elevated thermogenic response in adipose-specific Oct3 KO mice was observed, including a more significant increase in core body temperature (Fig 6A), O2 consumption (Fig 6B), and heat production (Fig 6C) following single NE treatment, as described before (Fig 2A–2H). However, propranolol pretreatment abolished the elevation of thermogenic response (Fig 6A–6C), indicating that the increased energy expenditure in adipose-specific Oct3 KO mice was mediated by the β-AR signaling pathway.
To estimate the functional relevance of Oct3 for thermogenic capacity of adipocytes, we performed plate-based respirometry on primary adipocytes from preadipocytes isolated from SVC of ingWAT, which maintained expression of beige-related genes. Although the OCR curves were nearly identical in primary adipocytes of both genotypes in the absence of NE (S6A Fig), Oct3 KO beige adipocytes displayed a higher OCR than that of Ctrl cells after NE stimulation (Fig 6D). Additionally, pharmacological inhibition of β-AR by propranolol significantly blunted the NE-stimulated OCR increase to a similar extent in both groups (Fig 6D).
Primary adipocytes derived from ingWAT were treated with NE and triiodothyronine (T3) to induce Ucp1 expression [26]. An approximate 1.9-fold increase in Ucp1 mRNA expression in Oct3 KO beige adipocytes was found even in the absence of NE compared with Ctrl, and another 3.7-fold increase occurred in response to NE treatment (Fig 6E). Meanwhile, acute propranolol-induced inhibition of β-AR similarly blunted the NE-induced increase in Ucp1 transcripts in the beige both groups of adipocytes (Fig 6E). As a control, no noticeable difference was observed in mRNA expression of the NE-insensitive gene Ap2 between groups (S6B Fig). Additionally, Ucp1 transcripts and Creb phosphorylation showed no response to treatment of other Oct3 substrates (epinephrine, dopamine, and serotonin) (S6C and S6D Fig).
As NE was known to activate β-AR and trigger PKA-Creb signaling, we examined whether the enhanced thermogenesis in Oct3 KO beige adipocytes was regulated by the PKA/Creb pathway. NE-stimulated Oct3 KO adipocytes showed robust phosphorylation of PKA substrates and Creb, without changes in total Creb (Fig 6F). Meanwhile, these effects were largely eliminated by blockade of β-AR signaling (Fig 6F), indicating that β-AR signaling was required for enhanced thermogenic activity under Oct3 deficiency.
We next aimed to determine the requirement of NE/β-AR pathway in lipolysis induced by Oct3 ablation. Propranolol at 10 μM partially but significantly inhibited the enhancing effect of NE on lipolysis and Hsl phosphorylation, while 100 μM propranolol showed complete inhibition in Oct3 KO adipocytes (Fig 6G and 6H), indicating that this effect relied on β-AR function.
Based on these collective results, we propose the following model (Fig 6I). Oct3 locates to cell membranes of adipocytes and uptakes NE, which thereby decreases NE concentration in AT microenvironments. Oct3 deficiency reduces clearance of extracellular NE, subsequently increasing extracellular NE concentration and leading to enhanced lipolysis, glycolysis, thermogenesis, and WAT browning via β-AR/cAMP/PKA pathway activation.
Next, we explored the effect of adipose-specific Oct3 KO on diet-induced obesity (DIO). There was no difference in body weight between adipose-specific Oct3 KO and littermate Ctrl mice fed with a high-fat diet (HFD) and housed at thermoneutrality (30°C) (S7A Fig) or at RT (S7B Fig) for 12 weeks. We then subjected HFD-fed mice of both genotypes (housed at thermoneutrality for 12 weeks) to 4°C cold exposure. Higher body temperature was evident in adipose-specific Oct3 KO mice after the first 4 hours (S7C Fig). After 1-week cold acclimation, the body weights of adipose-specific Oct3 KO mice decreased more (S7D Fig) despite no difference in food intake (S7E Fig), serum glucose (S7F Fig), triglyceride (S7G Fig), cholesterol (S7H Fig), insulin (S7I Fig), and adiponectin (S7J Fig). Meanwhile, serum FFAs increased about 21% during cold stimulation in adipose-specific Oct3 KO mice (S7K Fig), implying higher lipolysis owing to lower NE clearance under Oct3 deficiency. Furthermore, glucose tolerance was improved in adipose-specific Oct3 KO mice compared with Ctrl mice (S7L Fig). These data showed that thermogenesis and lipolysis under cold stimuli were still enhanced in adipose-specific Oct3 KO mice treated with HFD, and insulin sensitivity was also improved.
To determine the relevance of the studies in mice to human, BMR with polymorphisms within the human OCT3 locus was explored through mining of GWAS Catalog, dbGAP, and UK BioBank databases. Among the 11 single-nucleotide polymorphisms (SNPs) in the OCT3 locus, four independent SNPs were associated with BMR with p < 5 × 10−8 (S8A Fig and S3 Table). From the GTEx Portal, we noted that 10 of the 11 SNPs are significantly associated with human OCT3 expression levels in ATs (subcutaneous and/or visceral omentum) (S8B and S8C Fig). Each of the alleles, which was associated with lower expression levels of OCT3 in AT, was correlated with higher BMR in the UK Biobank participants.
To further confirm the relevance to humans of the studies in mice, we performed experiments on human AT and primary human AT-derived mesenchymal stem cells (ATMSCs) that were induced to differentiate into adipocytes. Both human AT and differentiated ATMSCs showed high expression levels of OCT3 and NE-degrading enzymes (S8D–S8F Fig). These results were consistent with studies on human adipocytes showing that NE uptake existed and was inhibited by OCT3 inhibitors [20]. Furthermore, inhibition of OCT3 led to an increased lipolytic response to NE in differentiated ATMSCs (S8G Fig), similar to the observations in mouse stromal vascular fraction (SVF)-derived adipocyte model (Fig 4E).
Although human functional genomic data and cell studies provided cues for metabolic roles of OCT3, further investigations are needed to define the subtle roles of human OCT3 in AT.
Brown and beige AT are two key drivers for the dissipation of chemical energy to stimulate thermogenesis [27]. The current study demonstrates that genetic deletion of Oct3 enhances ingWAT beiging and up-regulates NE-induced thermogenesis, lipolysis, and mitochondria biogenesis by attenuating NE clearance. These findings support the non-neural role of Oct3 as a catecholamine scavenger for beige AT microenvironments in adaptive thermogenesis.
Recent studies have proposed a neural mechanism for NE uptake by Slc6a2 in special macrophages in AT [14, 15]. However, a proinflammatory state seems necessary to enhance their NE uptake in macrophages [14]. In addition, considering the high-affinity, low-capacity nature of Slc6a2 (Km: 0.28 ± 0.03 μM; Vmax: 5.83 ± 0.49 pmol/mg protein/min) [23], a NE transport system with greater transport capacity may be needed to meet increased metabolic demand during long-term cold challenge. With high expression in adipocytes and high capacity for catecholamines, Oct3 (Km: 0.183 ± 0.0275 mM; Vmax: 3.57 ± 0.174 nmol/mg protein/min) may function in part to clear catecholamines in white AT microenvironments. However, the quantitative contribution and exact conditions of extraneuronal NE uptake through Oct3 versus neuronal NE clearance remains to be investigated.
Cold challenge induces stronger browning and NE/β-AR/PKA signaling in ingWAT of adipose-specific Oct3 KO mice. Consistent with sustained action of catecholamines, protein levels of pPKA substrate, pCreb, and downstream Ucp1 and Pgc1α are all up-regulated. Pgc1α is the crucial transcriptional co-activator that is highly induced after cold exposure [28]. In white adipocytes, ectopic expression of Pgc1α induces Ucp1 expression and essential enzymes in mitochondria ETC and also increases mitochondrial DNA content [28, 29]. Mitochondria biogenesis and expression of ETC is significantly increased in ingWAT of adipose-specific Oct3 KO mice after cold challenge, which may result from elevated Pgc1α protein expression. Moreover, genes encoding fatty acid β-oxidation were exclusively up-regulated in ingWAT of Oct3-deficient mice after cold exposure. Consequently, more FFAs may be utilized as the primary fuel source for thermogenesis through Ucp1. These chain reactions are likely driven by more NE available due to Oct3 deletion.
During browning of ingWAT in adipose-specific Oct3 KO mice, metabolic rewiring of glucose metabolism to glycolysis is another significant phenomenon observed. Upon cold exposure, key enzymes in glycolysis and pyruvate dehydrogenase kinase 4 (Pdk4, the key regulatory enzyme linking glycolysis to the citric acid cycle) are significantly up-regulated in ingWAT of adipose-specific Oct3 KO mice. Higher expression of Pdk4 restricts the conversion of glucose to acetyl-CoA and reprograms it towards a higher glycolytic metabolism [30]. Along with other key glycolytic enzymes, the enhanced glycolysis in adipose-specific Oct3 KO mice may compensate for the loss of mitochondrial ATP production due to heat-generating mitochondrial uncoupling [31]. Recent reports have demonstrated that expression of glycolytic genes and Pdk4 are induced by β-AR agonists in BAT [32]. The enhanced glycolytic activity in ingWAT of cold-exposed adipose-specific Oct3 KO mice most likely results from stronger adrenergic stimulation through blunted NE catabolism by Oct3 deficiency. To date, there is little information available regarding the role of glucose metabolism in WAT. RNA-seq data also showed that these glycolytic genes are more highly expressed in ingWAT in adipose-specific Oct3 KO mice after cold exposure (S1 Table). These findings revealed that Oct3-mediated glucose metabolism is important to WAT energy homeostasis.
The differential responsiveness of WAT and BAT in adipose-specific Oct3 KO mice may result from several factors. First, Oct3 has a much lower expression in BAT compared with WAT. Second, current studies suggest that BAT is more densely innervated compared with WAT [33], and neuronal Slc6a2 may be largely responsible for the clearance of catecholamines released in BAT after NE injection or cold exposure [34]. Although browning of WAT can lead to metabolic improvements, the marked increase in energy expenditure (Fig 4E–4G) may also come from other sources such as skeletal muscle, cardiac and respiratory work, and liver, stimulated by enhanced circulating NE level [30–32]. Under HFD treatment at RT or thermoneutrality, similar body weights between two genotypes are observed. It is probably due to drastically reduced density of sympathetic nerves in ingWAT under HFD [35]. On the other hand, the Uptake1 system might be sufficient to compensate for the KO effect of Oct3. The overall proposed mechanism for Oct3 effects on WAT browning is summarized in Fig 6I.
Exploration of the Gene Atlas database for genetic associations of human BMR with variants in OCT3 shows consistency with some of the results in the adipose-specific Oct3 KO mouse model. In particular, several reduced expression variants of OCT3 are positively correlated with BMR in human. Though a greater baseline BMR is not observed in the adipose-specific Oct3 KO mice, metabolic rate is greater after either NE administration or cold challenge. Our study suggests an important clinical implication as β3-AR plays a central role in regulating nonshivering thermogenesis, which enhances energy expenditure and has potential to treat obesity [36]. Inhibition of OCT3 might block NE transport into adipocytes and mimic β3-AR agonists, resulting in browning of WAT and enhanced lipolysis. In summary, our results demonstrate that Oct3 is essential in catecholamine clearance in adipocytes of WAT and blunts intracellular responses upon sustained catecholamine stimulation by regulating the activity of β-AR. Development of OCT3-specific inhibitors may shed light on improving metabolic disorders.
All experiments were performed in accordance with guidelines of the Institute for Laboratory Animal Research of Tsinghua University. The experimental procedures were approved by the Administrative Committee of Experimental Animal Care and Use of Tsinghua University, licensed by the Science and Technology Commission of Beijing Municipality (SYXK-2014-0024), and they conformed to the National Institutes of Health guidelines on the ethical use of animals.
Adiponectin-Cre mice were obtained from the Jackson Laboratory (stock NO.010803). Oct3fl/fl mice were generated in the Ying Xu lab (Soochow University), backcrossed with C57BL/6J mice for at least 8 generations, and subsequently intercrossed with Adiponectin-Cre mice. Mice were housed under a 12-hour light/dark cycle (light period 7:00–19:00) at constant temperature (22°C RT). Food and water were available ad libitum. Male mice were 8 to 10 weeks of age when used for experiments.
HEK-293 Flp-In cells (Invitrogen, Carlsbad, CA) stably overexpressing human or mouse OCT3 were generated, maintained, and used for uptake assays as previously described [37]. For comparison between human OCT3 and mouse Oct3 NE uptake kinetics (Fig 2E), uptake time was 2 minutes at 37°C. Km and Vmax were calculated by fitting the data to Michaelis-Menten equations using GraphPad Prism software (La Jolla, CA). For [3H]-NE uptake in primary adipocytes, the same procedures were performed as described above.
For cold challenge experiments, male mice were placed at 16°C for 1 week, and then 3 weeks at 4°C. Metabolic measurements were measured using CLAMS (Columbus Instruments, Columbus, OH). Mice were individually housed in cages. Food and water were provided ad libitum.
The amino acid sequence of the mouse Oct3 was retrieved from the UNIPROT database (Q9WTW5) [38]. Homology models of Oct3 was generated via four steps. First, the primary sequence of Oct3 was submitted to the RCSB Protein Data Bank [39] to search for suitable template structures using PSI-BLAST [40]. Human glucose transporter 3 (hGLUT3) having its crystal structure determined in complex with D-glucose at 1.5 Å resolution in an outward-occluded conformation (PDB ID: 4ZW9) [41] is the closest structurally characterized protein to Oct3, sharing 24% sequence identity. Second, a sequence alignment between Oct3 and hGLUT3 was computed by Multiple Sequence Comparison by Log-Expectation (MUSCLE) [42]. Third, a total of 500 homology models were generated for Oct3 with the standard “automodel” class in MODELLER [43] on the basis of hGLUT3 template. Finally, these 500 models were evaluated by the discrete optimized protein energy (DOPE) score [44], and the best ranking model was selected.
Molecular docking screens against the selected Oct3 homology model was performed with a semi-automatic docking procedure. All docking calculations were performed with DOCK 3.6 [45]. The docking database contains seven compounds, dextroamphetamine, NE, epinephrine, dopamine, serotonin, histamine, and tyramine. The docked compounds were ranked by the docking energy that is the sum of van der Waals, Poisson-Boltzmann electrostatic, and ligand desolvation penalty terms.
Fresh tissue samples were fixed in 4% paraformaldehyde, dehydrated in serial alcohol, embedded in paraffin, cut into 5-μm–thick sections, and stained with hematoxylin–eosin (HE). Ucp1 immunohistochemical staining was performed as previously described [46]. Specimens were prepared and observed by transmission electron microscope (TEM) (H-7650; Hitachi, Japan) as previously described [47]. Mitochondria numbers were counted in more than 10 images for each specimen.
Mouse ingWAT frozen sections were then blocked for 30 minutes in PBS containing 10% normal goat serum at RT. For Oct3 staining, sections were incubated with a rabbit anti-mouse Oct antibody (1:200; OriGene Technologies, Rockville, MD) followed by staining with a Goat anti-Rabbit IgG (H+L) Cross-Adsorbed Secondary Antibody, Alexa Fluor 488 (1:500; Invitrogen, Carlsbad, CA). Then, sections were stained with DAPI for 10 minutes. Images were obtained with a Nikon A1 confocal microscope (Nikon Corp., Japan).
For mitochondria staining, deparaffinized slices of ingWAT from Ctrl and cKO mice were incubated with 250 nM MitoTracker (Molecular Probes, OR, USA) for 1 hour at RT. Slides were washed with PBS and had coverslips mounted. Images were obtained with a Nikon A1 confocal microscope (Nikon Corp., Japan). Mitochondria area fraction percentage was calculated as previously described [48] using NIS-Elements software (Nikon Corp., Japan). More than 20 fields were scored per group.
For detection of intracellular NE by immunofluorescence, primary adipocytes were incubated with medium containing NE (100 ng/ml) for 20 minutes at 37°C. After blocking, the adipocytes were next incubated with anti-noradrenaline (ab8887, 1:50, Abcam, Cambridge, MA) overnight at 4°C. The following procedures of immunofluorescence were described earlier.
Isolation of stromal-vascular cells from ingWAT and BAT was performed as described previously [49, 50]. Cells were cultured in growth medium DMEM/F-12 containing 10% FBS and 1% penicillin/streptomycin at 37°C with 5% CO2. For induction to differentiate, cells were incubated with induction medium containing 5 mg/ml insulin, 1 mΜ dexamethasone, 1 μM rosiglitazone, 0.5 mM 3-isobutyl-1-methylxanthine, 1 nM T3, and 125 nM indomethacin for 2 days. Cells were then maintained with 1 μM rosiglitazone and 1 nM T3 until maturation. Where indicated, cells were treated with T3 (10 μM), norepinephrine (100 nM) for 4 hours to induce Ucp1 expression.
3T3-L1 preadipocytes were incubated with an induction medium DMEM containing 1 μg/ml insulin, 0.25 μM dexamethasone, 0.5 mM 3-isobutyl-1-methylxanthine, and 2 μM rosiglitazone for 2 days, and then with a differentiation medium DMEM containing 1 μg/ml insulin for another 2 days. Fresh DMEM was supplied every 2 days until maturation. For Oil Red staining, fully differentiated 3T3-L1 cells were fixed with 4% paraformaldehyde, followed by Oil Red O incubation for 2 hours.
Ctrl and cKO mice (males, 8 weeks, chow diet) were intraperitoneally injected with pentobarbital (75 mg/kg), which does not inhibit nonshivering thermogenesis [51, 52]. For NE treatment, mice were subcutaneously injected with 0.3 mg/kg NE under anesthesia. For β-AR inhibition assay, propranolol (5 mg/kg) was given 20 minutes before NE. The core body temperature was recorded with a rectal probe connected to a digital thermometer (Yellow Spring Instruments). Infrared thermal images were obtained using a FLIR E60 compact infrared thermal imaging camera and were analyzed using FLIR Tools software (FLIR, North Billerica, MA).
For in vivo lipolysis, 0.3 mg/kg NE or saline was injected intraperitoneally into Ctrl and cKO mice. Tail blood was taken from mice 20 minutes after injection of NE or saline. Serum was separated for the determination of FFA using NEFA C kits (Wako, Oxoid SA, France).
For ex vivo lipolysis, about 50 mg WAT samples from overnight fasted mice were incubated in KRB buffer (12 mM HEPES, 121 mM NaCl, 4.9 mM KCl, 1.2 mM MgSO4, and 0.33 mM CaCl2) containing 2% FA-free BSA and 2.5 mM glucose, and then stimulated with 1 μM NE. Glycerol release was measured by using free glycerol reagent (Sigma-Aldrich, USA).
For in vitro lipolysis, glycerol release from differentiated SVF-derived or 3T3-L1 adipocytes was measured as previously described [53]. Lipolysis was stimulated by incubation of NE (1 μM, Sigma-Aldrich, USA) or forskolin (10 μM, Sigma-Aldrich, USA).
Chemical sympathectomy was performed by unilateral denervation in ingWAT on one side and sham-operation on the collateral side to eliminate SNS NE release and uptake, as described previously [54]. The effectiveness of sympathectomy was confirmed by immunofluorescence of tyrosine hydroxylse (Th).
After 3 days of recovery from surgery, mice were given intraperitoneal injection of 0.5 mg/kg of NE spiked with [3H]-NE (PerkinElmer, CT, USA). After 30 minutes, mice were transcardially perfused with ice-cold PBS and washed for 4 to 5 minutes. IngWAT was immediately isolated with the major blood vessels removed and homogenized in Solvable (PerkinElmer CT, USA) overnight. Total radioactivity in ingWAT homogenates was determined by liquid scintillation counting. Protein content of homogenates was measured by BCA assay (Thermo Scientific, USA).
For RNA-seq analysis of ingWAT, total RNA was extracted using the RNeasy Lipid Mini Kit (Qiagen, Hilden, Germany) based on the manufacturer’s instructions. Quality of RNA was analyzed using the Agilent 2100 (Agilent Technologies, Palo Alto, CA). The poly-A-containing mRNA was purified by using poly-T oligo-attached magnetic beads. Purified mRNA was fragmented into small pieces using divalent cations under elevated temperature. RNA fragments were copied into first-strand cDNA using reverse transcriptase and random primers, followed by second-strand cDNA synthesis using DNA polymerase I and RNase H. A single “A” base was added to cDNA fragments, and the adapter was subsequently ligated. The products were then purified and enriched with PCR amplification. PCR yield was quantified by Qubit, and samples were pooled together to make a single-strand DNA circle (ssDNA circle), which gave the final library.
DNA nanoballs (DNBs) were generated with the ssDNA circle by rolling circle replication (RCR) to enlarge the fluorescent signals at the sequencing process. The DNBs were loaded into the patterned nanoarrays and had 50 bp of single end read through the BGISEQ-500 platform. The DNB-based nanoarrays and stepwise sequencing were combined using Combinational Probe-Anchor Synthesis Sequencing Method and were analyzed.
Heatmap for differential genes was produced using pheatmap package of R. GO enrichment analysis and KEGG analysis for differential genes were performed in clusterProfiler package [55] version 2.2.4 of R using default settings.
Tissues and cells were lysed with RIPA buffer containing Halt Protease/Phosphatase inhibitors (Thermo Fisher Scientifics, Pittsburgh, PA). The primary antibodies used in this study were as follows: anti-phospho (Ser563)-HSL (#4139), anti-phospho (Ser660)-HSL (#4126), anti-HSL (#4107), anti-pPKA Substrate (#9621), and anti-phospho (Ser133)-CREB (#9198) (all from Cell Signaling Technology); anti-SLC22A3 antibody (#ab191446), anti-ATP5A (#ab176569), anti-COX7B (#ab140629), anti-UQCRH (#ab134949), anti-UQCRB (#ab190360), and anti-beta Actin (ab8226) (all from Abcam); anti-UCP1 (#GTX112784) (Genetex); and anti-PGC1α (#20658) (Proteintech Group Inc).
Total RNA was extracted by RNeasy Mini Lipid Tissue Kit (QIAGEN, Hilden, Germany). Reverse transcription (Tiangen, Beijing, China) and SYBR green quantitative PCR (Transgen, Beijing, China) were performed according to the manufacturer’s instructions. Primers for target sequences were as shown in S2 Table. Relative gene expression level was normalized by TATA box-binding protein (Tbp) expression levels, unless otherwise indicated, since TBP expression was stable across ATs [56].
For mitochondrial DNA analysis, mouse tissues were homogenized, and genomic DNA was extracted by TIANamp genomic DNA kit (Tiangen, Beijing, China). SYBR green quantitative PCR (Transgen, Beijing, China) was performed in duplicate using mitochondrial DNA specific primers for mitochondrial Cox2 and normalized by amplification of the nuclear gene ribosomal protein s18 (Rps18).
Lactate concentration in the ATs was measured by using a Lactate Assay Kit (Biovision, USA) following the instructions of the manufacturer.
Cellular OCRs were determined using a Seahorse XFe96 Extracellular Flux Analyzer (Seahorse Biosciences, Chicopee, MA). Primary adipocytes from ingWAT were seeded at 20,000 cells/well. Differentiation was induced as described above, and the cells were analyzed on day 6. Before the measurements, the cells were washed twice with assay medium (XF DMEM + 25 mM glucose + 2 mM pyruvate + 4 mM glutamine) and incubated in 175 μL of assay medium for 1 hour in an incubator without CO2 at 37°C. Port injection solutions were prepared as follows: NE (1 μM final concentration) or NE with propranolol (50 μM final concentration), oligomycin (2 μM final concentration), FCCP (1 μM final concentration), and a cocktail containing rotenone (1 μM) and antimycin A (1 μM). Each cycle consisted of mix 5 minutes, wait 0 minutes, and measure 5 minutes.
Publicly available resources were surveyed to identify human genetic studies associated with BMR. The resources included GWAS Catalog (https://www.ebi.ac.uk/gwas/home), dbGAP (https://www.ncbi.nlm.nih.gov/gap), and UK BioBank resource (http://geneatlas.roslin.ed.ac.uk/, https://www.biorxiv.org/content/early/2017/08/18/176834).
In Gene Atlas, which housed the results of all the genetic association studies analyses of the UK Biobank cohort (N = 408,455), BMR was available for over 7,000 participants. HaploReg version 4.1 [57] and GTEx Portal [58] were used to determine the location, consequence of the variants, and the effect of the variants on transcript levels of the gene. Information about the methods used for determining BMR of the participants were available online (http://biobank.ctsu.ox.ac.uk/crystal/docs/Anthropometry.pdf).
Six-week-old adipose-specific Oct3 KO and Ctrl mice were treated with HFD for 12 weeks under RT or thermoneutrality (30°C). For the groups under thermoneutrality, mice were then cold-stimulated at 4°C for 7 days. Body temperatures of mice were monitored at the first 8 hours under cold exposure. Body weights were recorded before and after cold exposure. Food intake was recorded daily during cold exposure. Blood was collected into EDTA tubes, and plasma was separated by centrifugation. Plasma triglyceride and total cholesterol levels were determined by respective assay kits (Nanjing Jiancheng Biotechnology Institute, Nanjing, China). Commercially available ELISA kits were used to measure plasma adiponectin (Abcam), insulin (ALPCO, Salem, NH), serum and tissue NE (Rocky Mountain Diagnostics, Colorado Springs, CO) levels following instructions from the manufacturers. Plasma FFAs was measured NEFA C kits (Wako, Oxoid SA, France). Liver triglyceride content was measured using Triglyceride Quantification Assay Kit (Abcam).
Oral glucose tolerance test was performed on mice fed with 12-week HFD under thermoneutrality and then acclimated cold exposure for 7 days by oral gavage of glucose (2 g/kg) after 12-hour overnight fasting.
Human MSCs derived from human AT, a kind gift from Dr. Yanan Du (Tsinghua University), were cultured in human MSC growth medium (Wuhan Viraltherapy Technologies Co. Ltd, Beijing) containing 10% calf serum and penicillin G. Adipogenic differentiation of human MSCs were performed as previously described [59].
Statistical analysis was performed using GraphPad Prism 5.0 (http://www.graphpad.com/scientific-software/prism). ANCOVA was conducted for in vivo metabolic data. Other analysis was conducted by Student t test (for comparison of two experimental conditions) or one-way ANOVA followed by Tukey's test (for comparison of three or more experimental conditions). Data are represented as mean ± standard deviation. Statistical significance was calculated and indicated (*p < 0.05, **p < 0.01, ***p < 0.001).
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.